Binaural audio enriches immersive experiences by enabling the perception of the spatial locations of sounding objects in AR, VR, and embodied AI applications. While existing audio spatialization methods can generally map any available monaural audio to binaural audio signals, they often lack the flexible and interactive control needed in complex multi-object user-interactive environments. To address this, we propose a Text-guided Audio Spatialization (TAS) framework that utilizes diverse text prompts and evaluates our model from unified generation and comprehension perspectives. Due to the limited availability of high-quality, large-scale stereo data, we construct the SpatialTAS dataset, which encompasses 376,000 simulated binaural audio samples to facilitate the training of our model. Our model learns binaural differences guided by 3D spatial location and relative position prompts, enhanced with flipped-channel audio. Experimental results show that our model can generate high quality binaural audios for various audio types on both simulated and real-recorded datasets. Besides, we establish an assessment model based on Llama-3.1-8B, which evaluates the semantic accuracy of spatial locations through a spatial reasoning task. Results demonstrate that by utilizing text prompts for flexible and interactive control, we can generate binaural audio with both high quality and semantic consistency in spatial locations.

Autonomous agents have become increasingly important for interacting with the real world. Android agents, in particular, have been a frequently-mentioned interaction method. However, existing studies for training and evaluating Android agents lack systematic research on both open-source and closed-source models. In this work, we propose AndroidLab as a systematic Android agent framework. It includes an operation environment with different modalities, action space, and a reproducible benchmark. It supports both large language models (LLMs) and multimodal models (LMMs) in the same action space. AndroidLab benchmark includes predefined Android virtual devices and 138 tasks across nine apps built on these devices. By using the AndroidLab environment, we develop an Android Instruction dataset and train six open-source LLMs and LMMs, lifting the average success rates from 4.59% to 21.50% for LLMs and from 1.93% to 13.28% for LMMs. AndroidLab is open-sourced and publicly available at https://github.com/THUDM/Android-Lab.

Large language models have opened up a world of possibilities for various NLP tasks, sparking optimism for the future. Despite their potential, LLMs have yet to be widely used as agents on real mobile devices. The main challenge is the need for high-quality data sources. Time constraints and labor intensity often hinder human annotation. On the other hand, existing LLMs exhibit inadequate completion rates and need a robust data filtration strategy. Given these challenges, we develop a framework called AndroidGen to enhance the capabilities of LLM-based agents under data scarcity. In addition, we leverage AndroidGen to collect trajectories given human tasks and train open-source LLMs on these trajectories to develop an open-source mobile agent without manually labeled trajectories. We extensively evaluate AndroidGen with AndroidWorld, AitW, and various popular applications, demonstrating its improvements and revealing potential areas for future improvement. Code, model, and data are available at https://github.com/THUDM/AndroidGen.

Large language models (LLMs) have achieved remarkable proficiency in understanding and generating human natural languages, mainly owing to the “scaling law” that optimizes relationships among language modeling loss, model parameters, and pre-trained tokens. However, with the exhaustion of high-quality internet corpora and increasing computational demands, the sustainability of pre-training scaling needs to be addressed. This paper presents a comprehensive survey of post-training scaling, an emergent paradigm aiming to relieve the limitations of traditional pre-training by focusing on the alignment phase, which traditionally accounts for a minor fraction of the total training computation. Our survey categorizes post-training scaling into three key methodologies: Supervised Fine-tuning (SFT), Reinforcement Learning from Feedback (RLxF), and Test-time Compute (TTC). We provide an in-depth analysis of the motivation behind post-training scaling, the scalable variants of these methodologies, and a comparative discussion against traditional approaches. By examining the latest advancements, identifying promising application scenarios, and highlighting unresolved issues, we seek a coherent understanding and map future research trajectories in the landscape of post-training scaling for LLMs.

This paper introduces LongBench v2, a benchmark designed to assess the ability of LLMs to handle long-context problems requiring deep understanding and reasoning across real-world multitasks. LongBench v2 consists of 503 challenging multiple-choice questions, with contexts ranging from 8k to 2M words, across six major task categories: single-document QA, multi-document QA, long in-context learning, long-dialogue history understanding, code repository understanding, and long structured data understanding. To ensure the breadth and the practicality, we collect data from nearly 100 highly educated individuals with diverse professional backgrounds. We employ both automated and manual review processes to maintain high quality and difficulty, resulting in human experts achieving only 53.7% accuracy under a 15-minute time constraint. Our evaluation reveals that the best-performing model, when directly answers the questions, achieves only 50.1% accuracy. In contrast, the o1-preview model, which includes longer reasoning, achieves 57.7%, surpassing the human baseline by 4%. These results highlight the importance of enhanced reasoning ability and scaling inference-time compute to tackle the long-context challenges in LongBench v2.

Current inference scaling methods, such as Self-consistency and Best-of-N, have proven effective in improving the accuracy of LLMs on complex reasoning tasks. However, these methods rely heavily on the quality of candidate responses and are unable to produce correct answers when all candidates are incorrect. In this paper, we propose a novel inference scaling strategy, CoT-based Synthesizer, which leverages CoT reasoning to synthesize superior answers by analyzing complementary information from multiple candidate responses, even when all candidates are flawed. To support a lightweight and cost-effective implementation, we introduce an automated data generation pipeline that creates diverse training data. This enables smaller LLMs trained on this data to improve the inference accuracy of larger models, including API-based LLMs. Experimental results across four benchmark datasets with seven policy models demonstrate that our method significantly enhances performance, with gains of 11.8% for Llama3-8B and 10.3% for GPT-4o on the MATH dataset. The corresponding training data and code are publicly available on the [repository](https://github.com/RUCKBReasoning/CoT-based-Synthesizer).

Current large language models (LLMs) often struggle to produce accurate responses on the first attempt for complex reasoning tasks like code generation. Prior research tackles this challenge by generating multiple candidate solutions and validating them with LLM-generated unit tests. The execution results of unit tests serve as reward signals to identify correct solutions. As LLMs always confidently make mistakes, these unit tests are not reliable, thereby diminishing the quality of reward signals. Motivated by the observation that scaling the number of solutions improves LLM performance, we explore the impact of scaling unit tests to enhance reward signal quality. Our pioneer experiment reveals a positive correlation between the number of unit tests and reward signal quality, with greater benefits observed in more challenging problems. Based on these insights, we propose CodeRM-8B, a lightweight yet effective unit test generator that enables efficient and high-quality unit test scaling. Additionally, we implement a dynamic scaling mechanism that adapts the number of unit tests based on problem difficulty, further improving efficiency. Experimental results show that our approach significantly improves performance across various models on three benchmarks (e.g., with gains of 18.43 for Llama3-8B and 3.42 for GPT-4o-mini on HumanEval Plus). The parameters of CodeRM-8B and corresponding training data will be available upon publication.

Reinforcement learning (RL) with tree search has demonstrated superior performance in traditional reasoning tasks. Compared to conventional independent chain sampling strategies with outcome supervision, tree search enables better exploration of the reasoning space and provides dense, on-policy process rewards during RL training but remains under-explored in On-Policy LLM RL. We propose TreeRL, a reinforcement learning framework that directly incorporates on-policy tree search for RL training. Our approach includes intermediate supervision and eliminates the need for separate reward model training. Existing approaches typically train a separate process reward model, which can suffer from distribution mismatch and reward hacking. We also introduce a cost-effective tree search approach that achieves higher search efficiency under the same generation token budget by strategically branching from high-uncertainty intermediate steps rather than using random branching. Experiments on challenging math and code reasoning benchmarks demonstrate that TreeRL achieves superior performance compared to traditional ChainRL, highlighting the potential of tree search for LLM. TreeRL is open-sourced at https://github.com/THUDM/TreeRL.

As Large Language Models (LLMs) continue to advance in understanding and generating long sequences, new safety concerns have been introduced through the long context. However, the safety of LLMs in long-context tasks remains under-explored, leaving a significant gap in both evaluation and improvement of their safety. To address this, we introduce LongSafety, the first comprehensive benchmark specifically designed to evaluate LLM safety in open-ended long-context tasks. LongSafety encompasses 7 categories of safety issues and 6 user-oriented long-context tasks, with a total of 1,543 test cases, averaging 5,424 words per context. Our evaluation towards 16 representative LLMs reveals significant safety vulnerabilities, with most models achieving safety rates below 55%. Our findings also indicate that strong safety performance in short-context scenarios does not necessarily correlate with safety in long-context tasks, emphasizing the unique challenges and urgency of improving long-context safety. Moreover, through extensive analysis, we identify challenging safety issues and task types for long-context models. Furthermore, we find that relevant context and extended input sequences can exacerbate safety risks in long-context scenarios, highlighting the critical need for ongoing attention to long-context safety challenges. Our code and data will be publicly available.

Large Language Models (LLMs) have demonstrated notable capabilities across various tasks, showcasing complex problem-solving abilities. Understanding and executing complex rules, along with multi-step planning, are fundamental to logical reasoning and critical for practical LLM agents and decision-making systems. However, evaluating LLMs as effective rule-based executors and planners remains underexplored. In this paper, we introduce LogicGame, a novel benchmark designed to evaluate the comprehensive rule understanding, execution, and planning capabilities of LLMs. Unlike traditional benchmarks, LogicGame provides diverse games that contain a series of rules with an initial state, requiring models to comprehend and apply predefined regulations to solve problems. We create simulated scenarios in which models execute or plan operations to achieve specific outcomes. These game scenarios are specifically designed to distinguish logical reasoning from mere knowledge by relying exclusively on predefined rules. This separation allows for a pure assessment of rule-based reasoning capabilities. The evaluation considers not only final outcomes but also intermediate steps, providing a comprehensive assessment of model performance. Moreover, these intermediate steps are deterministic and can be automatically verified. LogicGame defines game scenarios with varying difficulty levels, from simple rule applications to complex reasoning chains, in order to offer a precise evaluation of model performance on rule understanding and multi-step execution. Utilizing LogicGame, we test various LLMs and identify notable shortcomings in their rule-based logical reasoning abilities.

Large language models (LLMs) have advanced rapidly from conversational problem solving to addressing real-world tasks involving tool use, such as software engineering (SWE). Recent LLM-powered toolkits, such as OpenAI Codex and Cursor, have offered end-to-end automation of the software development process. However, building effective SWE agents remains challenging due to the lack of high-quality training data and effective test cases. To address this issue, we present SWE-Dev, an SWE agent built upon open-source LLMs. First, we develop a robust pipeline to synthesize test cases for patch evaluation. Second, we scale up agent trajectories to construct the training data for building SWE-Dev. Experiments on the SWE-bench-Verified benchmark show that the SWE-Dev models can achieve top performance among all open SWE agents. Specifically, the success rates of the SWE-Dev 7B and 32B parameter models reach 23.4% and 36.6%, respectively, outperforming state-of-the-art open-source models. All code, models, and datasets are publicly available at https://github.com/THUDM/SWE-Dev.

We introduce TableLLM, a robust large language model (LLM) with 8 billion parameters, purpose-built for proficiently handling tabular data manipulation tasks, whether they are embedded within documents or spreadsheets, catering to real-world office scenarios. We propose a distant supervision method for training, which comprises a reasoning process extension strategy, aiding in training LLMs to understand reasoning patterns more effectively as well as a cross-way validation strategy, ensuring the quality of the automatically generated data. To evaluate the performance of TableLLM, we have crafted benchmarks tailored to address both document and spreadsheet formats as well as constructed a well-organized evaluation pipeline capable of handling both scenarios. Thorough evaluations underscore the advantages of TableLLM when compared to various existing general-purpose and tabular data-focused LLMs. We have publicly released the model checkpoint, source code, benchmarks, and a web application for user interaction on this anonymized repository.

Since the adoption of large language models (LLMs) for text evaluation has become increasingly prevalent in the field of natural language processing (NLP), a series of existing works attempt to optimize the prompts for LLM evaluators to improve their alignment with human judgment. However, their efforts are limited to optimizing individual factors of evaluation prompts, such as evaluation criteria or output formats, neglecting the combinatorial impact of multiple factors, which leads to insufficient optimization of the evaluation pipeline. Nevertheless, identifying well-behaved prompting strategies for adjusting multiple factors requires extensive enumeration. To this end, we comprehensively integrate 8 key factors for evaluation prompts and propose a novel automatic prompting strategy optimization method called Heuristic Prompting Strategy Search (HPSS). Inspired by the genetic algorithm, HPSS conducts an iterative search to find well-behaved prompting strategies for LLM evaluators. A heuristic function is employed to guide the search process, enhancing the performance of our algorithm. Extensive experiments across four evaluation tasks demonstrate the effectiveness of HPSS, consistently outperforming both human-designed evaluation prompts and existing automatic prompt optimization methods. Our code is available athttps://github.com/thu-coai/HPSS.

Although large language models (LLMs) demonstrate impressive performance for many language tasks, most of them can only handle texts a few thousand tokens long, limiting their applications on longer sequence inputs, such as books, reports, and codebases. Recent works have proposed methods to improve LLMs’ long context capabilities by extending context windows and more sophisticated memory mechanisms. However, comprehensive benchmarks tailored for evaluating long context understanding are lacking. In this paper, we introduce LongBench, the first bilingual, multi-task benchmark for long context understanding, enabling a more rigorous evaluation of long context understanding. LongBench comprises 21 datasets across 6 task categories in both English and Chinese, with an average length of 6,711 words (English) and 13,386 characters (Chinese). These tasks cover key long-text application areas including single-doc QA, multi-doc QA, summarization, few-shot learning, synthetic tasks, and code completion. All datasets in LongBench are standardized into a unified format, allowing for effortless automatic evaluation of LLMs. Upon comprehensive evaluation of 8 LLMs on LongBench, we find that: (1) Commercial model (GPT-3.5-Turbo-16k) outperforms other open-sourced models, but still struggles on longer contexts. (2) Scaled position embedding and fine-tuning on longer sequences lead to substantial improvement on long context understanding. (3) Context compression technique such as retrieval brings improvement for model with weak ability on long contexts, but the performance still lags behind models that have strong long context understanding capability.

Large language models (LLMs) have shown impressive success in various applications. However, these models are often not well aligned with human intents, which calls for additional treatments on them; that is, the alignment problem. To make LLMs better follow user instructions, existing alignment methods primarily focus on further training them. However, the extra training of LLMs is usually expensive in terms of GPU computing; even worse, some LLMs are not accessible for user-demanded training, such as GPTs. In this work, we take a different perspective—Black-Box Prompt Optimization (BPO)—to perform alignments. The idea is to optimize user prompts to suit LLMs’ input understanding, so as to best realize users’ intents without updating LLMs’ parameters. BPO leverages human preferences to optimize prompts, thus making it superior to LLM (e.g., ChatGPT) as a prompt engineer. Moreover, BPO is model-agnostic, and the empirical results demonstrate that the BPO-aligned ChatGPT yields a 22% increase in the win rate against its original version and 10% for GPT-4. Notably, the BPO-aligned LLMs can outperform the same models aligned by PPO and DPO, and it also brings additional performance gains when combining BPO with PPO or DPO. Code and datasets are released at https://github.com/thu-coai/BPO.

Alignment has become a critical step for instruction-tuned Large Language Models (LLMs) to become helpful assistants. However, effective evaluation of alignment for emerging Chinese LLMs is still significantly lacking, calling for real-scenario grounded, open-ended, challenging and automatic evaluations tailored for alignment. To fill in this gap, we introduce AlignBench, a comprehensive multi-dimensional benchmark for evaluating LLMs’ alignment in Chinese. We tailor a human-in-the-loop data curation pipeline, containing 8 main categories, 683 real-scenario rooted queries and corresponding human verified references.To ensure references’ correctness, each knowledge-intensive query is accompanied with evidences collected from reliable webpages (including the url and quotation) by our annotators.For automatic evaluation, our benchmark employs a rule-calibrated multi-dimensional LLM-as-Judge (CITATION) with Chain-of-Thought to generate explanations and final ratings as evaluations, ensuring high reliability and interpretability.All evaluation codes and data are publicly available at https://github.com/THUDM/AlignBench

This paper proposes PiNose, which trains a probing model on offline self-consistency checking results, thereby circumventing the need for human-annotated data and achieving transferability across diverse data distributions. As the consistency check process is offline, PiNose reduces the computational burden of generating multiple responses by online consistency verification. Additionally, it examines various aspects of internal states prior to response decoding, contributing to more effective detection of factual inaccuracies. Experiment results on both factuality detection and question answering benchmarks show that PiNose achieves surpassing results than existing factuality detection methods.

Since the natural language processing (NLP) community started to make large language models (LLMs) act as a critic to evaluate the quality of generated texts, most of the existing works train a critique generation model on the evaluation data labeled by GPT-4’s direct prompting. We observe that these models lack the ability to generate informative critiques in both pointwise grading and pairwise comparison especially without references. As a result, their generated critiques cannot provide fine-grained distinguishability on generated texts, causing unsatisfactory evaluation performance. In this paper, we propose a simple yet effective method called Eval-Instruct, which can first acquire pointwise grading critiques with pseudo references and then revise these critiques via multi-path prompting to obtain informative evaluation data in different tasks and settings, including pointwise grading and pairwise comparison with / without references. After fine-tuning on these data, the resulting model CritiqueLLM is empirically shown to outperform ChatGPT and all the open-source baselines and even achieve comparable evaluation performance to GPT-4 in system-level correlations of pointwise grading. We also demonstrate that our generated critiques can act as scalable feedback to further improve the generation quality of strong LLMs like ChatGPT.

With the rapid development of Large Language Models (LLMs), increasing attention has been paid to their safety concerns. Consequently, evaluating the safety of LLMs has become an essential task for facilitating the broad applications of LLMs. Nevertheless, the absence of comprehensive safety evaluation benchmarks poses a significant impediment to effectively assess and enhance the safety of LLMs. In this work, we present SafetyBench, a comprehensive benchmark for evaluating the safety of LLMs, which comprises 11,435 diverse multiple choice questions spanning across 7 distinct categories of safety concerns. Notably, SafetyBench also incorporates both Chinese and English data, facilitating the evaluation in both languages. Our extensive tests over 25 popular Chinese and English LLMs in both zero-shot and few-shot settings reveal a substantial performance advantage for GPT-4 over its counterparts, and there is still significant room for improving the safety of current LLMs. We also demonstrate that the measured safety understanding abilities in SafetyBench are correlated with safety generation abilities. Data and evaluation guidelines are available at https://github.com/thu-coai/SafetyBench. Submission entrance and leaderboard are available at https://llmbench.ai/safety.

We introduce OpenWebAgent, an open toolkit designed to optimize web automation by integrating both large language models (LLMs) and large multimodal models (LMMs). This toolkit focuses on enhancing human-computer interactions on the web, simplifying complex tasks through an advanced HTML parser, a rapid action generation module, and an intuitive user interface. At the core of OpenWebAgent is an innovative web agent framework that uses a modular design to allow developers to seamlessly integrate a variety of models and tools to process web information and automate tasks on the web. This enables the development of powerful, task-oriented web agents, significantly enhancing user experience and operational efficiency on the web. The OpenWebAgent framework, Chrome plugin, and demo video are available at https://github.com/THUDM/OpenWebAgent/.

Large Language Models (LLMs) have demonstrated exceptional abilities in comprehending and generating text, motivating numerous researchers to utilize them for Information Extraction (IE) purposes, including Relation Extraction (RE). Nonetheless, most existing methods are predominantly designed for Sentence-level Relation Extraction (SentRE) tasks, which typically encompass a restricted set of relations and triplet facts within a single sentence. Furthermore, certain approaches resort to treating relations as candidate choices integrated into prompt templates, leading to inefficient processing and suboptimal performance when tackling Document-Level Relation Extraction (DocRE) tasks, which entail handling multiple relations and triplet facts distributed across a given document, posing distinct challenges. To overcome these limitations, we introduce AutoRE, an end-to-end DocRE model that adopts a novel RE extraction paradigm named RHF (Relation-Head-Facts). Unlike existing approaches, AutoRE does not rely on the assumption of known relation options, making it more reflective of real-world scenarios. Additionally, we have developed an easily extensible RE framework using a Parameters Efficient Fine Tuning (PEFT) algorithm (QLoRA). Our experiments on the RE-DocRED dataset showcase AutoRE’s best performance, achieving state-of-the-art results, surpassing TAG by 10.03% and 9.03% respectively on the dev and test set. The code is available and the demonstration video is provided.

The applications of large language models (LLMs) have expanded well beyond the confines of text processing, signaling a new era where LLMs are envisioned as generalist agents capable of operating within complex environments. These environments are often highly expansive, making it impossible for the LLM to process them within its short-term memory. Motivated by recent research on extending the capabilities of LLMs with tools, we seek to investigate the intriguing potential of tools to augment LLMs in handling such complexity by introducing a novel class of tools, termed *middleware*, to aid in the proactive exploration within these massive environments. Such specialized tools can serve as a middleware layer shielding the LLM from environmental complexity. In two representative complex environments—knowledge bases (KBs) and databases—we demonstrate the significant potential of augmenting language agents with tools in complex environments. Notably, equipped with the middleware, GPT-4 achieves **2.8**X the performance of the best baseline in tasks requiring access to database content and **2.2**X in KB tasks. Our findings illuminate the path for advancing language agents in real-world applications.

Long-Context Question Answering (LCQA), a challenging task, aims to reason over long-context documents to yield accurate answers to questions. Existing long-context Large Language Models (LLMs) for LCQA often struggle with the “lost in the middle” issue. Retrieval-Augmented Generation (RAG) mitigates this issue by providing external factual evidence. However, its chunking strategy disrupts the global long-context information, and its low-quality retrieval in long contexts hinders LLMs from identifying effective factual details due to substantial noise. To this end, we propose LongRAG, a general, dual-perspective, and robust LLM-based RAG system paradigm for LCQA to enhance RAG’s understanding of complex long-context knowledge (i.e., global information and factual details). We design LongRAG as a plug-and-play paradigm, facilitating adaptation to various domains and LLMs. Extensive experiments on three multi-hop datasets demonstrate that LongRAG significantly outperforms long-context LLMs (up by 6.94%), advanced RAG (up by 6.16%), and Vanilla RAG (up by 17.25%). Furthermore, we conduct quantitative ablation studies and multi-dimensional analyses, highlighting the effectiveness of the system’s components and fine-tuning strategies.Data and code are available at [https://github.com/QingFei1/LongRAG](https://github.com/QingFei1/LongRAG).

Character-based dialogue (CharacterDial) has become essential in the industry (e.g., Character.AI), enabling users to freely customize social characters for social interactions. However, the generalizability and adaptability across various conversational scenarios inherent in customizing social characters still lack public industrial solutions. To address these challenges, by dissecting well-rounded social characters composed of both inherent social profiles and external social behaviors, we manually collect a large-scale Chinese corpus featuring characters with diverse categories and behaviors, and develop CharacterGLM models alongside well-designed refinement methods. Extensive experiments show that CharacterGLM outperforms most popular open- and closed-source LLMs and performs comparably to GPT-4. We will release our data and models for local development and deployment.

Open large language models (LLMs) with great performance in various tasks have significantly advanced the development of LLMs. However, they are far inferior to commercial models such as ChatGPT and GPT-4 when acting as agents to tackle complex tasks in the real world. These agent tasks employ LLMs as the central controller responsible for planning, memorization, and tool utilization, necessitating both fine-grained prompting methods and robust LLMs to achieve satisfactory performance. Though many prompting methods have been proposed to complete particular agent tasks, there is lack of research focusing on improving the agent capabilities of LLMs themselves without compromising their general abilities. In this work, we present AgentTuning, a simple and general method to enhance the agent abilities of LLMs while maintaining their general LLM capabilities. We construct AgentInstruct, a lightweight instruction-tuning dataset containing high-quality interaction trajectories. We employ a hybrid instruction-tuning strategy by combining AgentInstruct with open-source instructions from general domains. AgentTuning is used to instruction-tune the Llama 2 series, resulting in AgentLM. Our evaluations show that AgentTuning enables LLMs’ agent capabilities without compromising general abilities. The AgentLM-70B is comparable to GPT-3.5-turbo on unseen agent tasks, demonstrating generalized agent capabilities. We open source the AgentInstruct and AgentLM-7B, 13B, and 70B models at https://anonymous.4open.science/r/AgentTuning, serving open and powerful alternatives to commercial LLMs for agent tasks.

Large language models (LLMs) have manifested strong ability to generate codes for productive activities. However, current benchmarks for code synthesis, such as HumanEval, MBPP, and DS-1000, are predominantly oriented towards introductory tasks on algorithm and data science, insufficiently satisfying challenging requirements prevalent in real-world coding. To fill this gap, we propose NaturalCodeBench (NCB), a challenging code benchmark designed to mirror the complexity and variety of scenarios in real coding tasks. NCB comprises 402 high-quality problems in Python and Java, meticulously selected from natural user queries from online coding services, covering 6 different domains. Noting the extraordinary difficulty in creating testing cases for real-world queries, we also introduce a semi-automated pipeline to enhance the efficiency of test case construction. Comparing with manual solutions, it achieves an efficiency increase of more than 4 times. Our systematic experiments on 39 LLMs find that performance gaps on NCB between models with close HumanEval scores could still be significant, indicating a lack of focus on practical code synthesis scenarios or over-specified optimization on HumanEval. On the other hand, even the best-performing GPT-4 is still far from satisfying on NCB. The evaluation toolkit and development set are available at https://github.com/THUDM/NaturalCodeBench.

We identify two crucial limitations in the evaluation of recent parallel-integrated method Parallel Context Windows (PCW), which extends the maximum context lengths of language models, e.g., 2048 for LLaMA, by harnessing window-wise attention and positional embedding techniques. We first show that a simple yet strong baseline, weighted sum ensemble, is missing for the in-context few-shot classification. Moreover, on more challenging Chain-of-Thought (CoT) reasoning (e.g., HotpotQA), PCW would present unexpected deterioration regarding question miscomprehension and false inference. Based on our findings, we suggest that the existing PCW design may not guarantee sufficient improvement and practicality in handling lengthy documents in real-world applications. More community efforts on enabling language models’ long context understanding ability should be paid.

Extending large language models to effectively handle long contexts requires instruction fine-tuning on input sequences of similar length. To address this, we present LongAlign—a recipe of the instruction data, training, and evaluation for long context alignment. First, we construct a long instruction-following dataset using Self-Instruct. To ensure the data diversity, it covers a broad range of tasks from various long context sources. Second, we adopt the packing and sorted batching strategies to speed up supervised fine-tuning on data with varied length distributions. Additionally, we develop a loss weighting method to balance the contribution to the loss across different sequences during packing training. Third, we introduce the LongBench-Chat benchmark for evaluating instruction-following capabilities on queries of 10k-100k in length. Experiments show that LongAlign outperforms existing recipes for LLMs in long context tasks by up to 30%, while also maintaining their proficiency in handling short, generic tasks.

Although Large Language Models (LLMs) are becoming increasingly powerful, they still exhibit significant but subtle weaknesses, such as mistakes in instruction-following or coding tasks.As these unexpected errors could lead to severe consequences in practical deployments, it is crucial to investigate the limitations within LLMs systematically.Traditional benchmarking approaches cannot thoroughly pinpoint specific model deficiencies, while manual inspections are costly and not scalable. In this paper, we introduce a unified framework, AutoDetect, to automatically expose weaknesses in LLMs across various tasks. Inspired by the educational assessment process that measures students’ learning outcomes, AutoDetect consists of three LLM-powered agents: Examiner, Questioner, and Assessor.The collaboration among these three agents is designed to realize comprehensive and in-depth weakness identification. Our framework demonstrates significant success in uncovering flaws, with an identification success rate exceeding 30% in prominent models such as ChatGPT and Claude.More importantly, these identified weaknesses can guide specific model improvements, proving more effective than untargeted data augmentation methods like Self-Instruct. Our approach has led to substantial enhancements in popular LLMs, including the Llama series and Mistral-7b, boosting their performance by over 10% across several benchmarks.Code and data are publicly available at https://github.com/thu-coai/AutoDetect.

Large language models (LLMs) have shown excellent mastering of human language but still struggle in real-world applications that require mathematical problem-solving. While many strategies and datasets to enhance LLMs’ mathematics are developed, it remains a challenge to simultaneously maintain and improve both language and mathematical capabilities in deployed LLM systems. In this work, we tailor the Self-Critique pipeline, which addresses the challenge in the feedback learning stage of LLM alignment. We first train a general Math-Critique model from the LLM itself to provide feedback signals. Then, we sequentially employ rejective fine-tuning and direct preference optimization over the LLM’s own generations for data collection. Based on ChatGLM3-32B, we conduct experiments on both academic and our newly created challenging dataset, MathUserEval. Results show that our pipeline significantly enhances the LLM’s mathematical problem-solving while still improving its language ability, outperforming LLMs that could be two times larger. Related techniques have been deployed to ChatGLM, an online serving LLM. Related evaluation datasets and scripts are released at https://github.com/THUDM/ChatGLM-Math.

Currently, the reduction in the parameter scale of large-scale pre-trained language models (PLMs) through knowledge distillation has greatly facilitated their widespread deployment on various devices. However, the deployment of knowledge distillation systems faces great challenges in real-world industrial-strength applications, which require the use of complex distillation methods on even larger-scale PLMs (over 10B), limited by memory on GPUs and the switching of methods. To overcome these challenges, we propose GKD, a general knowledge distillation framework that supports distillation on larger-scale PLMs using various distillation methods. With GKD, developers can build larger distillation models on memory-limited GPUs and easily switch and combine different distillation methods within a single framework. Experimental results show that GKD can support the distillation of at least 100B-scale PLMs and 25 mainstream methods on 8 NVIDIA A100 (40GB) GPUs.

The large scale of pre-trained language models poses a challenge for their deployment on various devices, with a growing emphasis on methods to compress these models, particularly knowledge distillation. However, current knowledge distillation methods rely on the model’s intermediate layer features and the golden labels (also called hard labels), which usually require aligned model architecture and enough labeled data respectively. Moreover, the parameters of vocabulary are usually neglected in existing methods. To address these problems, we propose a general language model distillation (GLMD) method that performs two-stage word prediction distillation and vocabulary compression, which is simple and surprisingly shows extremely strong performance. Specifically, GLMD supports more general application scenarios by eliminating the constraints of dimension and structure between models and the need for labeled datasets through the absence of intermediate layers and golden labels. Meanwhile, based on the long-tailed distribution of word frequencies in the data, GLMD designs a strategy of vocabulary compression through decreasing vocabulary size instead of dimensionality. Experimental results show that our method outperforms 25 state-of-the-art methods on the SuperGLUE benchmark, achieving an average score that surpasses the best method by 3%.

Prompt tuning attempts to update few task-specific parameters in pre-trained models. It has achieved comparable performance to fine-tuning of the full parameter set on both language understanding and generation tasks. In this work, we study the problem of prompt tuning for neural text retrievers. We introduce parameter-efficient prompt tuning for text retrieval across in-domain, cross-domain, and cross-topic settings. Through an extensive analysis, we show that the strategy can mitigate the two issues—parameter-inefficiency and weak generalizability—faced by fine-tuning based retrieval methods. Notably, it can significantly improve the out-of-domain zero-shot generalization of the retrieval models. By updating only 0.1% of the model parameters, the prompt tuning strategy can help retrieval models achieve better generalization performance than traditional methods in which all parameters are updated. Finally, to facilitate research on retrievers’ cross-topic generalizability, we curate and release an academic retrieval dataset with 18K query-results pairs in 87 topics, making it the largest topic-specific one to date.

Evaluating open-domain dialogue systems is currently an open question. Automatic evaluation metrics have shown poor correlation with human assessment in dialogue generation tasks. Human evaluation, which involves annotators for multi-dimension scoring, is trustworthy but time-consuming. In this work, we propose FFAEval, a reliable and efficient human evaluation framework using Free-For-All ranking approach. By sharing the dialogue history, the framework enables annotators to converse with multiple dialogue systems simultaneously in a single-blind, multi-turn manner. The subsequent free-for-all allows annotators to select the most favourable model in each turn from among all the participating dialogue systems. The final performance of each model is represented by calculating the TrueSkill score derived from the free-for-all competition. Our empirical study on English and Chinese dialogue systems demonstrates that FFAEval achieves a strong correlation with score-based human assessment compared to existing evaluation methods. We further prove the efficiency and stability of our framework in additional experiments. The source code and data are available on Github.

There have been various types of pretraining architectures including autoencoding models (e.g., BERT), autoregressive models (e.g., GPT), and encoder-decoder models (e.g., T5). However, none of the pretraining frameworks performs the best for all tasks of three main categories including natural language understanding (NLU), unconditional generation, and conditional generation. We propose a General Language Model (GLM) based on autoregressive blank infilling to address this challenge. GLM improves blank filling pretraining by adding 2D positional encodings and allowing an arbitrary order to predict spans, which results in performance gains over BERT and T5 on NLU tasks. Meanwhile, GLM can be pretrained for different types of tasks by varying the number and lengths of blanks. On a wide range of tasks across NLU, conditional and unconditional generation, GLM outperforms BERT, T5, and GPT given the same model sizes and data, and achieves the best performance from a single pretrained model with 1.25× parameters of BERT Large , demonstrating its generalizability to different downstream tasks.

The few-shot natural language understanding (NLU) task has attracted much recent attention. However, prior methods have been evaluated under a disparate set of protocols, which hinders fair comparison and measuring the progress of the field. To address this issue, we introduce an evaluation framework that improves previous evaluation procedures in three key aspects, i.e., test performance, dev-test correlation, and stability. Under this new evaluation framework, we re-evaluate several state-of-the-art few-shot methods for NLU tasks. Our framework reveals new insights: (1) both the absolute performance and relative gap of the methods were not accurately estimated in prior literature; (2) no single method dominates most tasks with consistent performance; (3) improvements of some methods diminish with a larger pretrained model; and (4) gains from different methods are often complementary and the best combined model performs close to a strong fully-supervised baseline. We open-source our toolkit, FewNLU, that implements our evaluation framework along with a number of state-of-the-art methods.

Recent works on knowledge base question answering (KBQA) retrieve subgraphs for easier reasoning. The desired subgraph is crucial as a small one may exclude the answer but a large one might introduce more noises. However, the existing retrieval is either heuristic or interwoven with the reasoning, causing reasoning on the partial subgraphs, which increases the reasoning bias when the intermediate supervision is missing. This paper proposes a trainable subgraph retriever (SR) decoupled from the subsequent reasoning process, which enables a plug-and-play framework to enhance any subgraph-oriented KBQA model. Extensive experiments demonstrate SR achieves significantly better retrieval and QA performance than existing retrieval methods. Via weakly supervised pre-training as well as the end-to-end fine-tuning, SR achieves new state-of-the-art performance when combined with NSM (He et al., 2021), a subgraph-oriented reasoner, for embedding-based KBQA methods. Codes and datasets are available online (https://github.com/RUCKBReasoning/SubgraphRetrievalKBQA)

Most previous methods for text data augmentation are limited to simple tasks and weak baselines. We explore data augmentation on hard tasks (i.e., few-shot natural language understanding) and strong baselines (i.e., pretrained models with over one billion parameters). Under this setting, we reproduced a large number of previous augmentation methods and found that these methods bring marginal gains at best and sometimes degrade the performance much. To address this challenge, we propose a novel data augmentation method FlipDA that jointly uses a generative model and a classifier to generate label-flipped data. Central to the idea of FlipDA is the discovery that generating label-flipped data is more crucial to the performance than generating label-preserved data. Experiments show that FlipDA achieves a good tradeoff between effectiveness and robustness—it substantially improves many tasks while not negatively affecting the others.

Prompt tuning, which only tunes continuous prompts with a frozen language model, substantially reduces per-task storage and memory usage at training. However, in the context of NLU, prior work reveals that prompt tuning does not perform well for normal-sized pretrained models. We also find that existing methods of prompt tuning cannot handle hard sequence labeling tasks, indicating a lack of universality. We present a novel empirical finding that properly optimized prompt tuning can be universally effective across a wide range of model scales and NLU tasks. It matches the performance of finetuning while having only 0.1%-3% tuned parameters. Our method P-Tuning v2 is an implementation of Deep Prompt Tuning (CITATION) optimized and adapted for NLU. Given the universality and simplicity of P-Tuning v2, we believe it can serve as an alternative to finetuning and a strong baseline for future research.

We investigate the usage of entity linking (EL)in downstream tasks and present the first modularized EL toolkit for easy task adaptation. Different from the existing EL methods that dealwith all the features simultaneously, we modularize the whole model into separate parts witheach feature. This decoupled design enablesflexibly adding new features without retraining the whole model as well as flow visualization with better interpretability of the ELresult. We release the corresponding toolkit,HOSMEL, for Chinese, with three flexible usage modes, a live demo, and a demonstrationvideo. Experiments on two benchmarks forthe question answering task demonstrate thatHOSMEL achieves much less time and spaceconsumption as well as significantly better accuracy performance compared with existingSOTA EL methods. We hope the release ofHOSMEL will call for more attention to studyEL for downstream tasks in non-English languages.

In recent years, pretrained models revolutionized the paradigm of natural language understanding (NLU), where we append a randomly initialized classification head after the pretrained backbone, e.g. BERT, and finetune the whole model. As the pretrained backbone makes a major contribution to the improvement, we naturally expect a good pretrained classification head can also benefit the training. However, the final-layer output of the backbone, i.e. the input of the classification head, will change greatly during finetuning, making the usual head-only pretraining ineffective. In this paper, we find that parameter-efficient tuning makes a good classification head, with which we can simply replace the randomly initialized heads for a stable performance gain. Our experiments demonstrate that the classification head jointly pretrained with parameter-efficient tuning consistently improves the performance on 9 tasks in GLUE and SuperGLUE.

We introduce a method for improving the structural understanding abilities of language models. Unlike previous approaches that finetune the models with task-specific augmentation, we pretrain language models to generate structures from the text on a collection of task-agnostic corpora. Our structure pretraining enables zero-shot transfer of the learned knowledge that models have about the structure tasks. We study the performance of this approach on 28 datasets, spanning 10 structure prediction tasks including open information extraction, joint entity and relation extraction, named entity recognition, relation classification, semantic role labeling, event extraction, coreference resolution, factual probe, intent detection, and dialogue state tracking. We further enhance the pretraining with the task-specific training sets. We show that a 10B parameter language model transfers non-trivially to most tasks and obtains state-of-the-art performance on 21 of 28 datasets that we evaluate. Our code and datasets will be made publicly available.

We cast a suite of information extraction tasks into a text-to-triple translation framework. Instead of solving each task relying on task-specific datasets and models, we formalize the task as a translation between task-specific input text and output triples. By taking the task-specific input, we enable a task-agnostic translation by leveraging the latent knowledge that a pre-trained language model has about the task. We further demonstrate that a simple pre-training task of predicting which relational information corresponds to which input text is an effective way to produce task-specific outputs. This enables the zero-shot transfer of our framework to downstream tasks. We study the zero-shot performance of this framework on open information extraction (OIE2016, NYT, WEB, PENN), relation classification (FewRel and TACRED), and factual probe (Google-RE and T-REx). The model transfers non-trivially to most tasks and is often competitive with a fully supervised method without the need for any task-specific training. For instance, we significantly outperform the F1 score of the supervised open information extraction without needing to use its training set.

Within the prosperity of Massive Open Online Courses (MOOCs), the education applications that automatically provide extracurricular knowledge for MOOC users become rising research topics. However, MOOC courses’ diversity and rapid updates make it more challenging to find suitable new knowledge for students. In this paper, we present ExpanRL, an end-to-end hierarchical reinforcement learning (HRL) model for concept expansion in MOOCs. Employing a two-level HRL mechanism of seed selection and concept expansion, ExpanRL is more feasible to adjust the expansion strategy to find new concepts based on the students’ feedback on expansion results. Our experiments on nine novel datasets from real MOOCs show that ExpanRL achieves significant improvements over existing methods and maintain competitive performance under different settings.

The prosperity of Massive Open Online Courses (MOOCs) provides fodder for many NLP and AI research for education applications, e.g., course concept extraction, prerequisite relation discovery, etc. However, the publicly available datasets of MOOC are limited in size with few types of data, which hinders advanced models and novel attempts in related topics. Therefore, we present MOOCCube, a large-scale data repository of over 700 MOOC courses, 100k concepts, 8 million student behaviors with an external resource. Moreover, we conduct a prerequisite discovery task as an example application to show the potential of MOOCCube in facilitating relevant research. The data repository is now available at http://moocdata.cn/data/MOOCCube.

We present BlockBERT, a lightweight and efficient BERT model for better modeling long-distance dependencies. Our model extends BERT by introducing sparse block structures into the attention matrix to reduce both memory consumption and training/inference time, which also enables attention heads to capture either short- or long-range contextual information. We conduct experiments on language model pre-training and several benchmark question answering datasets with various paragraph lengths. BlockBERT uses 18.7-36.1% less memory and 12.0-25.1% less time to learn the model. During testing, BlockBERT saves 27.8% inference time, while having comparable and sometimes better prediction accuracy, compared to an advanced BERT-based model, RoBERTa.

In this paper, we propose a novel end-to-end framework called KBRD, which stands for Knowledge-Based Recommender Dialog System. It integrates the recommender system and the dialog generation system. The dialog generation system can enhance the performance of the recommendation system by introducing information about users’ preferences, and the recommender system can improve that of the dialog generation system by providing recommendation-aware vocabulary bias. Experimental results demonstrate that our proposed model has significant advantages over the baselines in both the evaluation of dialog generation and recommendation. A series of analyses show that the two systems can bring mutual benefits to each other, and the introduced knowledge contributes to both their performances.

We propose a new CogQA framework for multi-hop reading comprehension question answering in web-scale documents. Founded on the dual process theory in cognitive science, the framework gradually builds a cognitive graph in an iterative process by coordinating an implicit extraction module (System 1) and an explicit reasoning module (System 2). While giving accurate answers, our framework further provides explainable reasoning paths. Specifically, our implementation based on BERT and graph neural network efficiently handles millions of documents for multi-hop reasoning questions in the HotpotQA fullwiki dataset, achieving a winning joint F₁ score of 34.9 on the leaderboard, compared to 23.1 of the best competitor.

As Massive Open Online Courses (MOOCs) become increasingly popular, it is promising to automatically provide extracurricular knowledge for MOOC users. Suffering from semantic drifts and lack of knowledge guidance, existing methods can not effectively expand course concepts in complex MOOC environments. In this paper, we first build a novel boundary during searching for new concepts via external knowledge base and then utilize heterogeneous features to verify the high-quality results. In addition, to involve human efforts in our model, we design an interactive optimization mechanism based on a game. Our experiments on the four datasets from Coursera and XuetangX show that the proposed method achieves significant improvements(+0.19 by MAP) over existing methods.

Massive Open Online Courses (MOOCs), offering a new way to study online, are revolutionizing education. One challenging issue in MOOCs is how to design effective and fine-grained course concepts such that students with different backgrounds can grasp the essence of the course. In this paper, we conduct a systematic investigation of the problem of course concept extraction for MOOCs. We propose to learn latent representations for candidate concepts via an embedding-based method. Moreover, we develop a graph-based propagation algorithm to rank the candidate concepts based on the learned representations. We evaluate the proposed method using different courses from XuetangX and Coursera. Experimental results show that our method significantly outperforms all the alternative methods (+0.013-0.318 in terms of R-precision; p<<0.01, t-test).

What prerequisite knowledge should students achieve a level of mastery before moving forward to learn subsequent coursewares? We study the extent to which the prerequisite relation between knowledge concepts in Massive Open Online Courses (MOOCs) can be inferred automatically. In particular, what kinds of information can be leverage to uncover the potential prerequisite relation between knowledge concepts. We first propose a representation learning-based method for learning latent representations of course concepts, and then investigate how different features capture the prerequisite relations between concepts. Our experiments on three datasets form Coursera show that the proposed method achieves significant improvements (+5.9-48.0% by F1-score) comparing with existing methods.