Large language models (LLMs) are increasingly applied in specialized domains such as finance and healthcare, where they introduce unique safety risks. Domain-specific datasets of harmful prompts remain scarce and still largely rely on manual construction; public datasets mainly focus on explicit harmful prompts, which modern LLM defenses can often detect and refuse. In contrast, implicit harmful prompts—expressed through indirect domain knowledge—are harder to detect and better reflect real-world threats. We identify two challenges: transforming domain knowledge into actionable constraints and increasing the implicitness of generated harmful prompts. To address them, we propose an end-to-end framework that first performs knowledge-graph-guided harmful prompt generation to systematically produce domain-relevant prompts, and then applies two-strategy obfuscation rewriting to convert explicit harmful prompts into implicit variants via direct and context-enhanced rewriting. This framework yields high-quality datasets combining strong domain relevance with implicitness, enabling more realistic red-teaming and advancing LLM safety research. We release our code and datasets on GitHub.

Scaling data and model size has been proven effective for boosting the performance of large language models. In addition to training-time scaling, recent studies have revealed that increasing test-time computational resources can further improve performance. In this work, we introduce Aggregation Fine-Tuning (AFT), a supervised fine-tuning paradigm where the model learns to synthesize multiple draft responses, referred to as proposals, into a single, refined answer, termed aggregation. At inference time, we apply a propose-and-aggregate strategy that iteratively generates and aggregates proposals, effectively scaling inference-time computation without relying on external guidance such as a reward model. Empirical results across benchmark datasets demonstrate that AFT-trained models achieve substantial gains with test-time scaling, outperforming best-of-N baselines while eliminating the need for external reward signals. Notably, an AFT model, fine-tuned from Llama3.1-8B-Base with only 64k data, achieves a 41.3% LC win rate on AlpacaEval 2, surpassing significantly larger LLMs such as Llama3.1-405B-Instruct and GPT-4. By combining sequential refinement and parallel sampling, the propose-and-aggregate framework scales inference-time computation in a flexible manner.

Large language models (LLMs) often struggle with complex reasoning tasks due to their limitations in addressing the vast reasoning space and inherent ambiguities of natural language. We propose the Mixture-of-Search-Agents (MOSA) paradigm, a novel approach leveraging the collective expertise of multiple LLMs to enhance search-based reasoning. MOSA integrates diverse reasoning pathways by combining independent exploration with iterative refinement among LLMs, mitigating the limitations of single-model approaches. Using Monte Carlo Tree Search (MCTS) as a backbone, MOSA enables multiple agents to propose and aggregate reasoning steps, resulting in improved accuracy. Our comprehensive evaluation across four reasoning benchmarks demonstrates MOSA’s consistent performance improvements over single-agent and other multi-agent baselines, particularly in complex mathematical and commonsense reasoning tasks.

As large language models (LLMs) are increasingly applied to legal domain-specific tasks, evaluating their ability to perform legal work in real-world settings has become essential. However, existing legal benchmarks rely on simplified and highly standardized tasks, failing to capture the ambiguity, complexity, and reasoning demands of real legal practice. Moreover, prior evaluations often adopt coarse, single-dimensional metrics and do not explicitly assess fine-grained legal reasoning. To address these limitations, we introduce PLawBench, a Practical Law Benchmark designed to evaluate LLMs in realistic legal practice scenarios. Grounded in real-world legal workflows, PLawBench models the core processes of legal practitioners through three task categories: public legal consultation, practical case analysis, and legal document generation. These tasks assess a model’s ability to identify legal issues and key facts, perform structured legal reasoning, and generate legally coherent documents. PLawBench comprises 850 questions across 13 practical legal scenarios, with each question accompanied by expert-designed evaluation rubrics, resulting in approximately 12,500 rubric items for fine-grained assessment. Using an LLM-based evaluator aligned with human expert judgments, we evaluate 10 state-of-the-art LLMs. Experimental results show that none achieves strong performance on PLawBench, revealing substantial limitations in the fine-grained legal reasoning capabilities of current LLMs and highlighting important directions for future evaluation and development of legal LLMs. Data is available at: https://anonymous.4open.science/r/PLawbench-B524/.

Two lines of approaches are adopted for complex reasoning with LLMs. One line of work prompts LLMs with various reasoning structures, while the structural outputs can be naturally regarded as intermediate reasoning steps. Another line of work adopt LLM-free declarative solvers to do the reasoning task, rendering higher reasoning accuracy but lacking interpretability due to the black-box nature of the solvers. Aiming to resolve the trade-off between answer accuracy and interpretability, we present a simple extension to the latter line of work. Specifically, we showcase that the intermediate search logs generated by Prolog interpreters can be accessed and interpreted into human-readable reasoning proofs. As long as LLMs correctly translate problem descriptions into Prolog representations, the corresponding reasoning proofs are ensured to be causal and reliable. On two logical reasoning and one arithmetic reasoning datasets, our framework obtains significant improvements in terms of both answer accuracy and reasoning proof accuracy. We released our code at https://github.com/DAMO-NLP-SG/CaRing for future research regarding better reasoning proofs using LLMs.

The rise of Large Language Models (LLMs) has reshaped machine translation (MT), but multilingual MT still relies heavily on parallel data for supervised fine-tuning (SFT), facing challenges like data scarcity for low-resource languages and catastrophic forgetting. To address these issues, we propose TRANS-ZERO, a self-play framework that leverages only monolingual data and the intrinsic multilingual knowledge of LLM. TRANS-ZERO combines Genetic Monte-Carlo Tree Search (G-MCTS) with preference optimization, achieving strong translation performance that rivals supervised methods. Experiments demonstrate that this approach not only matches the performance of models trained on large-scale parallel data but also excels in non-English translation directions. Further analysis reveals that G-MCTS itself significantly enhances translation quality by exploring semantically consistent candidates through iterative translations, providing a robust foundation for the framework’s success.

While Large Language Models (LLMs) demonstrate impressive capabilities, they still struggle with generating factually incorrect content (i.e., hallucinations). A promising approach to mitigate this issue is enabling models to express uncertainty when unsure. Previous research on uncertainty modeling has primarily focused on short-form QA, but real-world applications often require much longer responses. In this work, we introduce the task of Long-form Generation with Uncertainty (LoGU). We identify two key challenges: Uncertainty Suppression, where models hesitate to express uncertainty, and Uncertainty Misalignment, where models convey uncertainty inaccurately. To tackle these challenges, we propose a refinement-based data collection framework and a two-stage training pipeline. Our framework adopts a divide-and-conquer strategy, refining uncertainty based on atomic claims. The collected data are then used in training through supervised fine-tuning (SFT) and direct preference optimization (DPO) to enhance uncertainty expression. Extensive experiments on three long-form instruction following datasets show that our method significantly improves accuracy, reduces hallucinations, and maintains the comprehensiveness of responses.

Large language models (LLMs) often suffer from hallucinations, posing significant challenges for real-world applications. Confidence calibration, as an effective indicator of hallucination, is thus essential to enhance the trustworthiness of LLMs. Prior work mainly focuses on short-form tasks using a single response-level score (macro calibration), which is insufficient for long-form outputs that may contain both accurate and inaccurate claims. In this work, we systematically study atomic calibration, which evaluates factuality calibration at a fine-grained level by decomposing long responses into atomic claims. We further categorize existing confidence elicitation methods into discriminative and generative types, and propose two new confidence fusion strategies to improve calibration. Our experiments demonstrate that LLMs exhibit poorer calibration at the atomic level during long-form generation. More importantly, atomic calibration uncovers insightful patterns regarding the alignment of confidence methods and the changes of confidence throughout generation. This sheds light on future research directions for confidence estimation in long-form generation.

Large language models (LLMs) have demonstrated strong machine translation capabilities for English-centric language pairs but underperform in direct non-English (x2x) translation. This work addresses this limitation through a synthetic data generation framework that leverages models’ established English-to-x (en2x) capabilities. By extending English parallel corpora into omnidirectional datasets and developing an English-referenced quality evaluation proxy, we enable effective collection of high-quality x2x training data. Combined with preference-based optimization, our method achieves significant improvement across 72 x2x directions for widely used LLMs, while generalizing to enhance en2x performance. The results demonstrate that strategic exploitation of English-centric strengths can bootstrap comprehensive multilingual translation capabilities in LLMs.

Despite the remarkable achievements of large language models (LLMs) in various tasks, there remains a linguistic bias that favors high-resource languages, such as English, often at the expense of low-resource and regional languages. To address this imbalance, we introduce SeaLLMs, an innovative series of language models that specifically focuses on Southeast Asian (SEA) languages. SeaLLMs are built upon popular English-centric models through continued pre-training with an extended vocabulary, specialized instruction and alignment tuning to better capture the intricacies of regional languages. This allows them to respect and reflect local cultural norms, customs, stylistic preferences, and legal considerations. Our comprehensive evaluation demonstrates that SeaLLM models exhibit superior performance across a wide spectrum of linguistic tasks and assistant-style instruction-following capabilities relative to comparable open-source models. Moreover, they outperform ChatGPT-3.5 in non-Latin languages, such as Thai, Khmer, Lao, and Burmese, by large margins while remaining lightweight and cost-effective to operate.

Iterative preference learning, though yielding superior performances, requires online annotated preference labels. In this work, we study strategies to save annotation budgets while achieving competitive or even better performances for iterative preference learning. Built on intuitions from active learning, we empirically show that annotating those response pairs with small margins is generally better than large or random. Besides, experiments under the multi-iteration scenario suggest allocating more annotation budgets in the earlier iterations rather than later ones.

Grammatical error correction systems improve written communication by detecting and correcting language mistakes. To help language learners better understand why the GEC system makes a certain correction, the causes of errors (evidence words) and the corresponding error types are two key factors. To enhance GEC systems with explanations, we introduce EXPECT, a large dataset annotated with evidence words and grammatical error types. We propose several baselines and anlysis to understand this task. Furthermore, human evaluation verifies our explainable GEC system’s explanations can assist second-language learners in determining whether to accept a correction suggestion and in understanding the associated grammar rule.

This paper describes our system used in SemEval-2023 Task-1: Visual Word Sense Disambiguation (VWSD). The VWSD task is to identify the correct image that corresponds to an ambiguous target word given limited textual context. To reduce word ambiguity and enhance image selection, we proposed several text augmentation techniques, such as prompting, WordNet synonyms, and text generation. We experimented with different vision-language pre-trained models to capture the joint features of the augmented text and image. Our approach achieved the best performance using a combination of GPT-3 text generation and the CLIP model. On the multilingual test sets, our system achieved an average hit rate (at top-1) of 51.11 and a mean reciprocal rank of 65.69.

In recent years, deep neural networks (DNNs) have achieved state-of-the-art performance on a wide range of tasks. However, limitations in interpretability have hindered their applications in the real world. This work proposes to interpret neural networks by linear decomposition and finds that the ReLU-activated Transformer can be considered as a linear model on a single input. We further leverage the linearity of the model and propose a linear decomposition of the model output to generate local explanations. Our evaluation of sentiment classification and machine translation shows that our method achieves competitive performance in efficiency and fidelity of explanation. In addition, we demonstrate the potential of our approach in applications with examples of error analysis on multiple tasks.

Our physical world is constantly evolving over time, rendering challenges for pre-trained language models to understand and reason over the temporal contexts of texts. Existing work focuses on strengthening the direct association between a piece of text and its time-stamp. However, the knowledge-time association is usually insufficient for the downstream tasks that require reasoning over temporal dependencies between knowledge. In this work, we make use of the underlying nature of time, all temporally-scoped sentences are strung together through a one-dimensional time axis, and suggest creating a graph structure based on the relative placements of events along the time axis. Inspired by the graph view, we propose RemeMo ( ̲Relative Ti ̲me  ̲Modeling), which explicitly connects all temporally-scoped facts by modeling the time relations between any two sentences. Experimental results show that RemeMo outperforms the baseline T5 on multiple temporal question answering datasets under various settings. Further analysis suggests that RemeMo is especially good at modeling long-range complex temporal dependencies.

Thanks to the strong representation power of neural encoders, neural chart-based parsers have achieved highly competitive performance by using local features. Recently, it has been shown that non-local features in CRF structures lead to improvements. In this paper, we investigate injecting non-local features into the training process of a local span-based parser, by predicting constituent n-gram non-local patterns and ensuring consistency between non-local patterns and local constituents. Results show that our simple method gives better results than the self-attentive parser on both PTB and CTB. Besides, our method achieves state-of-the-art BERT-based performance on PTB (95.92 F1) and strong performance on CTB (92.31 F1). Our parser also outperforms the self-attentive parser in multi-lingual and zero-shot cross-domain settings.

Neural constituency parsers have reached practical performance on news-domain benchmarks. However, their generalization ability to other domains remains weak. Existing findings on cross-domain constituency parsing are only made on a limited number of domains. Tracking this, we manually annotate a high-quality constituency treebank containing five domains. We analyze challenges to open-domain constituency parsing using a set of linguistic features on various strong constituency parsers. Primarily, we find that 1) BERT significantly increases parsers’ cross-domain performance by reducing their sensitivity on the domain-variant features.2) Compared with single metrics such as unigram distribution and OOV rate, challenges to open-domain constituency parsing arise from complex features, including cross-domain lexical and constituent structure variations.

Abstract Meaning Representation (AMR) parsing aims to predict an AMR graph from textual input. Recently, there has been notable growth in AMR parsing performance. However, most existing work focuses on improving the performance in the specific domain, ignoring the potential domain dependence of AMR parsing systems. To address this, we extensively evaluate five representative AMR parsers on five domains and analyze challenges to cross-domain AMR parsing. We observe that challenges to cross-domain AMR parsing mainly arise from the distribution shift of words and AMR concepts. Based on our observation, we investigate two approaches to reduce the domain distribution divergence of text and AMR features, respectively. Experimental results on two out-of-domain test sets show the superiority of our method.

Although showing promising values to downstream applications, generating question and answer together is under-explored. In this paper, we introduce a novel task that targets question-answer pair generation from visual images. It requires not only generating diverse question-answer pairs but also keeping the consistency of them. We study different generation paradigms for this task and propose three models: the pipeline model, the joint model, and the sequential model. We integrate variational inference into these models to achieve diversity and consistency. We also propose region representation scaling and attention alignment to improve the consistency further. We finally devise an evaluator as a quantitative metric for consistency. We validate our approach on two benchmarks, VQA2.0 and Visual-7w, by automatically and manually evaluating diversity and consistency. Experimental results show the effectiveness of our models: they can generate diverse or consistent pairs. Moreover, this task can be used to improve visual question generation and visual question answering.

Sentiment analysis provides a useful overview of customer review contents. Many review websites allow a user to enter a summary in addition to a full review. Intuitively, summary information may give additional benefit for review sentiment analysis. In this paper, we conduct a study to exploit methods for better use of summary information. We start by finding out that the sentimental signal distribution of a review and that of its corresponding summary are in fact complementary to each other. We thus explore various architectures to better guide the interactions between the two and propose a hierarchically-refined review-centric attention model. Empirical results show that our review-centric model can make better use of user-written summaries for review sentiment analysis, and is also more effective compared to existing methods when the user summary is replaced with summary generated by an automatic summarization system.

Deep learning has led to significant improvement in text summarization with various methods investigated and improved ROUGE scores reported over the years. However, gaps still exist between summaries produced by automatic summarizers and human professionals. Aiming to gain more understanding of summarization systems with respect to their strengths and limits on a fine-grained syntactic and semantic level, we consult the Multidimensional Quality Metric (MQM) and quantify 8 major sources of errors on 10 representative summarization models manually. Primarily, we find that 1) under similar settings, extractive summarizers are in general better than their abstractive counterparts thanks to strength in faithfulness and factual-consistency; 2) milestone techniques such as copy, coverage and hybrid extractive/abstractive methods do bring specific improvements but also demonstrate limitations; 3) pre-training techniques, and in particular sequence-to-sequence pre-training, are highly effective for improving text summarization, with BART giving the best results.

Traditional approaches to the task of ACE event extraction usually depend on manually annotated data, which is often laborious to create and limited in size. Therefore, in addition to the difficulty of event extraction itself, insufficient training data hinders the learning process as well. To promote event extraction, we first propose an event extraction model to overcome the roles overlap problem by separating the argument prediction in terms of roles. Moreover, to address the problem of insufficient training data, we propose a method to automatically generate labeled data by editing prototypes and screen out generated samples by ranking the quality. Experiments on the ACE2005 dataset demonstrate that our extraction model can surpass most existing extraction methods. Besides, incorporating our generation method exhibits further significant improvement. It obtains new state-of-the-art results on the event extraction task, including pushing the F1 score of trigger classification to 81.1%, and the F1 score of argument classification to 58.9%.