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.

Hallucination remains a critical challenge for multimodal large language models (MLLMs), undermining their reliability in real-world applications. While fine-grained hallucination detection (FHD) holds promise for enhancing high-quality vision-language data construction and model alignment through enriched feedback signals, automated solutions for this task have yet to be systematically explored. Inspired by the concept of “MLLM as a Judge”, we introduce MHALO, the first comprehensive benchmark specifically designed for evaluating MLLMs’ capability in performing token-level FHD. Our benchmark encompasses 12 distinct hallucination types spanning both multimodal perception and reasoning domains. Through extensive evaluations of 9 selected MLLMs, we reveal substantial performance limitations, with the leading model achieving an average F1_IoU of only 40.59%. To address this limitation, we develop HaloDet-4B, a specialized model trained on our curated training data, which significantly outperforms existing models. We hope the benchmark can provide valuable insights for future research on hallucination mitigation in MLLMs. The code and dataset will be publicly available.

The integration of sophisticated Vision-Language Models (VLMs) in vehicular systems is revolutionizing vehicle interaction and safety, performing tasks such as Visual Question Answering (VQA). However, a critical gap persists due to the lack of a comprehensive benchmark for multimodal VQA models in vehicular scenarios. To address this, we propose IntelliCockpitBench, a benchmark that encompasses diverse automotive scenarios. It includes images from front, side, and rear cameras, various road types, weather conditions, and interior views, integrating data from both moving and stationary states. Notably, all images and queries in the benchmark are verified for high levels of authenticity, ensuring the data accurately reflects real-world conditions. A sophisticated scoring methodology combining human and model-generated assessments enhances reliability and consistency. Our contributions include a diverse and authentic dataset for automotive VQA and a robust evaluation metric aligning human and machine assessments. All code and data can be found at https://github.com/Lane315/IntelliCockpitBench.

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.

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

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.

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.

Recent studies have achieved inspiring success in unsupervised grammar induction using masked language modeling (MLM) as the proxy task. Despite their high accuracy in identifying low-level structures, prior arts tend to struggle in capturing high-level structures like clauses, since the MLM task usually only requires information from local context. In this work, we revisit LM-based constituency parsing from a phrase-centered perspective. Inspired by the natural reading process of human, we propose to regularize the parser with phrases extracted by an unsupervised phrase tagger to help the LM model quickly manage low-level structures. For a better understanding of high-level structures, we propose a phrase-guided masking strategy for LM to emphasize more on reconstructing non-phrase words. We show that the initial phrase regularization serves as an effective bootstrap, and phrase-guided masking improves the identification of high-level structures. Experiments on the public benchmark with two different backbone models demonstrate the effectiveness and generality of our method.

As the excessive pre-training cost arouses the need to improve efficiency, considerable efforts have been made to train BERT progressively–start from an inferior but low-cost model and gradually increase the computational complexity. Our objective is to help advance the understanding of such Transformer growth and discover principles that guide progressive training. First, we find that similar to network architecture selection, Transformer growth also favors compound scaling. Specifically, while existing methods only conduct network growth in a single dimension, we observe that it is beneficial to use compound growth operators and balance multiple dimensions (e.g., depth, width, and input length of the model). Moreover, we explore alternative growth operators in each dimension via controlled comparison to give practical guidance for operator selection. In light of our analyses, the proposed method CompoundGrow speeds up BERT pre-training by 73.6% and 82.2% for the base and large models respectively while achieving comparable performances.

Walk-based models have shown their advantages in knowledge graph (KG) reasoning by achieving decent performance while providing interpretable decisions. However, the sparse reward signals offered by the KG during a traversal are often insufficient to guide a sophisticated walk-based reinforcement learning (RL) model. An alternate approach is to use traditional symbolic methods (e.g., rule induction), which achieve good performance but can be hard to generalize due to the limitation of symbolic representation. In this paper, we propose RuleGuider, which leverages high-quality rules generated by symbolic-based methods to provide reward supervision for walk-based agents. Experiments on benchmark datasets shows that RuleGuider clearly improves the performance of walk-based models without losing interpretability.

Despite of the recent success of collective entity linking (EL) methods, these “global” inference methods may yield sub-optimal results when the “all-mention coherence” assumption breaks, and often suffer from high computational cost at the inference stage, due to the complex search space. In this paper, we propose a simple yet effective solution, called Dynamic Context Augmentation (DCA), for collective EL, which requires only one pass through the mentions in a document. DCA sequentially accumulates context information to make efficient, collective inference, and can cope with different local EL models as a plug-and-enhance module. We explore both supervised and reinforcement learning strategies for learning the DCA model. Extensive experiments show the effectiveness of our model with different learning settings, base models, decision orders and attention mechanisms.

Recently, distant supervision has gained great success on Fine-grained Entity Typing (FET). Despite its efficiency in reducing manual labeling efforts, it also brings the challenge of dealing with false entity type labels, as distant supervision assigns labels in a context-agnostic manner. Existing works alleviated this issue with partial-label loss, but usually suffer from confirmation bias, which means the classifier fit a pseudo data distribution given by itself. In this work, we propose to regularize distantly supervised models with Compact Latent Space Clustering (CLSC) to bypass this problem and effectively utilize noisy data yet. Our proposed method first dynamically constructs a similarity graph of different entity mentions; infer the labels of noisy instances via label propagation. Based on the inferred labels, mention embeddings are updated accordingly to encourage entity mentions with close semantics to form a compact cluster in the embedding space, thus leading to better classification performance. Extensive experiments on standard benchmarks show that our CLSC model consistently outperforms state-of-the-art distantly supervised entity typing systems by a significant margin.

Many efforts have been made to facilitate natural language processing tasks with pre-trained language models (LMs), and brought significant improvements to various applications. To fully leverage the nearly unlimited corpora and capture linguistic information of multifarious levels, large-size LMs are required; but for a specific task, only parts of these information are useful. Such large-sized LMs, even in the inference stage, may cause heavy computation workloads, making them too time-consuming for large-scale applications. Here we propose to compress bulky LMs while preserving useful information with regard to a specific task. As different layers of the model keep different information, we develop a layer selection method for model pruning using sparsity-inducing regularization. By introducing the dense connectivity, we can detach any layer without affecting others, and stretch shallow and wide LMs to be deep and narrow. In model training, LMs are learned with layer-wise dropouts for better robustness. Experiments on two benchmark datasets demonstrate the effectiveness of our method.

Recent advances in deep neural models allow us to build reliable named entity recognition (NER) systems without handcrafting features. However, such methods require large amounts of manually-labeled training data. There have been efforts on replacing human annotations with distant supervision (in conjunction with external dictionaries), but the generated noisy labels pose significant challenges on learning effective neural models. Here we propose two neural models to suit noisy distant supervision from the dictionary. First, under the traditional sequence labeling framework, we propose a revised fuzzy CRF layer to handle tokens with multiple possible labels. After identifying the nature of noisy labels in distant supervision, we go beyond the traditional framework and propose a novel, more effective neural model AutoNER with a new Tie or Break scheme. In addition, we discuss how to refine distant supervision for better NER performance. Extensive experiments on three benchmark datasets demonstrate that AutoNER achieves the best performance when only using dictionaries with no additional human effort, and delivers competitive results with state-of-the-art supervised benchmarks.

We present a novel end-to-end reinforcement learning approach to automatic taxonomy induction from a set of terms. While prior methods treat the problem as a two-phase task (i.e.,, detecting hypernymy pairs followed by organizing these pairs into a tree-structured hierarchy), we argue that such two-phase methods may suffer from error propagation, and cannot effectively optimize metrics that capture the holistic structure of a taxonomy. In our approach, the representations of term pairs are learned using multiple sources of information and used to determine which term to select and where to place it on the taxonomy via a policy network. All components are trained in an end-to-end manner with cumulative rewards, measured by a holistic tree metric over the training taxonomies. Experiments on two public datasets of different domains show that our approach outperforms prior state-of-the-art taxonomy induction methods up to 19.6% on ancestor F1.