For the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), it was decided once again to organise a dedicated Industry Track. Similar to the main research track of the conference, the industry track attracted an unprecedented number of 421 paper submissions. In total, 453 reviewers and 21 area chairs participated in the evaluation of these papers. After a thorough, double-blind peer-review evaluation with three reviews for each submission followed by reviewer discussions and additional deliberations, 108 papers were selected for presentation at the ACL 2025 Industry Track. Large language models were front and center of almost all submissions with trustworthiness, domain-adaptation, retrieval-augmented generation, and agentic architectures – across domains such as medical, legal, and finance – being popular topics.

Effective decision-making in Large Language Models (LLMs) is essential for handling intricate tasks. However, existing approaches prioritize performance but often overlook the balance between effectiveness and computational cost. To address this, we first introduce the 3E Criteria to systematically assess the cost-effectiveness of search strategies, revealing that existing methods often trade significant efficiency for marginal performance gains. To improve LLM decision-making while maintaining efficiency, we propose the Speculative Reward Model (SRM), a plug-and-play framework that seamlessly integrates with existing search strategies. Specifically, SRM employs an external reward assigner to predict optimal actions, reducing reliance on LLMs’ internal self-evaluation. And a speculative verification mechanism is used to prune suboptimal choices and guide the search toward more promising steps. We evaluate SRM on several complex decision-making tasks including mathematical reasoning, planning and numerical reasoning in specialized domains. Experimental results show that SRM reduces costs to 1/10 of the original search framework on average while maintaining effectiveness.

Advertisement (Ad) video violation detection is critical for ensuring platform compliance, but existing methods struggle with precise temporal grounding, noisy annotations, and limited generalization. We propose RAVEN, a novel framework that integrates curriculum reinforcement learning with multimodal large language models (MLLMs) to enhance reasoning and cognitive capabilities for violation detection. RAVEN employs a progressive training strategy, combining precisely and coarsely annotated data, and leverages Group Relative Policy Optimization (GRPO) to develop emergent reasoning abilities without explicit reasoning annotations. Multiple hierarchical sophisticated reward mechanism ensures precise temporal grounding and consistent category prediction. Experiments on industrial datasets and public benchmarks show that RAVEN achieves superior performances in violation category accuracy and temporal interval localization. We also design a pipeline to deploy the RAVEN on the online Ad services, and online A/B testing further validates its practical applicability, with significant improvements in precision and recall. RAVEN also demonstrates strong generalization, mitigating the catastrophic forgetting issue associated with supervised fine-tuning.

Enhancing computational efficiency and reducing deployment costs for large language models (LLMs) have become critical challenges in various resource-constrained scenarios. In this work, we present DistilQwen2.5, a family of distilled, lightweight LLMs derived from the public Qwen2.5 models. These distilled models exhibit enhanced instruction-following capabilities compared to the original models based on a series of distillation techniques that incorporate knowledge from much larger LLMs. In our industrial practice, we first leverage powerful proprietary LLMs with varying capacities as multi-agent teachers to select, rewrite, and refine instruction-response pairs that are more suitable for student LLMs to learn. After standard fine-tuning, we further leverage a computationally efficient model fusion approach that enables student models to progressively integrate fine-grained hidden knowledge from their teachers. Experimental evaluations demonstrate that the distilled models possess significantly stronger capabilities than their original checkpoints. Additionally, we present use cases to illustrate the applications of our framework in real-world scenarios. To facilitate practical use, we have released all the DistilQwen2.5 models to the open-source community.

Recommender systems play a central role in numerous real-life applications, yet evaluating their performance remains a significant challenge due to the gap between offline metrics and online behaviors. Given the scarcity and limits (e.g., privacy issues) of real user data, we introduce SimUSER, an agent framework that serves as believable and cost-effective human proxies. SimUSER first identifies self-consistent personas from historical data, enriching user profiles with unique backgrounds and personalities. Then, central to this evaluation are users equipped with persona, memory, perception, and brain modules, engaging in interactions with the recommender system. SimUSER exhibits closer alignment with genuine humans than prior work, both at micro and macro levels. Additionally, we conduct insightful experiments to explore the effects of thumbnails on click rates, the exposure effect, and the impact of reviews on user engagement. Finally, we refine recommender system parameters based on offline A/B test results, resulting in improved user engagement in the real world.

We present MegaBeam-Mistral-7B, a language model that supports 512K-token context length. Our work addresses practical limitations in long-context training, supporting real-world tasks such as compliance monitoring and verification. Evaluated on three long-context benchmarks, our 7B-parameter model demonstrates superior in-context learning performance on HELMET and robust retrieval and tracing capability on RULER. It is currently the only open model to achieve competitive long-range reasoning on BABILong at 512K context length without RAG or targeted fine-tuning. Released as fully open source under the Apache 2.0 license, the model has been downloaded over 100,000 times on Hugging Face.

Mathematical error detection in educational settings presents a significant challenge for Multimodal Large Language Models (MLLMs), requiring a sophisticated understanding of both visual and textual mathematical content along with complex reasoning capabilities. Though effective in mathematical problem-solving, MLLMs often struggle with the nuanced task of **identifying and categorizing student errors in multimodal mathematical contexts**. Therefore, we introduce **MathAgent, a novel Mixture-of-Math-Agent framework** specifically designed to address these challenges. Our approach decomposes error detection into three phases with specialized agents: an image-text consistency validator, a visual semantic interpreter, and an integrative error analyzer. This architecture enables more accurate processing of multimodal mathematical content by explicitly modeling the relationships between multimodal problems and student solution steps. We evaluate MathAgent on real-world educational data, demonstrating approximately 5% higher accuracy in error step identification and 3% improvement in error categorization compared to baseline models. Furthermore, MathAgent has been successfully deployed in an educational platform serving over one million K-12 students, achieving nearly 90% student satisfaction while generating significant cost savings by reducing manual error detection.

Despite advances in unsupervised log anomaly detection, current models require dataset-specific training, causing costly procedures, limited scalability, and performance bottlenecks. Furthermore, numerous models lack cognitive reasoning abilities, limiting their transferability to similar systems. Additionally, these models often encounter the **“identical shortcut”** predicament, erroneously predicting normal classes when confronted with rare anomaly logs due to reconstruction errors. To address these issues, we propose **MLAD**, a novel **M**ulti-system **L**og **A**nomaly **D**etection model incorporating semantic relational reasoning. Specifically, we extract cross-system semantic patterns and encode them as high-dimensional learnable vectors. Subsequently, we revamp attention formulas to discern keyword significance and model the overall distribution through vector space diffusion. Lastly, we employ a Gaussian mixture model to highlight rare word uncertainty, optimizing the vector space with maximum expectation. Experiments on real-world datasets demonstrate the superiority of MLAD.

This paper presents a novel approach to e-commerce payment fraud detection by integrating reinforcement learning (RL) with Large Language Models (LLMs). By framing transaction risk as a multi-step Markov Decision Process (MDP), RL optimizes risk detection across multiple payment stages. Crafting effective reward functions, essential for RL model success, typically requires significant human expertise due to the complexity and variability in design. LLMs, with their advanced reasoning and coding capabilities, are well-suited to refine these functions, offering improvements over traditional methods. Our approach leverages LLMs to iteratively enhance reward functions, achieving better fraud detection accuracy and demonstrating zero-shot capability. Experiments with real-world data confirm the effectiveness, robustness, and resilience of our LLM-enhanced RL framework through long-term evaluations, underscoring the potential of LLMs in advancing industrial RL applications.

Operations research (OR) is widely deployed to solve critical decision-making problems with complex objectives and constraints, impacting manufacturing, logistics, finance, and healthcare outcomes. While Large Language Models (LLMs) have shown promising results in various domains, their practical application in industry-relevant operations research (OR) problems presents significant challenges and opportunities. Preliminary industrial applications of LLMs for operations research face two critical deployment challenges: 1) Self-correction focuses on code syntax rather than mathematical accuracy, causing costly errors; 2) Complex expert selection creates unpredictable workflows that reduce transparency and increase maintenance costs, making them impractical for time-sensitive business applications. To address these business limitations, we introduce ORMind, a cognitive-inspired framework that enhances optimization through counterfactual reasoning. Our approach emulates human cognition—implementing an end-to-end workflow that systematically transforms requirements into mathematical models and executable solver code. It is currently being tested internally in Lenovo’s AI Assistant, with plans to enhance optimization capabilities for both business and consumer customers. Experiments demonstrate that ORMind outperforms existing methods, achieving a 9.5% improvement on the NL4Opt dataset and a 14.6% improvement on the ComplexOR dataset.

System-level testing is a critical phase in the development of large, safety-dependent systems, such as those in the automotive industry. However, creating test specifications can be a time-consuming and error-prone process. This paper presents an AI-based assistant to aid users in creating test specifications for system-level requirements. The system mimics the working process of a test developer by utilizing a LLM and an agentic framework, and by introducing intermediate test artifacts - structured intermediate representations derived from input requirements. Our user study demonstrates a 30 to 40% reduction in effort required for test development. For test specification generation, our quantitative analysis reveals that iteratively providing the model with more targeted information, like examples of similar test specifications, based on comparable requirements and purposes, can boost the performance by up to 30% in ROUGE-L. Overall, our approach has the potential to improve the efficiency, accuracy, and reliability of system-level testing and can be applied to various industries where safety and functionality are paramount.

Referred to as LLM-as-judge, a generative large language model (LLM) has demonstrated considerable efficacy as an evaluator in various tasks, including Machine Translation (LAJ-MT) by predicting scores or identifying error types for individual sentences. However, its dependability in practical application has yet to be demonstrated, as there is only an approximated match due to the task’s open-ended nature. To address this problem, we introduce a straightforward and novel meta-evaluation strategy PromptCUE and evaluate cutting-edge LAJ-MT models such as GEMBA-MQM. We identify their fundamental deficits, including certain label biases and the inability to assess near-perfect translations.To improve reliability, we investigate more trustworthy and less biased models using multidimensional prompt engineering. Our findings indicate that the combination of span-level error quantification and a rubric-style prompt tailored to the characteristics of LLMs has efficiently addressed the majority of the challenges current LAJ-MT models face. Furthermore, it demonstrates a considerably enhanced alignment with human values. Accordingly, we present Rubric-MQM, the LAJ-MT for high-end models and an updated version of GEMBA-MQM.

Social media platforms have enabled large-scale influence campaigns, impacting democratic processes. To fight against these threats, continuous training is needed. A typical training session is based on a fictive scenario describing key elements which are instantiated into a dedicated platform.Such a platform simulates social networks, which host a huge amount of content aligned with the training scenario. However, directly using Large Language Models to create appropriate content result in low content diversity due to coarse-grained and high-level scenario constraints, which compromises the trainees’ immersion.We address this issue with SocialForge, a system designed toenhance the diversity and realism of the generated content while ensuring its adherence to the original scenario.Specifically, SocialForge refines and augments the initial scenario constraints by generating detailed subnarratives, personas, and events.We assess diversity, realism, and adherence to the scenario through custom evaluation protocol. We also propose an automatic method to detect erroneous constraint generation, ensuring optimal alignment of the content with the scenario.SocialForge has been used in real trainings and in several showcases, with great end-user satisfaction. We release an open-source dataset generated with SocialForge for the research community.

Behavioral therapy notes are important for both legal compliance and patient care. Unlike progress notes in physical health, quality standards for behavioral therapy notes remain underdeveloped. To address this gap, we collaborated with licensed therapists to design a comprehensive rubric for evaluating therapy notes across key dimensions: completeness, conciseness, and faithfulness. Further, we extend a public dataset of behavioral health conversations with therapist-written notes and LLM-generated notes, and apply our evaluation framework to measure their quality. We find that: (1) A rubric-based manual evaluation protocol offers more reliable and interpretable results than traditional Likert-scale annotations. (2) LLMs can mimic human evaluators in assessing completeness and conciseness but struggle with faithfulness. (3) Therapist-written notes often lack completeness and conciseness, while LLM-generated notes contain hallucinations. Surprisingly, in a blind test, therapists prefer and judge LLM-generated notes to be superior to therapist-written notes. As recruiting therapists for annotation is expensive, we will release the rubric, therapist-written notes, and expert annotations to support future research.

LoRA is a technique that reduces the number of trainable parameters in a neural network by introducing low-rank adapters to linear layers. This technique is used for fine-tuning and even training large transformer models from scratch. This paper presents the RunLoRA framework for efficient implementations of LoRA, which significantly improves the speed of neural network training and fine-tuning with low-rank adapters. The proposed implementation optimizes the computation of LoRA operations based on the shape of the corresponding linear layer weights, the input dimensions, and the LoRA rank by selecting the best forward and backward computation graphs based on FLOPs and time estimations. This results in faster training without sacrificing accuracy. The experimental results show a speedup ranging from 10% to 28% on various transformer models.

Large Language Models (LLMs) require high quality instruction data for effective alignment, particularly in code generation tasks where expert curated datasets are expensive to produce. We present Genetic-Instruct, a scalable algorithm for synthesizing large-scale, high quality coding instructions using evolutionary principles. Starting from a small set of seed instructions, Genetic-Instruct generates diverse and challenging instruction-code pairs by leveraging an Instructor-LLM for generation, a Coder-LLM for code synthesis, and a Judge-LLM for automatic quality evaluation. Our proposed approach is highly parallelizable and effective even with a small seed data and weaker generator models. We generated more than 7.5 million coding instructions with the proposed approach. Then we evaluated it by fine-tuning LLMs with the synthetic samples and demonstrated a significant improvement in their code generation capability compared to the other synthetic generation approaches and publicly available datasets. Our results highlight the efficiency, scalability, and generalizability of the Genetic-Instruct framework.

Construction of a general-purpose post-recognition error corrector poses a crucial question: how can we most effectively train a model on a large mixture of domain datasets? The answer would lie in learning dataset-specific features and digesting their knowledge in a single model. Previous methods achieve this by having separate correction language models, resulting in a significant increase in parameters. In this work, we present Mixture-of-Experts as a solution, highlighting that MoEs are much more than a scalability tool. We propose a Multi-Task Correction MoE, where we train the experts to become an “expert” of speech-to-text, language-to-text and vision-to-text datasets by learning to route each dataset’s tokens to its mapped expert. Experiments on the Open ASR Leaderboard show that we explore a new state-of-the-art performance by achieving an average relative 5.0% WER reduction and substantial improvements in BLEU scores for speech and translation tasks. On zero-shot evaluation, NeKo outperforms GPT-3.5 and Claude-3.5-Sonnet with 15.5% to 27.6% relative WER reduction in the Hyporadise benchmark. NeKo performs competitively on grammar and post-OCR correction as a multi-task model.

AI agents and business automation tools interacting with external web services require standardized, machine-readable information about their APIs in the form of API specifications. However, the information about APIs available online is often presented as unstructured, free-form HTML documentation, requiring external users to spend significant time manually converting it into a structured format. To address this, we introduce , a novel framework that transforms long and diverse API documentation pages into consistent, machine-readable API specifications. This is achieved through a carefully crafted pipeline that integrates large language models and rule-based algorithms which are guided by domain knowledge of the structure of documentation webpages. Our experiments demonstrate that generalizes well across hundreds of APIs, and produces valid OpenAPI specifications that encapsulate most of the information from the original documentation. has been successfully implemented in an enterprise environment, saving thousands of hours of manual effort and making hundreds of complex enterprise APIs accessible as tools for LLMs.

With the rapid expansion of e-commerce and continuous urban evolution, Geospatial Repartition, dividing geographical regions into delivery zones, is essential to optimize various objectives, e.g., on-time delivery rate, for last-mile delivery. Recently, large language models (LLMs) have offered promising capabilities for integrating diverse contextual information that is beneficial for geospatial repartition. However, given the inherent uncertainty in LLMs, adapting them to practical usage in real-world repartition is nontrivial. Thus, we introduce CoAlign, a novel three-stage framework that calibrates LLM uncertainty to enable robust geospatial repartition by transforming the task into a ranking problem, integrating historical data with LLM-generated candidates. It first generates explainable candidate partitions with a multi-criteria strategy and then designs a novel conformal method to rank these candidates relative to historical partitions with coverage guarantees. Finally, CoAlign delivers candidates through an interactive decision support system. Extensive evaluation with real-world data shows that CoAlign effectively calibrates LLM uncertainty and generates partitions that better align with human feedback. Moreover, we have deployed CoAlign in one of the world’s largest logistics companies, significantly enhancing their delivery operations by increasing candidate acceptance rates by 300% and improving on-time delivery rates by 3%. Our work provides a novel angle to address industrial geospatial decision-making tasks by calibrating LLM uncertainty.

The vast portion of workloads employing LLMs involves answering questions grounded on PDF or scanned content. We introduce the Arctic-TILT achieving accuracy on par with models 1000× its size on these use cases. It can be finetuned and deployed on a single 24GB GPU, lowering operational costs while processing rich documents with up to 400k tokens. The model establishes state-of-the-art results on seven diverse Document Understanding benchmarks, as well as provides reliable confidence scores and quick inference, essential for processing files in large-scale or time-sensitive enterprise environments. We release Arctic-TILT weights and an efficient vLLM-based implementation on a permissive license.

We introduce a next-generation vandalism detection system for Wikidata, one of the largest open-source structured knowledge bases on the Web. Wikidata is highly complex: its items incorporate an ever-expanding universe of factual triples and multilingual texts. While edits can alter both structured and textual content, our approach converts all edits into a single space using a method we call Graph2Text. This allows for evaluating all content changes for potential vandalism using a single multilingual language model. This unified approach improves coverage and simplifies maintenance. Experiments demonstrate that our solution outperforms the current production system. Additionally, we are releasing the code under an open license along with a large dataset of various human-generated knowledge alterations, enabling further research.

Large Vision-Language Models (LVLMs) have transformed image captioning, shifting from concise captions to detailed descriptions. We introduce LOTUS, a leaderboard for evaluating detailed captions, addressing three main gaps in existing evaluations: lack of standardized criteria, bias-aware assessments, and user preference considerations. LOTUS comprehensively evaluates various aspects, including caption quality (e.g., alignment, descriptiveness), risks (e.g., hallucination), and societal biases (e.g., gender bias) while enabling preference-oriented evaluations by tailoring criteria to diverse user preferences. Our analysis of recent LVLMs reveals no single model excels across all criteria, while correlations emerge between caption detail and bias risks. Preference-oriented evaluations demonstrate that optimal model selection depends on user priorities.

Retrieval Augmented Generation (RAG) has emerged as a powerful application of Large Language Models (LLMs), revolutionizing information search and consumption. RAG systems combine traditional search capabilities with LLMs to generate comprehensive answers to user queries, ideally with accurate citations. However, in our experience of developing a RAG product, LLMs often struggle with source attribution, aligning with other industry studies reporting citation accuracy rates of only about 74% for popular generative search engines. To address this, we present efficient post-processing algorithms to improve citation accuracy in LLM-generated responses, with minimal impact on latency and cost. Our approaches cross-check generated citations against retrieved articles using methods including keyword + semantic matching, fine tuned model with BERTScore, and a lightweight LLM-based technique. Our experimental results demonstrate a relative improvement of 15.46% in the overall accuracy metrics of our RAG system. This significant enhancement potentially enables a shift from our current larger language model to a relatively smaller model that is approximately 12x more cost-effective and 3x faster in inference time, while maintaining comparable performance. This research contributes to enhancing the reliability and trustworthiness of AI-generated content in information retrieval and summarization tasks which is critical to gain customer trust especially in commercial products.

This paper introduces Light-R1, an opensource suite for training long reasoning modelsusing reproducible and cost-effective methodology. Given the proprietary nature of data usedin the DeepSeek-R1 series, we develop an alternative approach leveraging exclusively publicdata and models. Our curriculum training progressively increases data difficulty, combinedwith multi-staged post-training. Our LightR1-32B model, trained from Qwen2.5-32BInstruct, outperforms DeepSeek-R1-DistillQwen-32B in math reasoning. Experimental results show that this curriculum approachbecomes more effective when distinct, diverse datasets are available for different training stages: fine-tuning DeepSeek-R1-Distilledmodels (pre-tuned by DeepSeek team on proprietary data) with 3,000 challenging examplesfrom our curriculum dataset yielded state-ofthe-art 7B and 14B models, while the 32Bmodel, Light-R1-32B-DS performed comparably to QwQ-32B and DeepSeek-R1. Furthermore, we extend our work by applying GRPOon long reasoning models. Our final Light-R1-14B-DS achieves SOTA performance among14B models in math, with AIME24 & 25 scoresof 74.0 and 60.2 respectively, surpassing many32B models and DeepSeek-R1-Distill-Llama70B. Despite math-focused training, Light-R1-14B-DS demonstrates strong cross-domain generalization. Light-R1 represents a significantadvancement in making sophisticated reasoning models more accessible and implementablein real-world applications. Our models, training data and code have been made available.

Out-of-scope (OOS) intent detection is a critical challenge in task-oriented dialogue systems (TODS), as it ensures robustness to unseen and ambiguous queries. In this work, we propose a novel but simple modular framework that combines uncertainty modeling with fine-tuned large language models (LLMs) for efficient and accurate OOS detection. The first step applies uncertainty estimation to the output of an in-scope intent detection classifier, which is currently deployed in a real-world TODS handling tens of thousands of user interactions daily. The second step then leverages an emerging LLM-based approach, where a fine-tuned LLM is triggered to make a final decision on instances with high uncertainty.Unlike prior approaches, our method effectively balances computational efficiency and performance, combining traditional approaches with LLMs and yielding state-of-the-art results on key OOS detection benchmarks, including real-world OOS data acquired from a deployed TODS.

Discovering and evaluating long-form talk content such as videos and podcasts poses a significant challenge for users, as it requires a considerable time investment. Previews offer a practical solution by providing concise snippets that showcase key moments of the content, enabling users to make more informed and confident choices. We propose an LLM-based approach for generating podcast episode previews and deploy the solution at scale, serving hundreds of thousands of podcast previews in a real-world application. Comprehensive offline evaluations and online A/B testing demonstrate that LLM-generated previews consistently outperform a strong baseline built on top of various ML expert models, showcasing a significant reduction in the need for meticulous feature engineering. The offline results indicate notable enhancements in understandability, contextual clarity, and interest level, and the online A/B test shows a 4.6% increase in user engagement with preview content, along with a 5x boost in processing efficiency, offering a more streamlined and performant solution compared to the strong baseline of feature-engineered expert models.

The advent of Large Language Models (LLMs) has transformed how complex tasks across various domains can be automated. One of the industry trends today is Agentic AI, which leverages LLMs to operate multiple tools and provide automatic configuration.In the domain of cloud computing, Agentic AI might be used, for example, with the generation of Kubernetes manifests – structured configuration files that define containerized environments.However, effectively applying LLMs to domain-specific tasks often reveals knowledge gaps that impact the accuracy and reliability of the generated output.To address these challenges, we propose KGen, a pipeline for generating K8s manifests directly from user-described intents expressed in natural language using LLMs. Our approach leverages an extensive n-shot learning analysis to choose the appropriate number of examples that can better guide the adopted models in generating the manuscripts while also looking at the computational cost. Our results validate the use of LLM in this task and show that (as expected) increasing the number of n-shot examples can improve the quality of the generated configurations when adopting more specialized models, such as Mixtral-8x7B (which uses the Mixture of Experts approach) and Prometheus-8x7B-v2.0, but (surprisingly) for more general-purpose models like Llama3-8B and Llama3-70B, it can lead to smaller number of valid K8s manifests. These results underscore the complexities of adapting LLMs for domain-specific structured generation and emphasize the need for an in-depth analysis to determine the most effective setup, also suggesting that smaller models sometimes outperform their larger counterparts for each domain-specific task.

Tabular data analysis is crucial in many scenarios, yet efficiently identifying relevant queries and results for new tables remains challenging due to data complexity, diverse analytical operations, and high-quality analysis requirements. To address these challenges, we aim to recommend query–code–result triplets tailored for new tables in tabular data analysis workflows. In this paper, we present TablePilot, a pioneering tabular data analysis framework leveraging large language models to autonomously generate comprehensive and superior analytical results without relying on user profiles or prior interactions. Additionally, we propose Rec-Align, a novel method to further improve recommendation quality and better align with human preferences. Experiments on DART, a dataset specifically designed for comprehensive tabular data analysis recommendation, demonstrate the effectiveness of our framework. Based on GPT-4o, the tuned TablePilot achieves 77.0% top-5 recommendation recall. Human evaluations further highlight its effectiveness in optimizing tabular data analysis workflows.

Anomaly Detection (AD) focuses on detecting samples that differ from the standard pattern, making it a vital tool in process control. Logical anomalies may appear visually normal yet violate predefined constraints on object presence, arrangement, or quantity, depending on reasoning and explainability. We introduce LogicQA, a framework that enhances AD by providing industrial operators with explanations for logical anomalies. LogicQA compiles automatically generated questions into a checklist and collects responses to identify violations of logical constraints. LogicQA is training-free, annotation-free, and operates in a few-shot setting. We achieve state-of-the-art (SOTA) Logical AD performance on public benchmarks, MVTec LOCO AD, with an AUROC of 87.6% and an F₁-max of 87.0% along with the explanations of anomalies. Also, our approach has shown outstanding performance on semiconductor SEM corporate data, further validating its effectiveness in industrial applications.

Large Language Models (LLMs) require continuous updates to maintain accurate and current knowledge as the world evolves. While existing knowledge editing approaches offer various solutions for knowledge updating, they often struggle with sequential editing scenarios and harm the general capabilities of the model, thereby significantly hampering their practical applicability.This paper proposes a two-stage framework combining robust supervised fine-tuning (R-SFT) with model merging for knowledge editing. Our method first fine-tunes the LLM to internalize new knowledge fully, then merges the fine-tuned model with the original foundation model to preserve newly acquired knowledge and general capabilities. Experimental results demonstrate that our approach significantly outperforms existing methods in sequential editing while better preserving the original performance of the model, all without requiring any architectural changes. Code is available at [Applied-Machine-Learning-Lab/MM4KE](https://github.com/Applied-Machine-Learning-Lab/MM4KE).

Food delivery search aims to quickly retrieve deliverable items that meet users’ needs, typically requiring faster and more accurate query understanding compared to traditional e-commerce search. Generative retrieval (GR), an emerging search paradigm, harnesses the advanced query understanding capabilities of large language models (LLMs) to enhance the retrieval of results for complex and long-tail queries in food delivery search scenarios. However, there are still challenges in deploying GR to online scenarios: 1) **the large scale of items**; 2) **latency constraints unmet by LLM inference in online retrieval**; and 3) **strong location-based service restrictions on generated items**. To explore the application of GR in food delivery search, we optimize both offline training and online deployment, proposing **Hier**archical semantic representation enhancement for **G**enerative **R**etrieval (HierGR). Specifically, for the generation of semantic IDs, we propose an optimization method that refines the residual quantization process to generate hierarchically semantic IDs for items. Additionally, to successfully deploy on a well-known food delivery platform, we utilize the query cache mechanism and integrate the GR model with the online dense retrieval model to fulfill real-world search requirements. Online A/B testing results show that our proposed method increases **the number of online orders by 0.68%** for complex search intents. The source code is available at https://github.com/zhangfw123/HierGR.

The pretraining of code LLMs typically begins with general data and progresses to domain-specific data through sequential stages. In the latter stages, a challenging issue is that the data of a target domain can be limited in size, and conventional approach of increasing the number of epochs does not lead to a performance gain. In this paper, we propose a novel packing method, which is extracting overlapping contexts from the training data using variable-length stride. Our method can mitigate the data-scarcity issue by providing more diverse and abundant examples of next token prediction than non-overlapping contexts. While the training time of our approach is increased proportionally to the amount of augmented examples, we present space-efficient implementations to store overlapping contexts. Extensive experiments with real datasets show that our approach outperforms the conventional approach of controlling the number of epochs in terms of the pass@k rate.

We introduce DataMorgana, a tool for generating synthetic Q&A benchmarks tailored to RAG applications in enterprise settings. DataMorgana enables customization of the generated benchmark according to the expected diverse traffic of the RAG application. It allows for specifying question types and their associated distribution via a lightweight configuration mechanism. We demonstrate via a series of quantitative and qualitative experiments that DataMorgana surpasses existing tools in terms of lexical, syntactic, and semantic diversity of the generated benchmark while maintaining high quality. We run our experiments over domain-specific and general-knowledge public datasets, as well as two private datasets from governmental RAG applications: one for citizens and the other for government employees. The private datasets have been shared with us by AI71, an AI company, which has integrated DataMorgana into its offerings. In addition, DataMorgana has been offered to about 150 researchers worldwide as part of the SIGIR’2025 LiveRAG Challenge held in Spring 2025.

Large Language Models (LLMs) have shown capabilities in various natural language processing tasks, yet they often struggle with logical reasoning, particularly when dealing with complex natural language statements. To address this challenge, approaches that combine LLMs with symbolic reasoners have been proposed, where the LLM translates the natural language statements into symbolic representations, which are then verified by an external symbolic solver. However, ensuring syntactic correctness in these translations remains a significant challenge. To address this, we propose to constrain the outputs of the LLMs using Grammar-Constrained Decoding, showing that it consistently improves both syntactic correctness and semantic accuracy in logical parsing tasks. Our findings suggest that grammar constraints can serve as an effective substitute for in-context examples, especially beneficial for resource-constrained applications using smaller models.

We present AUTOSUMM, a large language model (LLM)-based summarization system deployed in a regulated banking environment to generate accurate, privacy-compliant summaries of customer-advisor conversations. The system addresses challenges unique to this domain, including speaker attribution errors, hallucination risks, and short or low-information transcripts. Our architecture integrates dynamic transcript segmentation, thematic coverage tracking, and a domain specific multi-layered hallucination detection module that combines syntactic, semantic, and entailment-based checks. Human-in-the-loop feedback from over 300 advisors supports continuous refinement and auditability.Empirically, AUTOSUMM achieves a 94% factual consistency rate and a significant reduction in hallucination rate. In production, 89% of summaries required no edits, and only 1% required major corrections. A structured model version management pipeline ensures stable upgrades with minimal disruption. We detail our deployment methodology, monitoring strategy, and ethical safeguards, showing how LLMs can be reliably integrated into high-stakes, regulated workflows.

Ensuring clinical data privacy while preserving utility is critical for AI-driven healthcare and data analytics. Existing de-identification (De-ID) methods, including rule-based techniques, deep learning models, and large language models (LLMs), often suffer from recall errors, limited generalization, and inefficiencies, limiting their real-world applicability. We propose a fully automated, multi-modal framework, RedactOR for de-identifying structured and unstructured electronic health records, including clinical audio records. Our framework employs cost-efficient De-ID strategies, including intelligent routing, hybrid rule and LLM based approaches, and a two-step audio redaction approach. We present a retrieval-based entity relexicalization approach to ensure consistent substitutions of protected entities, thereby enhancing data coherence for downstream applications. We discuss key design desiderata, de-identification and relexicalization methodology, and modular architecture of RedactOR and its integration with Oracle Health Clinical AI system. Evaluated on the i2b2 2014 De-ID dataset using standard metrics with strict recall, our approach achieves competitive performance while optimizing token usage to reduce LLM costs. Finally, we discuss key lessons and insights from deployment in real-world AI-driven healthcare data pipelines.

Industrial applications pose heightened requirements for consistency and reliability of large language models (LLMs). While LLMs are being tested with increasingly complex reasoning tasks, we argue that much can be learned via diagnostic tools that probe a fundamentally basic type of reasoning: conceptual consistency, e.g., a rule applying to “all surgeons” must also apply to “cardiac surgeons” since a cardiac surgeon is a type of surgeon. In this emerging industry track submission, we propose a method that takes concept hierarchies from a knowledge graph (KG) and automatically generates benchmarks that test conceptual consistency in LLMs. We develop a multi-domain benchmark that reveals rates of conceptual inconsistencies in several state of the art LLMs. Additionally, we use measured levels of inconsistency and disagreement in LLMs to find potentially problematic subgraphs in the reference KG. As such, it offers a scalable complement to symbolic curation, maintenance, and refinement of knowledge graphs, which is a critical activity in KG-based industrial applications.

Large language models (LLMs) have been widely used for relevance assessment in information retrieval. However, our study demonstrates that combining two distinct small language models (SLMs) with different architectures can outperform LLMs in this task. Our approach—QUPID—integrates a generative SLM with an embedding-based SLM, achieving higher relevance judgment accuracy while reducing computational costs compared to state-of-the-art LLM solutions. This computational efficiency makes QUPID highly scalable for real-world search systems processing millions of queries daily. In experiments across diverse document types, our method demonstrated consistent performance improvements (Cohen’s Kappa of 0.646 versus 0.387 for leading LLMs) while offering 60x faster inference times. Furthermore, when integrated into production search pipelines, QUPID improved nDCG@5 scores by 1.9%. These findings underscore how architectural diversity in model combinations can significantly enhance both search relevance and operational efficiency in information retrieval systems.

Recently, Large Language Models (LLMs) have been widely studied by researchers for their roles in various downstream NLP tasks. As a fundamental task in the NLP field, Chinese Grammatical Error Correction (CGEC) aims to correct all potential grammatical errors in the input sentences. Previous studies have shown that LLMs’ performance as correctors on CGEC remains unsatisfactory due to the challenging nature of the task. To promote the CGEC field to better adapt to the era of LLMs, we rethink the roles of LLMs in the CGEC task so that they can be better utilized and explored in CGEC. Considering the rich grammatical knowledge stored in LLMs and their powerful semantic understanding capabilities, we utilize LLMs as explainers to provide explanation information to the CGEC small models during error correction, aiming to enhance performance. We also use LLMs as evaluators to bring more reasonable CGEC evaluations, thus alleviating the troubles caused by the subjectivity of the CGEC task. In particular, our work is also an active exploration of how LLMs and small models better collaborate in downstream tasks. Extensive experiment and detailed analyses on widely used datasets verify the effectiveness of our intuition and the proposed methods.

Transformer-based large language models (LLMs) encounter challenges in processing long sequences on edge devices due to the quadratic complexity of attention mechanisms and growing memory demands from Key-Value (KV) cache. Existing KV cache optimizations struggle with irreversible token eviction in long-output tasks, while alternative sequence modeling architectures prove costly to adopt within established Transformer infrastructure. We present EdgeInfinite, a memory-efficient solution for infinite contexts that integrates compressed memory into Transformer-based LLMs through a trainable memory-gating module. This approach maintains full compatibility with standard Transformer architectures, requiring fine-tuning only a small part of parameters, and enables selective activation of the memory-gating module for long and short context task routing. The experimental result shows that EdgeInfinite achieves comparable performance to baseline Transformer-based LLM on long context benchmarks while optimizing memory consumption and time to first token.

Task-oriented dialogue systems employ natural language understanding (NLU) modules to manage the intricate and continually evolving business requirements of production systems.Although the development of Large Language Models (LLMs) introduced extraordinary chitchat capabilities, implementing LLMs into such systems brought new difficulties.One of the main challenges is the lack of specific datasets for training and evaluation of systems that offer both capabilities: chat and task. As NLU modules are designed to handle complex task requests and LLMs are utilized to specifically answer chitchat interactions, the system must correctly identify the functional intent of the user to utilize an applicable module. This paper presents CTFusion, a multi-turn dialogue generation framework designed to assist the evaluation and training of production systems that offer both capabilities. Utilizing the framework, we generate a multi-turn dialogue dataset for in-vehicle speech recognition system, which includes 41,211 dialogues of 240 real-world in-vehicle intents, and train In-vehicle Context Sensor (ICS), a lightweight model that successfully identifies the functional intent of the driver.ICS outperforms all baseline models across various experimental settings, which demonstrates that CTFusion can help generate relevant datasets with a complex business logic, which can subsequently assist production systems in leveraging LLMs for their chitchat capabilities.

Text-to-speech (TTS) technology has achieved impressive results for widely spoken languages, yet many under-resourced languages remain challenged by limited data and linguistic complexities. In this paper, we present a novel methodology that integrates a data-optimized framework with an advanced acoustic model to build high-quality TTS systems for low-resource scenarios. We demonstrate the effectiveness of our approach using Thai as an illustrative case, where intricate phonetic rules and sparse resources are effectively addressed. Our method enables zero-shot voice cloning and improved performance across diverse client applications, ranging from finance to healthcare, education, and law. Extensive evaluations—both subjective and objective—confirm that our model meets state-of-the-art standards, offering a scalable solution for TTS production in data-limited settings, with significant implications for broader industry adoption and multilingual accessibility. All demos are available in https://luoji.cn/static/thai/demo.html.

The architectural industry produces extensive documents, including method statements—expository documents that integrate multi-source data into actionable guidance. Manual drafting however is labor-intensive and time-consuming. This paper introduces ArchiDocGen, a multi-agent framework automating method statement generation. Unlike traditional approaches relying on static templates or single-pass generation, ArchiDocGen decomposes the task into three steps: outline generation, section-based content generation, and polishing, each handled by specialized agents. To provide domain expertise, ArchiDocGen employs a section-based retriever to fetch and synthesize relevant documents from its custom knowledge base. Each section is generated through iterative reasoning of a section-based chain-of-thought (SeCoT) scheme, followed by refinement to meet professional standards. To evaluate the generated method statements, we partner with the industry to establish a multi-dimensional evaluation system by combining automatic and empirical methods. Experiments show that ArchiDocGen achieves 4.38 ContentScore, excelling in specialization, completeness, organization, and clarity. Additionally, a web-based application for ArchiDocGen is developed and deployed with industry partners.

Current Large Language Model (LLM) evaluation frameworks utilize the same static prompt template across all models under evaluation. This differs from the common industry practice of using prompt optimization (PO) techniques to optimize the prompt for each model to maximize application performance. In this paper, we investigate the effect of PO towards LLM evaluations. Our results on public academic and internal industry benchmarks show that PO greatly affects the final ranking of models. This highlights the importance of practitioners performing PO per model when conducting evaluations to choose the best model for a given task.

As Large Language Models (LLMs) become deeply integrated into human life and increasingly influence decision-making, it’s crucial to evaluate whether and to what extent they exhibit subjective preferences, opinions, and beliefs. These tendencies may stem from biases within the models, which may shape their behavior, influence the advice and recommendations they offer to users, and potentially reinforce certain viewpoints. This paper presents the Preference, Opinion, and Belief survey (POBs), a benchmark developed to assess LLMs’ subjective inclinations across societal, cultural, ethical, and personal domains. We applied our benchmark to evaluate leading open- and closed-source LLMs, measuring desired properties such as reliability, neutrality, and consistency. In addition, we investigated the effect of increasing the test-time compute, through reasoning and self-reflection mechanisms, on those metrics. While effective in other tasks, our results show that these mechanisms offer only limited gains in our domain. Furthermore, we reveal that newer model versions are becoming less consistent and more biased toward specific viewpoints, highlighting a blind spot and a concerning trend.POBS: https://ibm.github.io/POBS

Electronic Discovery (eDiscovery) requires identifying relevant documents from vast collections for legal production requests. While artificial intelligence (AI) and natural language processing (NLP) have improved document review efficiency, current methods still struggle with legal entities, citations, and complex legal artifacts. To address these challenges, we introduce DISCOvery Graph (DISCOG), an emerging system that integrates knowledge graphs for enhanced document ranking and classification, augmented by LLM-driven reasoning. DISCOG outperforms strong baselines in F1-score, precision, and recall across both balanced and imbalanced datasets. In real-world deployments, it has reduced litigation-related document review costs by approximately 98%, demonstrating significant business impact.

Analyzing large volumes of case law to uncover evolving legal principles, across multiple cases, on a given topic is a demanding task for legal professionals. Structured topical reports provide an effective solution by summarizing key issues, principles, and judgments, enabling comprehensive legal analysis on a particular topic. While prior works have advanced query-based individual case summarization, none have extended to automatically generating multi-case structured reports. To address this, we introduce LexGenie, an automated LLM-based pipeline designed to create structured reports using the entire body of case law on user-specified topics within the European Court of Human Rights jurisdiction. LexGenie retrieves, clusters, and organizes relevant passages by topic to generate a structured outline and cohesive content for each section. Expert evaluation confirms LexGenie’s utility in producing structured reports that enhance efficient, scalable legal analysis.

In high-stakes industrial NLP applications, balancing generation quality with speed and efficiency presents significant challenges. We address them by investigating two complementary optimization approaches: Medusa for speculative decoding and knowledge distillation (KD) for model compression. We demonstrate the practical application of these techniques in real-world travel domain tasks, including trip planning, smart filters, and generating accommodation descriptions. We introduce modifications to the Medusa implementation, starting with base pre-trained models rather than conversational fine-tuned ones, and developing a simplified single-stage training process for Medusa-2 that maintains performance while reducing computational requirements. Lastly, we present a novel framework that combines Medusa with knowledge distillation, achieving compounded benefits in both model size and inference speed. Our experiments with TinyLlama-1.1B as the student model and Llama-3.1-70B as the teacher show that the combined approach maintains the teacher’s performance quality while reducing inference latency by 10-20x.

The discovery of novel antibiotics is critical to address the growing antimicrobial resistance (AMR). However, pharmaceutical industries face high costs (over $1 billion), long timelines, and a high failure rate, worsened by the rediscovery of known compounds. We propose an LLM-based pipeline that acts as an alert system, detecting prior evidence of antibiotic activity to prevent costly rediscoveries. The system integrates literature on organisms and chemicals into a Knowledge Graph (KG), ensuring taxonomic resolution, synonym handling, and multi-level evidence classification. We tested the pipeline on a private list of 73 potential antibiotic-producing organisms, disclosing 12 negative hits for evaluation. The results highlight the effectiveness of the pipeline for evidence reviewing, reducing false negatives, and accelerating decision-making. The KG for negative hits as well as the user interface for interactive exploration are available at https://github.com/idiap/abroad-kg-store and https://github.com/idiap/abroad-demo-webapp.

The evolution of Large Language Models (LLMs) has significantly advanced multi-turn conversation systems, emphasizing the need for proactive guidance to enhance users’ interactions. However, these systems face challenges in dynamically adapting to shifts in users’ goals and maintaining low latency for real-time interactions. In the Baidu Search AI assistant, an industrial-scale multi-turn search system, we propose a novel two-phase framework to provide proactive guidance. The first phase, Goal-adaptive Supervised Fine-Tuning (G-SFT), employs a goal adaptation agent that dynamically adapts to user goal shifts and provides goal-relevant contextual information. G-SFT also incorporates scalable knowledge transfer to distill insights from LLMs into a lightweight model for real-time interaction. The second phase, Click-oriented Reinforcement Learning (C-RL), adopts a generate-rank paradigm, systematically constructs preference pairs from user click signals, and proactively improves click-through rates through more engaging guidance. This dual-phase architecture achieves complementary objectives: G-SFT ensures accurate goal tracking, while C-RL optimizes interaction quality through click signal-driven reinforcement learning. Extensive experiments demonstrate that our framework achieves 86.10% accuracy in offline evaluation (+23.95% over baseline) and 25.28% CTR in online deployment (149.06% relative improvement), while reducing inference latency by 69.55% through scalable knowledge distillation.

High-quality Text-to-Speech (TTS) model training requires extensive and diverse text and speech data. It is challenging to procure such data from real sources due to issues of domain specificity, licensing, and scalability. Large language models (LLMs) can certainly generate textual data, but they create repetitive text with insufficient variation in the prompt during the generation process. Another important aspect in TTS training data is text normalization. Tools for normalization might occasionally introduce anomalies or overlook valuable patterns, and thus impact data quality. Furthermore, it is also impractical to rely on voice artists for large scale speech recording in commercial TTS systems with standardized voices. To address these challenges, we propose SpeechWeave, a synthetic speech data generation pipeline that is capable of automating the generation of multilingual, domain-specific datasets for training TTS models. Our experiments reveal that our pipeline generates data that is 10–48% more diverse than the baseline across various linguistic and phonetic metrics, along with speaker-standardized speech audio while generating approximately 97% correctly normalized text. Our approach enables scalable, high-quality data generation for TTS training, improving diversity, normalization, and voice consistency in the generated datasets.

Effectively selecting data from subgroups where a model performs poorly is crucial for improving its performance. Traditional methods for identifying these subgroups often rely on sensitive information, raising privacy issues. Additionally, gathering such information at runtime might be impractical. This paper introduces a cost-effective strategy that addresses these concerns. We identify underperforming subgroups and train a model to predict if an utterance belongs to these subgroups without needing sensitive information. This model helps mitigate bias by selecting and adding new data, which is labeled as challenging, for re-training the speech model. Experimental results on intent classification and automatic speech recognition tasks show the effectiveness of our approach in reducing biases and enhancing performance, with improvements in reducing error rates of up to 39% for FSC, 16% for ITALIC, and 22% for LibriSpeech.

Community Question Answering (CQA) platforms can be deemed as important knowledge bases in community, but effectively leveraging historical interactions and domain knowledge in real-time remains a challenge. Existing methods often underutilize external knowledge, fail to incorporate dynamic historical QA context, or lack memory mechanisms suited for industrial deployment. We propose ComRAG, a retrieval-augmented generation framework for real-time industrial CQA that integrates static knowledge with dynamic historical QA pairs via a centroid-based memory mechanism designed for retrieval, generation, and efficient storage. Evaluated on three industrial CQA datasets, ComRAG consistently outperforms all baselines—achieving up to 25.9% improvement in vector similarity, reducing latency by 8.7%–23.3%, and lowering chunk growth from 20.23% to 2.06% over iterations.

In the field of urban planning, general-purpose large language models often struggle to meet the specific needs of planners. Tasks like generating urban planning texts, retrieving related information, and evaluating planning documents pose unique challenges. To enhance the efficiency of urban professionals and overcome these obstacles, we introduce PlanGPT, the first specialized AI agent framework tailored for urban and spatial planning. Developed through collaborative efforts with professional urban planners, PlanGPT integrates a customized local database retrieval system, domain-specific knowledge activation capabilities, and advanced tool orchestration mechanisms. Through its comprehensive agent architecture, PlanGPT coordinates multiple specialized components to deliver intelligent assistance precisely tailored to the intricacies of urban planning workflows. Empirical tests demonstrate that PlanGPT framework has achieved advanced performance, providing comprehensive support that significantly enhances professional planning efficiency.

We investigate the utility of Large Language Models for automated taxonomy generation and completion specifically applied to taxonomies from the food technology industry. We explore the extent to which taxonomies can be completed from a seed taxonomy or generated without a seed from a set of known concepts, in an iterative fashion using recent prompting techniques.Experiments on five taxonomies using an open-source LLM (Llama-3), while promising, point to the difficulty of correctly placing inner nodes.

This paper presents a user-centered, empirically guided approach to multilingual metadata enrichment for children’s books. We combine LLMs with human-in-the-loop quality control in a scalable CI/CD pipeline to curate brand collections that enhance book discovery and engagement for young readers across multiple European markets. Our results demonstrate that this hybrid approach delivers high-quality, child-appropriate labels, improves user experience, and accelerates deployment in real-world production environments. This work offers practical insights for applying generative NLP in the media and publishing industry.

Understanding and effectively responding to email communication remains a critical yet complex challenge for current AI techniques, especially in corporate environments. These tasks are further complicated by the need for domain-specific knowledge, accurate entity recognition, and high precision to prevent costly errors. While recent advances in AI, specifically Large Language Models (LLMs), have made strides in natural language understanding, they often lack business-specific expertise required in such settings. In this work, we present Advanced Messaging Platform (AMP), a production-grade AI pipeline that automates email response generation at scale in real-world enterprise settings. AMP has been in production for more than a year, processing thousands of emails daily while maintaining high accuracy and adaptability to evolving business needs.

Modern text processing pipelines demand robust methods to remove extraneous content while preserving a document’s core message. Traditional approaches—such as HTML boilerplate extraction or keyword filters—often fail in multilingual settings and struggle with context-sensitive nuances, whereas Large Language Models (LLMs) offer improved quality at high computational cost. We introduce SORE (Semantic Outlier Removal), a cost-effective, transparent method that leverages multilingual sentence embeddings and approximate nearest-neighbor search to identify and excise unwanted text segments. By first identifying core content via metadata embedding and then flagging segments that either closely match predefined outlier groups or deviate significantly from the core, SORE achieves near-LLM extraction precision at a fraction of the cost. Experiments on HTML datasets demonstrate that SORE outperforms structural methods and yield high precision in diverse scenarios. Our system is currently deployed in production, processing millions of documents daily across multiple languages while maintaining both efficiency and accuracy. To facilitate reproducibility and further research, we will publicly release our implementation and evaluation datasets.

Named Entity Recognition (NER) for low-resource variants of English remains challenging, as most NER models are trained on datasets predominantly focused on American or British English. While recent work has shown that proprietary Large Language Models (LLMs) can perform NER effectively in low-resource settings through in-context learning, practical deployment is limited by their high computational costs and privacy concerns. Open-source Small Language Models (SLMs) offer promising alternatives, but the tendency of these Language Models (LM) to hallucinate poses challenges for production use. To address this, we introduce SLENDER, a novel output format for LM-based NER that achieves a three-fold reduction in inference time on average compared to JSON format, which is widely used for structured outputs. Our approach using Gemma-2-9B-it with the SLENDER output format and constrained decoding in zero-shot settings outperforms the en_core_web_trf model from SpaCy, an industry-standard NER tool, in all five regions of the Worldwide test set.

Virtual Teaching Assistants (VTAs) powered by Large Language Models (LLMs) have the potential to enhance student learning by providing instant feedback and facilitating multi-turn interactions. However, empirical studies on their effectiveness and acceptance in real-world classrooms are limited, leaving their practical impact uncertain. In this study, we develop an LLM-based VTA and deploy it in an introductory AI programming course with 477 graduate students. To assess how student perceptions of the VTA’s performance evolve over time, we conduct three rounds of comprehensive surveys at different stages of the course. Additionally, we analyze 3,869 student–VTA interaction pairs to identify common question types and engagement patterns. We then compare these interactions with traditional student-human instructor interactions to evaluate the VTA’s role in the learning process. Through a large-scale empirical study and interaction analysis, we assess the feasibility of deploying VTAs in real-world classrooms and identify key challenges for broader adoption. Finally, we release the source code of our VTA system, fostering future advancements in AI-driven education.

Practitioners working on dense retrieval today face a bewildering number of choices. Beyond selecting the embedding model, another consequential choice is the actual implementation of nearest-neighbor vector search. While best practices recommend HNSW indexes, flat vector indexes with brute-force search represent another viable option, particularly for smaller corpora and for rapid prototyping. In this paper, we provide experimental results on the BEIR dataset using the open-source Lucene search library that explicate the tradeoffs between HNSW and flat indexes (including quantized variants) from the perspectives of indexing time, query evaluation performance, and retrieval quality. With additional comparisons between dense and sparse retrievers, our results provide guidance for today’s search practitioner in understanding the design space of dense and sparse retrievers. To our knowledge, we are the first to provide operational advice supported by empirical experiments in this regard.

Effective content moderation is essential for video platforms to safeguard user experience and uphold community standards. While traditional video classification models effectively handle well-defined moderation tasks, they struggle with complicated scenarios such as implicit harmful content and contextual ambiguity. Multimodal large language models (MLLMs) offer a promising solution to these limitations with their superior cross-modal reasoning and contextual understanding. However, two key challenges hinder their industrial adoption. First, the high computational cost of MLLMs makes full-scale deployment impractical. Second, adapting generative models for discriminative classification remains an open research problem. In this paper, we first introduce an efficient method to transform a generative MLLM into a multimodal classifier using minimal discriminative training data. To enable industry-scale deployment, we then propose a router-ranking cascade system that integrates MLLMs with a lightweight router model. Offline experiments demonstrate that our MLLM-based approach improves F1 score by 66.50% over traditional classifiers while requiring only 2% of the fine-tuning data. Online evaluations show that our system increases automatic content moderation volume by 41%, while the cascading deployment reduces computational cost to only 1.5% of direct full-scale deployment.

Ambiguous user queries pose a significant challenge in task-oriented dialogue systems relying on information retrieval. While Large Language Models (LLMs) have shown promise in generating clarification questions to tackle query ambiguity, they rely solely on the top-k retrieved documents for clarification which fails when ambiguity is too high to retrieve relevant documents in the first place. Traditional approaches lack principled mechanisms to determine when to use broad domain knowledge vs specific retrieved document context for clarification. We propose AsK, a novel hybrid approach that dynamically chooses between document-based or aspect-based clarification based on query ambiguity. Our approach requires no labeled clarification data and introduces: (1) Weakly-supervised Longformer-based ambiguity analysis, (2) Automated domain-specific aspect generation using clustering and LLMs and (3) LLM-powered clarification generation. AsK demonstrates significant improvements over baselines in both single-turn and multi-turn settings (recall@5 gain of ~20%) when evaluated on product troubleshooting and product search datasets.

Foundational models endowed with emergent abilities are increasingly deployed as autonomous agents to navigate intricate environments. Despite their capability to comprehend human intentions, even when paired with reasoning traces, they struggle to achieve robust autonomy. In this work, we introduce **LEAP & LEAN**, a novel paradigm designed to enhance the performance of Large Language Models (LLMs) as autonomous agents. LEAP employs look-ahead planning to refine action selection, while LEAN streamlines navigation through agile prompt construction, enabling more efficient task completion. Together, LEAP & LEAN address the explore-exploit dilemma, fostering optimal decision-making and improving task performance. We evaluate our framework across diverse, multi-faceted task-oriented domains (WebShop, ALFWorld, and TravelPlanner) using both proprietary and open-source LLM agents. Notably, without any fine-tuning, our framework outperforms agents trained via imitation learning, reinforcement learning, and reasoning-based approaches. Our findings underscore the importance of action and prompt curation to create robust and efficient fully autonomous LLM agents.

In the retrieval stage of recommendation systems, two-tower models are widely adopted for their efficiency as a predominant paradigm. However, this method, which relies on collaborative filtering signals, exhibits limitations in modeling similarity for long-tail items. To address this issue, we propose a Motivation-aware Retrieval for Long-Tail Recommendation, named MotiR. The purchase motivations generated by LLMs represent a condensed abstraction of items’ intrinsic attributes. By effectively integrating them with traditional item features, this approach enables the two-tower model to capture semantic-level similarities among long-tail items. Furthermore, a gated network-based adaptive weighting mechanism dynamically adjusts representation weights: emphasizing semantic modeling for long-tail items while preserving collaborative signal advantages for popular items. Experimental results demonstrate 60.5% Hit@10 improvements over existing methods on Amazon Books. Industrial deployment in Taobao&Tmall Group 88VIP scenarios achieves over 4% CTR and CVR improvement, validating the effectiveness of our method.

Electronic Health Records contain vast amounts of valuable clinical data, much of which is stored as unstructured text. Extracting meaningful clinical events (e.g., disorders, symptoms, findings, medications, and procedures etc.) in context within real-world healthcare settings is crucial for enabling downstream applications such as disease prediction, clinical coding for billing and decision support.After Named Entity Recognition and Linking (NER+L) methodology, the identified concepts need to be further classified (i.e. contextualized) for distinct properties such as their relevance to the patient, their temporal and negated status for meaningful clinical use. We present a solution that, using an existing NER+L approach - MedCAT, classifies and contextualizes medical entities at scale. We evaluate the NLP approaches through 14 distinct real-world clinical text classification projects, testing our suite of models tailored to different clinical NLP needs. For tasks requiring high minority class recall, BERT proves the most effective when coupled with class imbalance mitigation techniques, outperforming Bi-LSTM with up to 28%. For majority class focused tasks, Bi-LSTM offers a lightweight alternative with, on average, 32% faster training time and lower computational cost. Importantly, these tools are integrated into an openly available library, enabling users to select the best model for their specific downstream applications.

We introduce an active-sampling-based framework for automatic prompt optimization, designed to enhance the performance of Large Language Model (LLM)-as-a-judge systems, which use LLMs to evaluate the quality of text or other outputs, in label-scarce settings. Unlike existing approaches that rely on extensive annotations, our method starts with no labeled data and iteratively selects and labels a small, diverse, and informative subset of samples to guide prompt refinement. At each iteration, our method evaluates the current prompt based on selected data and automatically updates the prompt, enabling efficient prompt optimization with minimal supervision. Moreover, we formulate sample selection as a convex optimization problem that balances uncertainty and diversity, maximizing the utility of limited labeling budgets. We validate our framework across four popular LLMs and three real-world datasets, including one from a deployed industry product. Results show that our optimized prompts consistently outperform baselines, achieving significant gains in evaluation quality and robustness while substantially reducing labeling costs.

With the emergence of ChatGPT, Transformer models have significantly advanced text classification and related tasks. Decoder-only models such as Llama exhibit strong performance and flexibility, yet they suffer from inefficiency on inference due to token-by-token generation, and their effectiveness in text classification tasks heavily depends on prompt quality. Moreover, their substantial GPU resource requirements often limit widespread adoption. Thus, the question of whether smaller language models are capable of effectively handling text classification tasks emerges as a topic of significant interest. However, the selection of appropriate models and methodologies remains largely underexplored. In this paper, we conduct a comprehensive evaluation of prompt engineering and supervised fine-tuning methods for transformer-based text classification. Specifically, we focus on practical industrial scenarios, including email classification, legal document categorization, and the classification of extremely long academic texts. We examine the strengths and limitations of smaller models, with particular attention to both their performance and their efficiency in Video Random-Access Memory (VRAM) utilization, thereby providing valuable insights for the local deployment and application of compact models in industrial settings.

Text chunking is fundamental to modern retrieval-augmented systems, yet existing methods often struggle with maintaining semantic coherence, both within and across chunks, while dealing with document structure and noise. We present AutoChunker, a bottom-up approach for text chunking that combines document structure awareness with noise elimination. AutoChunker leverages language models to identify and segregate logical units of information (a chunk) while preserving document hierarchy through a tree-based representation. To evaluate the chunking operator, we introduce a comprehensive evaluation framework based on five core tenets: noise reduction, completeness, context coherence, task relevance, and retrieval performance. Experimental results on Support and Wikipedia articles demonstrate that AutoChunker significantly outperforms existing methods, reducing noise while improving chunk completeness compared to state-of-the-art baselines. When integrated with an online product support system, our approach led to improvements in retrieval performance and customer return rates. Our work not only advances the state of text chunking but also provides a standardized framework for evaluating chunking strategies, addressing a critical gap in the field.

Exploration, the act of broadening user experiences beyond their established preferences, is challenging in large-scale recommendation systems due to feedback loops and limited signals on user exploration patterns. Large Language Models (LLMs) offer potential solutions by leveraging their world knowledge to recommend novel content outside these loops. A key challenge is aligning LLMs with user preferences while preserving their knowledge and reasoning. To enhance planning for new user interests using LLMs, this paper introduces a novel approach that combines hierarchical planning with LLM inference-time scaling. This method aims to improve recommendation relevancy without compromising novelty. We decouple novelty and user-alignment, training separate LLMs for each objective. We then scale up the novelty-focused LLM’s inference and select the best-of-n predictions using the user-aligned LLM. Live experiments demonstrate efficacy, showing significant gains in both user satisfaction (measured by watch activity and active user counts) and exploration diversity.

Large Language Models (LLMs) enable natural language to SQL conversion, allowing users to query databases without SQL expertise. However, generating accurate, efficient queries is challenging due to ambiguous intent, domain knowledge requirements, and database constraints. Extensive reasoning improves SQL quality but increases computational costs and latency. We propose SQLGenie, a practical system for reliable SQL generation. It consists of three components: (1) Table Onboarder, which analyzes new tables, optimizes indexing, partitions data, identifies foreign key relationships, and stores schema details for SQL generation; (2) SQL Generator, an LLM-based system producing accurate SQL; and (3) Feedback Augmentation, which filters correct query-SQL pairs, leverages multiple LLM agents for complex SQL, and stores verified examples. SQLGenie achieves state-of-the-art performance on public benchmarks (92.8% execution accuracy on WikiSQL, 82.1% of Spider, 73.8% on BIRD) and internal datasets, surpassing the best single-LLM baseline by 21.5% and the strongest pipeline competitor by 5.3%. Its hybrid variant optimally balances accuracy and efficiency, reducing generation time by 64% compared to traditional multi-LLM approaches while maintaining competitive accuracy.

Enterprise search systems often struggle to retrieve accurate, domain-specific information due to semantic mismatches and overlapping terminologies. These issues can degrade the performance of downstream applications such as knowledge management, customer support, and retrieval-augmented generation agents. To address this challenge, we propose a scalable hard-negative mining framework tailored specifically for domain-specific enterprise data. Our approach dynamically selects semantically challenging but contextually irrelevant documents to enhance deployed re-ranking models.Our method integrates diverse embedding models, performs dimensionality reduction, and uniquely selects hard negatives, ensuring computational efficiency and semantic precision. Evaluation on our proprietary enterprise corpus (cloud services domain) demonstrates substantial improvements of 15% in MRR@3 and 19% in MRR@10 compared to state-of-the-art baselines and other negative sampling techniques. Further validation on public domain-specific datasets (FiQA, Climate Fever, TechQA) confirms our method’s generalizability and readiness for real-world applications.

Determining company similarity is a vital task in finance, underpinning risk management, hedging, and portfolio diversification. Practitioners often rely on sector and industry classifications such as SIC and GICS codes to gauge similarity, the former being used by the U.S. Securities and Exchange Commission (SEC), and the latter widely used by the investment community. Since these classifications lack granularity and need regular updating, using clusters of embeddings of company descriptions has been proposed as a potential alternative, but the lack of interpretability in token embeddings poses a significant barrier to adoption in high-stakes contexts. Sparse Autoencoders (SAEs) have shown promise in enhancing the interpretability of Large Language Models (LLMs) by decomposing Large Language Model (LLM) activations into interpretable features. Moreover, SAEs capture an LLM’s internal representation of a company description, as opposed to semantic similarity alone, as is the case with embeddings. We apply SAEs to company descriptions, and obtain meaningful clusters of equities. We benchmark SAE features against SIC-codes, Industry codes, and Embeddings. Our results demonstrate that SAE features surpass sector classifications and embeddings in capturing fundamental company characteristics. This is evidenced by their superior performance in correlating logged monthly returns – a proxy for similarity – and generating higher Sharpe ratios in co-integration trading strategies, which underscores deeper fundamental similarities among companies. Finally, we verify the interpretability of our clusters, and demonstrate that sparse features form simple and interpretable explanations for our clusters.

We present the e-Llama models: 8 billion and 70 billion parameter large language models that are adapted towards the e-commerce domain.These models are meant as foundation models with deep knowledge about e-commerce, that form a base for instruction- and fine-tuning.The e-Llama models are obtained by continuously pretraining the Llama 3.1 base models on 1 trillion tokens of domain-specific data.We discuss our approach and motivate our choice of hyperparameters with a series of ablation studies.To quantify how well the models have been adapted to the e-commerce domain, we define and implement a set of multilingual, e-commerce specific evaluation tasks.We show that, when carefully choosing the training setup, the Llama 3.1 models can be adapted towards the new domain without sacrificing significant performance on general domain tasks.We also explore the possibility of merging the adapted model and the base model for a better control of the performance trade-off between domains.

Large Language Models (LLMs) are increasingly deployed as computer-use agents, autonomously performing tasks within real desktop or web environments. While this evolution greatly expands practical use cases for humans, it also creates serious security exposures. We present SUDO (Screen-based Universal Detox2tox Offense), a novel attack framework that systematically bypasses refusal-trained safeguards in commercial computer-use agents, such as Claude for Computer Use. The core mechanism, Detox2tox, transforms harmful requests (that agents initially reject) into seemingly benign requests via detoxification, secures detailed instructions from advanced vision language models (VLMs), and then reintroduces malicious content via toxification just before execution. Unlike conventional jailbreaks, SUDO iteratively refines its attacks based on a built-in refusal feedback, making it increasingly effective against robust policy filters. In extensive tests spanning 50 real-world tasks and multiple state-of-the-art VLMs, SUDO achieves a stark attack success rate of 24.41% (with no refinement), and up to 41.33% (by its iterative refinement) in Claude for Computer Use. By revealing these vulnerabilities and demonstrating the ease with which they can be exploited in real-world computing environments, this paper highlights an immediate need for robust, context-aware safeguards. WARNING: This paper includes harmful or offensive model outputs.

Despite recent success in applying large language models (LLMs) to electronic health records (EHR), most systems focus primarily on assessment rather than treatment planning. We identify three critical limitations in current approaches: they generate treatment plans in a single pass rather than following the sequential reasoning process used by clinicians; they rarely incorporate patient-specific historical context; and they fail to effectively distinguish between subjective and objective clinical information. Motivated by the SOAP methodology (Subjective, Objective, Assessment, Plan), we introduce MedPlan, a novel framework that structures LLM reasoning to align with real-life clinician workflows. Our approach employs a two-stage architecture that first generates a clinical assessment based on patient symptoms and objective data, then formulates a structured treatment plan informed by this assessment and enriched with patient-specific information through retrieval-augmented generation. Comprehensive evaluation demonstrates that our method significantly outperforms baseline approaches in both assessment accuracy and treatment plan quality. Our demo system and code are available at https://github.com/JustinHsu1019/MedPlan.

Fine-tuning large language models (LLMs) for specific tasks requires diverse, high-quality training data. However, obtaining sufficient relevant data remains a significant challenge. Existing data synthesis methods either depend on extensive seed datasets or struggle to balance task relevance and data diversity. To address these challenges, we propose Attribute-guided multI-hop Data Expansion (AIDE), a novel data synthesis framework that uses a multi-hop process to expand very few seed data points while ensuring data diversity and task relevance. AIDE extracts the main topic and key knowledge attributes from the seeds to guide the synthesis steps. The process repeats for K hops, using the generated data as seeds. To prevent irrelevant data generation as the hop depth increases, AIDE incorporates a residual connection mechanism. Our empirical results show that AIDE enables fine-tuning of Mistral-7B, Llama-3.1-8B and Llama-3.2-3B from 10 seeds, surpassing the models fine-tuned on human curated data. Furthermore, AIDE outperforms state-of-the-art data synthesis methods, such as Evol-Instruct, by over 30% in task-specific fine-tuning. Code is available at https://github.com/Code4Graph/AIDE.

Error correction is an important capability when applying large language models (LLMs) to facilitate user typing on mobile devices. In this paper, we use LLMs to synthesize a high-quality dataset of error correction pairs to evaluate and improve LLMs for mobile applications. We first prompt LLMs with error correction domain knowledge to build a scalable and reliable addition to the existing data synthesis pipeline. We then adapt the synthetic data distribution to match the mobile application domain by reweighting the samples. The reweighting model is learnt by predicting (a handful of) live A/B test metrics when deploying LLMs in production, given the LLM performance on offline evaluation data and scores from a small privacy-preserving on-device language model.Finally, we present best practices for mixing our synthetic data with other data sources to improve model performance on error correction in both offline evaluation and production live A/B testing.

Multilingual automatic speech recognition (ASR) in the medical domain serves as a foundational task for various downstream applications such as speech translation, spoken language understanding, and voice-activated assistants. This technology improves patient care by enabling efficient communication across language barriers, alleviating specialized workforce shortages, and facilitating improved diagnosis and treatment, particularly during pandemics. In this work, we introduce MultiMed, the first multilingual medical ASR dataset, along with the first collection of small-to-large end-to-end medical ASR models, spanning five languages: Vietnamese, English, German, French, and Mandarin Chinese. To our best knowledge, MultiMed stands as the world’s largest medical ASR dataset across all major benchmarks: total duration, number of recording conditions, number of accents, and number of speaking roles. Furthermore, we present the first multilinguality study for medical ASR, which includes reproducible empirical baselines, a monolinguality-multilinguality analysis, Attention Encoder Decoder (AED) vs Hybrid comparative study and a linguistic analysis. We present practical ASR end-to-end training schemes optimized for a fixed number of trainable parameters that are common in industry settings. All code, data, and models are available online.

Attribute value extraction plays a crucial role in enhancing e-commerce search, filtering, and recommendation systems. However, prior visual attribute value extraction methods typically rely on both product images and textual information such as product descriptions and titles. In practice, text can be ambiguous, inaccurate, or unavailable, which can degrade model performance. We propose Mixture of Image Captioning Experts (MICE), a novel augmentation framework for product attribute value extraction. MICE leverages a curated pool of image captioning models to generate accurate captions from product images, resulting in robust attribute extraction solely from an image. Extensive experiments on the public ImplicitAVE dataset and a proprietary women’s tops dataset demonstrate that MICE significantly improves the performance of state-of-the-art large multimodal models (LMMs) in both zero-shot and fine-tuning settings. An ablation study validates the contribution of each component in the framework. MICE’s modular design offers scalability and adaptability, making it well-suited for diverse industrial applications with varying computational and latency requirements.

In this work, we present the first open leaderboard for evaluating Korean large language models focused on finance. Operated for abouteight weeks, the leaderboard evaluated 1,119 submissions on a closed benchmark covering five MCQA categories: finance and accounting, stock price prediction, domestic company analysis, financial markets, and financial agent tasks and one open-ended qa task. Building on insights from these evaluations, we release an open instruction dataset of 80k instances and summarize widely used training strategies observed among top-performing models. Finally, we introduce FINKRX, a fully open and transparent LLM built using these best practices. We hope our contributions help advance the development of better and safer financial LLMs for Korean and other languages.

Transparency in AI healthcare decision-makingis crucial. By incorporating rationales to explain reason for each predicted label, userscould understand Large Language Models(LLMs)’s reasoning to make better decision.In this work, we introduce a new task - Sentiment Reasoning - for both speech and textmodalities, and our proposed multimodal multitask framework and the world’s largest multimodal sentiment analysis dataset. Sentiment Reasoning is an auxiliary task in sentiment analysis where the model predicts boththe sentiment label and generates the rationale behind it based on the input transcript.Our study conducted on both human transcriptsand Automatic Speech Recognition (ASR) transcripts shows that Sentiment Reasoning helpsimprove model transparency by providing rationale for model prediction with quality semantically comparable to humans while alsoimproving model’s classification performance(+2% increase in both accuracy and macro-F1) via rationale-augmented fine-tuning. Also,no significant difference in the semantic quality of generated rationales between human andASR transcripts. All code, data (five languages - Vietnamese, English, Chinese, German, andFrench) and models are published online.

Charts are ubiquitous as they help people understand and reason with data. Recently, various downstream tasks, such as chart question answering, chart2text, and fact-checking, have emerged. Large Vision-Language Models (LVLMs) show promise in tackling these tasks, but their evaluation is costly and time-consuming, limiting real-world deployment. While using LVLMs as judges to assess chart comprehension capabilities of other LVLMs could streamline evaluation processes, challenges like proprietary datasets, restricted access to powerful models, and evaluation costs hinder their adoption in industrial settings. To this end, we present a comprehensive evaluation of 13 open-source LVLMs as judges for diverse chart comprehension and reasoning tasks. We design both pairwise and pointwise evaluation tasks covering criteria like factual correctness, informativeness, and relevancy. Additionally, we analyze LVLM judges based on format adherence, positional consistency, length bias, and instruction-following. We focus on cost-effective LVLMs (<10B parameters) suitable for both research and commercial use, following a standardized evaluation protocol and rubric to measure the LVLM judge accuracy. Experimental results reveal notable variability: while some open LVLM judges achieve GPT-4-level evaluation performance (about 80% agreement with GPT-4 judgments), others struggle (below ~10% agreement). Our findings highlight that state-of-the-art open-source LVLMs can serve as cost-effective automatic evaluators for chart-related tasks, though biases such as positional preference and length bias persist.

Occupational Health (OH) triage is a systematic process for evaluating and prioritising workplace health concerns to determine appropriate care and interventions. This research addresses critical triage challenges through our novel AI agent orchestration framework, OccuTriage, developed in collaboration with Healthcare Provider. Our framework simulates healthcare professionals’ reasoning using specialized LLM agents, retrieval augmentation with domain-specific knowledge, and a bidirectional decision architecture. Experimental evaluation on 2,589 OH cases demonstrates OccuTriage outperforms single-agent approaches with a 20.16% average discordance rate compared to baseline rates of 43.05%, while matching or exceeding human expert performance (25.11%). The system excels in reducing under-triage rates, achieving 9.84% and 3.1% for appointment and assessor type decisions respectively. These results establish OccuTriage’s efficacy in performing complex OH triage while maintaining safety and optimizing resource allocation.

With the release of R1, a publicly available large reasoning model (LRM), researchers commonly train new LRMs by training language models on R1’s long chain-of-thought (CoT) inferences. While prior works show that LRMs’ capabilities can be reproduced through direct distillation, the continued reliance on the existing models (e.g., R1) remains a critical limitation in advancing the field.As a first step toward independent LRM development, this paper explores the possibility of constructing a long CoT dataset with LLMs that are not trained for inference-time scaling.To this end, we present the Long CoT Collection, a dataset of 100K CoT rationales annotated using existing short CoT LLMs. We develop a pipeline that induces o1’s novel reasoning strategies into short CoT LLMs, enabling them to think longer and introducing controllability over the thought budget to better manage the overthinking problem.Our extensive analyses validate that our dataset achieves quality comparable to—or slightly below—R1. Furthermore, our experiments demonstrate that training on our dataset not only strengthens general reasoning skills, but also provides a strong foundation for reinforcement learning—models initialized on our data achieve 2-3x larger gains with RLVR. We make the codes, datasets, and models publicly available at LINK.

The integration of generative artificial intelligence into educational applications has enhanced personalized and interactive learning experiences, and it shows strong potential to promote young learners language acquisition. However, it is still challenging to ensure consistent and robust performance across different languages and cultural contexts, and kids-friendly design requires simplified instructions, engaging interactions, and age-appropriate scaffolding to maintain motivation and optimize learning outcomes.In this work, we introduce SingaKids, a dialogic tutor designed to facilitate language learning through picture description tasks. Our system integrates dense image captioning, multilingual dialogic interaction, speech understanding, and engaging speech generation to create an immersive learning environment in four languages: English, Mandarin, Malay, and Tamil. We further improve the system through multilingual pre-training, task-specific tuning, and scaffolding optimization. Empirical studies with elementary school students demonstrate that SingaKids provides effective dialogic teaching, benefiting learners at different performance levels.

There has been increasing interest in unifying streaming and non-streaming automatic speech recognition (ASR) models to reduce development, training, and deployment costs. We present a unified framework that trains a single end-to-end ASR model for both streaming and non-streaming applications, leveraging future context information. We propose to use dynamic right-context through the chunked attention masking in the training of zipformer-based ASR models. We demonstrate that using right-context is more effective in zipformer models compared to other conformer models due to its multi-scale nature. We analyze the effect of varying the number of right-context frames on accuracy and latency of the streaming ASR models. We use Librispeech and large in-house conversational datasets to train different versions of streaming and non-streaming models and evaluate them in a production grade server-client setup across diverse testsets of different domains. The proposed strategy reduces word error by relative 7.9% with a small degradation in user-perceived latency. By adding more right-context frames, we are able to achieve streaming performance close to that of non-streaming models. Our approach also allows flexible control of the latency-accuracy tradeoff according to customers requirements.

Query classification, including multiple subtasks such as intent and category prediction, is a vital part of e-commerce applications. E-commerce queries are usually short and lack context, and the information between labels cannot be used, resulting in insufficient prior information for modeling. Most existing industrial query classification methods rely on users’ posterior click behavior to construct training samples, resulting in a Matthew vicious cycle. Furthermore, the subtasks of query classification lack a unified framework, leading to low efficiency for algorithm improvement.In this paper, we propose a novel Semi-supervised Scalable Unified Framework (SSUF), containing multiple enhanced modules to unify the query classification tasks. The knowledge-enhanced module uses world knowledge to enhance query representations and solve the problem of insufficient query information. The label-enhanced module uses label semantics and semi-supervised signals to reduce the dependence on posterior labels. The structure-enhanced module enhances the label representation based on the complex label relations. Each module is highly pluggable, and input features can be added or removed as needed according to each subtask. We conduct extensive offline and online A/B experiments, and the results show that SSUF significantly outperforms the state-of-the-art models.

With the rapid advancement of Large Language Models (LLMs), the demand for robust instruction-following capabilities in code generation tasks has grown significantly. Code generation not only facilitates faster prototyping and automated testing, but also augments developer efficiency through improved maintainability and reusability of code. In this paper, we introduce CodeIF, the first benchmark specifically designed to assess the abilities of LLMs to adhere to task-oriented instructions within diverse code generation scenarios. CodeIF encompasses a broad range of tasks, including function synthesis, error debugging, algorithmic refactoring, and code explanation, thereby providing a comprehensive suite to evaluate model performance across varying complexity levels and programming domains. We conduct extensive experiments with LLMs, analyzing their strengths and limitations in meeting the demands of these tasks. The experimental results offer valuable insights into how well current models align with human instructions, as well as the extent to which they can generate consistent, maintainable, and contextually relevant code. Our findings not only underscore the critical role that instruction-following LLMs can play in modern software development, but also illuminate pathways for future research aimed at enhancing their adaptability, reliability, and overall effectiveness in automated code generation.

A comprehensive benchmark is crucial for evaluating automated Business Intelligence (BI) systems and their real-world effectiveness. We propose BI-Bench, a holistic, end-to-end benchmarking framework that assesses BI systems based on the quality, relevance, and depth of insights. It categorizes queries into descriptive, diagnostic, predictive, and prescriptive types, aligning with practical BI needs. Our fully automated approach enables custom benchmark generation tailored to specific datasets. Additionally, we introduce an automated evaluation mechanism within BI-Bench that removes reliance on strict ground truth, ensuring scalable and adaptable assessments. By addressing key limitations, it offers a flexible and robust, user-centered methodology for advancing next-generation BI systems.

Accurate mapping of queries to product categories is crucial for efficient retrieval and ranking of relevant products in e-commerce search. Conventionally, such query classification models rely on supervised learning using historical user interactions, but their effectiveness diminishes in cold-start scenarios, where new categories or products lack sufficient training data. This results in poor query-to-category mappings, negatively affecting retrieval and ranking. Synthetic query generation has emerged as a promising solution by augmenting training data; however, existing methods do not incorporate feedback from the query relevance model, limiting their ability to generate queries that enhance product retrieval. To address this, we propose an adversarial reinforcement learning framework that optimizes an LLM-based generator to expose weaknesses in query classification models. The generator produces synthetic queries to augment the classifier’s training set, ultimately improving its performance. Additionally, we introduce a structured reward signal to ensure stable training. Experiments on public datasets show an average PR-AUC improvement of +1.82% on benchmarks and +3.26% on a proprietary dataset, demonstrating the framework’s effectiveness in enhancing query classification and mitigating cold-start challenges.

Automating benefit verification phone calls saves time in healthcare and helps patients receive treatment faster. It is critical to obtain highly accurate information in these phone calls, as it can affect a patient’s healthcare journey. Given the noise in phone call transcripts, we have a two-stage system that involves a post-call review phase for potentially noisy fields, where human reviewers manually verify the extracted data—a labor-intensive task. To automate this stage, we introduce Auto Review, which significantly reduces manual effort while maintaining a high bar for accuracy. This system, being highly reliant on call transcripts, suffers a performance bottleneck due to automatic speech recognition (ASR) issues. This problem is further exacerbated by the use of domain-specific jargon in the calls. In this work, we propose a second-stage postprocessing pipeline for accurate information extraction. We improve accuracy by using multiple ASR alternatives and a pseudo-labeling approach that does not require manually corrected transcripts. Experiments with general-purpose large language models and feature-based model pipelines demonstrate substantial improvements in the quality of corrected call transcripts, thereby enhancing the efficiency of Auto Review.

In-car AI assistants enhance driving by enabling hands-free interactions, yet they often struggle with multi-turn conversations and fail to handle cognitively complex follow-up questions. This limits their effectiveness in real-world deployment. To address this limitation, we propose a framework that leverages Bloom’s Taxonomy to systematically generate follow-up questions with increasing cognitive complexity and a Gricean-inspired evaluation framework to assess their Logical Consistency, Informativeness, Relevance, and Clarity. We introduce a dataset comprising 750 human-annotated seed questions and 3750 follow-up questions, with human evaluation confirming that 96.68% of the generated questions adhere to the intended Bloom’s Taxonomy levels. Our approach, validated through both LLM-based and human assessments, also identifies the specific cognitive complexity level at which in-car AI assistants begin to falter information that can help developers measure and optimize key cognitive aspects of conversational performance.

The development of large language models (LLMs) offers a feasible approach to simulating complex behavioral patterns of individuals, enabling the reconstruction of microscopic and realistic human societal dynamics. However, this approach demands a realistic environment to provide feedback for the evolving of agents, as well as a parallelized framework to support the massive and uncertain interactions among agents and environments. To address the gaps in existing works, which lack real-world environments and struggle with complex interactions, we design a scalable framework named **AgentSociety**, which integrates realistic societal environments and parallelized interactions to support simulations of large-scale agents. Experiments demonstrate that the framework can support simulations of 30,000 agents that are faster than the wall-clock time with 24 NVIDIA A800 GPUs and the performance grows linearly with the increase of LLM computational resources. We also show that the integration of realistic environments significantly enhances the authenticity of the agents’ behaviors. Through the framework and experimental results, we are confident that deploying large-scale LLM Agents to simulate human societies becomes feasible. This will help practitioners in fields such as social sciences and management sciences to obtain new scientific discoveries via language generation technologies, and even improve planning and decision-making in the real world. The code is available at https://github.com/tsinghua-fib-lab/agentsociety/.

Recent advances in large language models (LLMs) have drawn attention for their potential to automate and optimize processes across various sectors.However, the adoption of LLMs in the plant construction industry remains limited, mainly due to its highly specialized nature and the lack of resources for domain-specific training and evaluation.In this work, we propose ENGinius, the first LLM designed for plant construction engineering.We present procedures for data construction and model training, along with the first benchmarks tailored to this underrepresented domain.We show that ENGinius delivers optimized responses to plant engineers by leveraging enriched domain knowledge.We also demonstrate its practical impact and use cases, such as technical document processing and multilingual communication.

Modern digital platforms rely on related search query recommendations to enhance engagement, yet existing methods fail to reconcile click-through rate (CTR) optimization with topic expansion. We propose **CMAQ**, a **C**onsistent **M**ulti-Objective **A**ligned **Q**uery generation framework that harmonizes these goals through three components: (1) reward modeling to quantify objectives, (2) style alignment for format compliance, and (3) consistency-aware optimization to coordinate joint improvements. CMAQ employs adaptive 𝛽-scaled DPO with geometric mean rewards, balancing CTR and expansion while mitigating objective conflicts. Extensive offline and online evaluations in a large-scale industrial setting demonstrate CMAQ’s superiority, achieving significant CTR gains (+2.3%) and higher human-rated query quality compared to state-of-the-art methods. Our approach enables high-quality query generation while sustaining user engagement and platform ecosystem health.

Generating high-quality geometry problems is both an important and challenging task in education. Compared to math word problems, geometry problems further emphasize multi-modal formats and the translation between informal and formal languages. In this paper, we introduce a novel task for geometry problem generation and propose a new pipeline method: the Symbolic Deduction Engine-based Geometry Problem Generation framework (SDE-GPG). The framework leverages a symbolic deduction engine and contains four main steps: (1) searching a predefined mapping table from knowledge points to extended definitions, (2) sampling extended definitions and performing symbolic deduction, (3) filtering out unqualified problems, and (4) generating textual problems and diagrams. Specifically, our method supports to avoid inherent biases in translating natural language into formal language by designing the mapping table, and guarantees to control the generated problems in terms of knowledge points and difficulties by an elaborate checking function. With obtained formal problems, they are translated to natural language and the accompanying diagrams are automatically drew by rule-based methods. We conduct experiments using real-world combinations of knowledge points from two public datasets. The results demonstrate that the SDE-GPG can effectively generate readable, solvable and controllable geometry problems.

This paper introduces a task aimed at extracting structured tables from text using natural language (NL) instructions. We present TableCoder, an approach that leverages the symbolic nature of code to enhance the robustness of table structure construction and content extraction. TableCoder first generates Python classes or SQL statements to explicitly construct table structures, capturing semantic ontology, computational dependencies, numerical properties, and format strings. This approach reliably mitigates issues such as structural errors, erroneous computations, and mismatched value types. Subsequently, TableCoder proposes grounded content extraction, populating table cells sequentially and maintaining the exact order in which they are mentioned in the source text. By simulating a grounded “translation” from text to code, this method reduces the likelihood of omissions and hallucinations.Experimental results demonstrate that TableCoder significantly improves F1 scores and mitigates hallucination and computational errors, crucial for high-stakes applications like government data analytics and financial compliance reporting. Moreover, the code-generation-based method naturally integrates with standard SQL databases and Python workflows, ensuring seamless deployment in existing enterprise data pipelines.

Due to the legal and ethical responsibilities of healthcare providers (HCPs) for accurate documentation and protection of patient data privacy, the natural variability in the responses of large language models (LLMs) presents challenges for incorporating clinical note generation (CNG) systems, driven by LLMs, into real-world clinical processes. The complexity is further amplified by the detailed nature of texts in CNG. To enhance the confidence of HCPs in tools powered by LLMs, this study evaluates the reliability of 12 open-weight and proprietary LLMs from Anthropic, Meta, Mistral, and OpenAI in CNG in terms of their ability to generate notes that are string equivalent (consistency rate), have the same meaning (semantic consistency) and are correct (semantic similarity), across several iterations using the same prompt. The results show that (1) LLMs from all model families are stable, such that their responses are semantically consistent despite being written in various ways, and (2) most of the LLMs generated notes close to the corresponding notes made by experts. Overall, Meta’s Llama 70B was the most reliable, followed by Mistral’s Small model. With these findings, we recommend the local deployment of these relatively smaller open-weight models for CNG to ensure compliance with data privacy regulations, as well as to improve the efficiency of HCPs in clinical documentation.

Recent advances in large language models (LLMs) have significantly improved Document AI, demonstrating remarkable performance on document understanding tasks such as question answering. However, existing approaches primarily focus on solving specific tasks, lacking the capability to structurally organize and systematically manage document information. To address this limitation, we propose Revise, a framework that systematically corrects errors introduced by OCR at the character, word, and structural levels. Specifically, Revise employs a comprehensive hierarchical taxonomy of common OCR errors and a synthetic data generation strategy that realistically simulates such errors to train an effective correction model. Experimental results demonstrate that Revise effectively corrects OCR outputs, enabling more structured representation and systematic management of document contents. Consequently, our method significantly enhances downstream performance in document retrieval and question answering tasks, highlighting the potential to overcome the structural management limitations of existing Document AI frameworks.

The key-value (KV) cache in transformer models is a critical component for efficient decoding or inference, yet its memory demands scale poorly with sequence length, posing a major challenge for scalable deployment of large language models. Among several approaches to KV cache compression, quantization of key and value activations has been widely explored. Most KV cache quantization methods still need to manage sparse and noncontiguous outliers separately. To address this, we introduce TaDA, a training-free recipe for KV cache compression with quantization precision that adapts to error sensitivity across layers and a mean centering to eliminate separate outlier handling. Our approach yields substantial accuracy improvements for multiple models supporting various context lengths. Moreover, our approach does not need to separately manage outlier elements—a persistent hurdle in most traditional quantization methods. Experiments on standard benchmarks demonstrate that our technique reduces KV cache memory footprint to 27% of the original 16-bit baseline while achieving comparable accuracy. Our method paves the way for scalable and high-performance reasoning in language models by potentially enabling inference for longer context length models, reasoning models, and longer chain of thoughts.

Emotional support conversation (ESC) aims to alleviate the emotional distress of individuals through effective conversations. Although large language models (LLMs) have obtained remarkable progress on ESC, most of these studies might not define the diagram from the state model perspective, therefore providing a suboptimal solution for long-term satisfaction. To address such an issue, we leverage the Q-learning on LLMs, and propose a framework called straQ*. Our framework allows a plug-and-play LLM to bootstrap the planning during ESC, determine the optimal strategy based on long-term returns, and finally guide the LLM to response. Substantial experiments on ESC datasets suggest that straQ* outperforms many baselines, including direct inference, self-refine, chain of thought, finetuning, and finite state machines.

The rapid advancement of generative AI technologies is driving the integration of diverse AI-powered services into smartphones, transforming how users interact with their devices. To simplify access to predefined AI services, this paper introduces MIRA, a pioneering framework for task instruction recommendation that enables intuitive one-touch AI tasking on smartphones. With MIRA, users can long-press on images or text objects to receive contextually relevant instruction recommendations for executing AI tasks. Our work introduces three key innovations: 1) A multimodal large language model (MLLM)-based recommendation pipeline with structured reasoning to extract key entities, infer user intent, and generate precise instructions; 2) A template-augmented reasoning mechanism that integrates high-level reasoning templates, enhancing task inference accuracy; 3) A prefix-tree-based constrained decoding strategy that restricts outputs to predefined instruction candidates, ensuring coherence and intent alignment. Through evaluation using a real-world annotated datasets and a user study, MIRA has demonstrated substantial improvements in recommendation accuracy. The encouraging results highlight MIRA’s potential to revolutionize the way users engage with AI services on their smartphones, offering a more seamless and efficient experience.

System-level programming is essential for modern enterprise infrastructure, enabling the automation and management of complex systems through declarative code. Developers write this code based on schemas, which themselves are a form of code that defines constraints like data types and required fields. These schemas help ensure operational correctness and smooth integration across systems. However, as enterprise schemas become complex, manually writing code adhering to these constraints becomes challenging for developers. Large Language Models (LLMs) have demonstrated potential in code generation and natural language understanding, particularly in zero-shot and few-shot settings. However, applying LLMs to handle constraints represented in code, essential for system-level programming rather than natural language, has not been explored. Hence, we introduce ConCodeEval, a study across two key dimensions: format and constraint efficacy, with a first-of-its-kind benchmark involving two novel experiments for code constraints across five representations (JSON, YAML, XML, Python, and natural language). Our findings suggest that conscious choice of representations can lead to optimal use of LLMs in enterprise use cases involving constraints. Nonetheless, LLMs continue to struggle significantly with code constraints, motivating the need for innovation in this direction.

Consumers often face inconsistent product quality, particularly when identical products vary between markets, a situation known as the dual quality problem. To identify and address this issue, automated techniques are needed. This paper explores how natural language processing (NLP) can aid in detecting such discrepancies and presents the full process of developing a solution. First, we describe in detail the creation of a new Polish-language dataset with 1,957 reviews, 540 highlighting dual quality issues. We then discuss experiments with various approaches like SetFit with sentence-transformers, transformer-based encoders, and LLMs, including error analysis and robustness verification. Additionally, we evaluate multilingual transfer using a subset of opinions in English, French, and German. The paper concludes with insights on deployment and practical applications.

A key ethical challenge in Automated Essay Scoring (AES) is ensuring that scores are only released when they meet high reliability standards. Confidence modelling addresses this by assigning a reliability estimate measure, in the form of a confidence score, to each automated score. In this study, we frame confidence estimation as a classification task: predicting whether an AES-generated score correctly places a candidate in the appropriate CEFR level. While this is a binary decision, we leverage the inherent granularity of the scoring domain in two ways. First, we reformulate the task as an n-ary classification problem using score binning. Second, we introduce a set of novel Kernel Weighted Ordinal Categorical Cross Entropy (KWOCCE) loss functions that incorporate the ordinal structure of CEFR labels. Our best-performing model achieves an F1 score of 0.97, and enables the system to release 47% of scores with 100% CEFR agreement and 99% with at least 95% CEFR agreement — compared to ≈ 92 % CEFR agreement from the standalone AES model where we release all AM predicted scores.

The advancement of Large Language Models (LLMs) has led to significant improvements in various service domains, including search, recommendation, and chatbot applications.However, applying state-of-the-art (SOTA) research to industrial settings presents challenges, as it requires maintaining flexible conversational abilities while also strictly complying with service-specific constraints.This can be seen as two conflicting requirements due to the probabilistic nature of LLMs.In this paper, we propose our approach to addressing this challenge and detail the strategies we employed to overcome their inherent limitations in real-world applications.We conduct a practical case study of a conversational agent designed for the e-commerce domain, detailing our implementation workflow and optimizations.Our findings provide insights into bridging the gap between academic research and real-world application, introducing a framework for developing scalable, controllable, and reliable AI-driven agents.

Document question answering plays a crucial role in enhancing employee productivity by providing quick and accurate access to information. Two primary approaches have been developed: retrieval-augmented generation (RAG), which reduces input tokens and inference costs, and long-context question answering (LC), which processes entire documents for higher accuracy. We introduce EXPLAIN (EXtracting, Pre-summarizing, Linking and enhAcINg RAG), a novel retrieval-augmented generation method that automatically extracts useful entities and generates summaries from documents. EXPLAIN improves accuracy by retrieving more informative entity summaries, achieving precision comparable to LC while maintaining low token consumption. Experimental results on internal dataset (ROUGE-L from 30.14% to 30.31%) and three public datasets (HotpotQA, 2WikiMQA, and Quality, average score from 62% to 64%) demonstrate the efficacy of EXPLAIN. Human evaluation in ant group production deployment indicates EXPLAIN surpasses baseline RAG in comprehensiveness.

Enterprises are increasingly adopting Generative AI applications to extract insights from large volumes of multimodal documents in domains such as finance, law, healthcare, and industry. These documents contain structured and unstructured data (images, charts, handwritten texts, etc.) requiring robust AI systems for effective retrieval and comprehension. Recent advancements in Retrieval-Augmented Generation (RAG) frameworks and Vision-Language Models (VLMs) have improved retrieval performance on multimodal documents by processing pages as images. However, large-scale deployment remains challenging due to the high cost of LLM API usage and the slower inference speed of image-based processing of pages compared to text-based processing. To address these challenges, we propose EcoDoc, a cost-effective multimodal document processing system that dynamically selects the processing modalities for each page as an image or text based on page characteristics and query intent. Our experimental evaluation on TAT-DQA and DocVQA benchmarks shows that EcoDoc reduces average query processing latency by up to 2.29× and cost by up to 10×, without compromising accuracy.