Xiaozhi Wang

Feng Yao, Chaojun Xiao, Xiaozhi Wang et al. · tsinghua

Recognizing facts is the most fundamental step in making judgments, hence detecting events in the legal documents is important to legal case analysis tasks. However, existing Legal Event Detection (LED) datasets only concern incomprehensive event types and have limited annotated data, which restricts the development of LED methods and their downstream applications. To alleviate these issues, we present LEVEN a large-scale Chinese LEgal eVENt detection dataset, with 8,116 legal documents and 150,977 human-annotated event mentions in 108 event types. Not only charge-related events, LEVEN also covers general events, which are critical for legal case understanding but neglected in existing LED datasets. To our knowledge, LEVEN is the largest LED dataset and has dozens of times the data scale of others, which shall significantly promote the training and evaluation of LED methods. The results of extensive experiments indicate that LED is challenging and needs further effort. Moreover, we simply utilize legal events as side information to promote downstream applications. The method achieves improvements of average 2.2 points precision in low-resource judgment prediction, and 1.5 points mean average precision in unsupervised case retrieval, which suggests the fundamentality of LED. The source code and dataset can be obtained from https://github.com/thunlp/LEVEN.

Xiaozhi Wang, Yulin Chen, Ning Ding et al. · tencent-ai, tsinghua

The diverse relationships among real-world events, including coreference, temporal, causal, and subevent relations, are fundamental to understanding natural languages. However, two drawbacks of existing datasets limit event relation extraction (ERE) tasks: (1) Small scale. Due to the annotation complexity, the data scale of existing datasets is limited, which cannot well train and evaluate data-hungry models. (2) Absence of unified annotation. Different types of event relations naturally interact with each other, but existing datasets only cover limited relation types at once, which prevents models from taking full advantage of relation interactions. To address these issues, we construct a unified large-scale human-annotated ERE dataset MAVEN-ERE with improved annotation schemes. It contains 103,193 event coreference chains, 1,216,217 temporal relations, 57,992 causal relations, and 15,841 subevent relations, which is larger than existing datasets of all the ERE tasks by at least an order of magnitude. Experiments show that ERE on MAVEN-ERE is quite challenging, and considering relation interactions with joint learning can improve performances. The dataset and source codes can be obtained from https://github.com/THU-KEG/MAVEN-ERE.

Jifan Yu, Xiaozhi Wang, Shangqing Tu et al. · tsinghua

The unprecedented performance of large language models (LLMs) necessitates improvements in evaluations. Rather than merely exploring the breadth of LLM abilities, we believe meticulous and thoughtful designs are essential to thorough, unbiased, and applicable evaluations. Given the importance of world knowledge to LLMs, we construct a Knowledge-oriented LLM Assessment benchmark (KoLA), in which we carefully design three crucial factors: (1) For \textbf{ability modeling}, we mimic human cognition to form a four-level taxonomy of knowledge-related abilities, covering $19$ tasks. (2) For \textbf{data}, to ensure fair comparisons, we use both Wikipedia, a corpus prevalently pre-trained by LLMs, along with continuously collected emerging corpora, aiming to evaluate the capacity to handle unseen data and evolving knowledge. (3) For \textbf{evaluation criteria}, we adopt a contrastive system, including overall standard scores for better numerical comparability across tasks and models and a unique self-contrast metric for automatically evaluating knowledge-creating ability. We evaluate $28$ open-source and commercial LLMs and obtain some intriguing findings. The KoLA dataset and open-participation leaderboard are publicly released at https://kola.xlore.cn and will be continuously updated to provide references for developing LLMs and knowledge-related systems.

Ji Qi, Jifan Yu, Teng Tu et al. · pku, tsinghua

Despite the recent emergence of video captioning models, how to generate vivid, fine-grained video descriptions based on the background knowledge (i.e., long and informative commentary about the domain-specific scenes with appropriate reasoning) is still far from being solved, which however has great applications such as automatic sports narrative. In this paper, we present GOAL, a benchmark of over 8.9k soccer video clips, 22k sentences, and 42k knowledge triples for proposing a challenging new task setting as Knowledge-grounded Video Captioning (KGVC). Moreover, we conduct experimental adaption of existing methods to show the difficulty and potential directions for solving this valuable and applicable task. Our data and code are available at https://github.com/THU-KEG/goal.

Xuekang Wang, Zhuoyuan Hao, Shuo Hou et al.

Rubric-based reinforcement learning (RL) uses an LLM-as-a-Judge (LaaJ) to score model outputs according to rubrics as rewards. However, policy models may exploit latent biases in the judge, leading to reward hacking and ineffective or unsafe training outcomes. In real-world rubric-based RL, such hacking behaviors are often subtle and entangled with multiple judge biases, making them difficult to analyze, detect, and mitigate. In this paper, we introduce CHERRL, a controllable hacking environment for rubric-based RL. By injecting known biases into LaaJ, CHERRL enables stable reproduction of reward hacking, explicit observation of reward divergence, and precise identification of hacking onset. This provides a clean experimental testbed for studying the mechanisms and mitigations of reward hacking in rubric-based RL. To demonstrate its utility, we analyze different judge biases from the perspectives of discoverability and exploitability, and explore an agent-based system for automatically detecting reward hacking onset from training logs. The code and environment are publicly available at https://github.com/THUAIS-Lab/CHERRL.

Xiaozhi Wang, Kaiyue Wen, Zhengyan Zhang et al. · tsinghua

Transformer-based pre-trained language models have demonstrated superior performance on various natural language processing tasks. However, it remains unclear how the skills required to handle these tasks distribute among model parameters. In this paper, we find that after prompt tuning for specific tasks, the activations of some neurons within pre-trained Transformers are highly predictive of the task labels. We dub these neurons skill neurons and confirm they encode task-specific skills by finding that: (1) Skill neurons are crucial for handling tasks. Performances of pre-trained Transformers on a task significantly drop when corresponding skill neurons are perturbed. (2) Skill neurons are task-specific. Similar tasks tend to have similar distributions of skill neurons. Furthermore, we demonstrate the skill neurons are most likely generated in pre-training rather than fine-tuning by showing that the skill neurons found with prompt tuning are also crucial for other fine-tuning methods freezing neuron weights, such as the adapter-based tuning and BitFit. We also explore the applications of skill neurons, including accelerating Transformers with network pruning and building better transferability indicators. These findings may promote further research on understanding Transformers. The source code can be obtained from https://github.com/THU-KEG/Skill-Neuron.

Ning Ding, Yujia Qin, Guang Yang et al. · tsinghua

Despite the success, the process of fine-tuning large-scale PLMs brings prohibitive adaptation costs. In fact, fine-tuning all the parameters of a colossal model and retaining separate instances for different tasks are practically infeasible. This necessitates a new branch of research focusing on the parameter-efficient adaptation of PLMs, dubbed as delta tuning in this paper. In contrast with the standard fine-tuning, delta tuning only fine-tunes a small portion of the model parameters while keeping the rest untouched, largely reducing both the computation and storage costs. Recent studies have demonstrated that a series of delta tuning methods with distinct tuned parameter selection could achieve performance on a par with full-parameter fine-tuning, suggesting a new promising way of stimulating large-scale PLMs. In this paper, we first formally describe the problem of delta tuning and then comprehensively review recent delta tuning approaches. We also propose a unified categorization criterion that divide existing delta tuning methods into three groups: addition-based, specification-based, and reparameterization-based methods. Though initially proposed as an efficient method to steer large models, we believe that some of the fascinating evidence discovered along with delta tuning could help further reveal the mechanisms of PLMs and even deep neural networks. To this end, we discuss the theoretical principles underlying the effectiveness of delta tuning and propose frameworks to interpret delta tuning from the perspective of optimization and optimal control, respectively. Furthermore, we provide a holistic empirical study of representative methods, where results on over 100 NLP tasks demonstrate a comprehensive performance comparison of different approaches. The experimental results also cover the analysis of combinatorial, scaling and transferable properties of delta tuning.

Chenglei Si, Zhengyan Zhang, Yingfa Chen et al. · tsinghua

For many real-world applications, the user-generated inputs usually contain various noises due to speech recognition errors caused by linguistic variations1 or typographical errors (typos). Thus, it is crucial to test model performance on data with realistic input noises to ensure robustness and fairness. However, little study has been done to construct such benchmarks for Chinese, where various language-specific input noises happen in the real world. In order to fill this important gap, we construct READIN: a Chinese multi-task benchmark with REalistic And Diverse Input Noises. READIN contains four diverse tasks and requests annotators to re-enter the original test data with two commonly used Chinese input methods: Pinyin input and speech input. We designed our annotation pipeline to maximize diversity, for example by instructing the annotators to use diverse input method editors (IMEs) for keyboard noises and recruiting speakers from diverse dialectical groups for speech noises. We experiment with a series of strong pretrained language models as well as robust training methods, we find that these models often suffer significant performance drops on READIN even with robustness methods like data augmentation. As the first large-scale attempt in creating a benchmark with noises geared towards user-generated inputs, we believe that READIN serves as an important complement to existing Chinese NLP benchmarks. The source code and dataset can be obtained from https://github.com/thunlp/READIN.

Xiaozhi Wang, Hao Peng, Yong Guan et al. · tsinghua

Understanding events in texts is a core objective of natural language understanding, which requires detecting event occurrences, extracting event arguments, and analyzing inter-event relationships. However, due to the annotation challenges brought by task complexity, a large-scale dataset covering the full process of event understanding has long been absent. In this paper, we introduce MAVEN-Arg, which augments MAVEN datasets with event argument annotations, making the first all-in-one dataset supporting event detection, event argument extraction (EAE), and event relation extraction. As an EAE benchmark, MAVEN-Arg offers three main advantages: (1) a comprehensive schema covering 162 event types and 612 argument roles, all with expert-written definitions and examples; (2) a large data scale, containing 98,591 events and 290,613 arguments obtained with laborious human annotation; (3) the exhaustive annotation supporting all task variants of EAE, which annotates both entity and non-entity event arguments in document level. Experiments indicate that MAVEN-Arg is quite challenging for both fine-tuned EAE models and proprietary large language models (LLMs). Furthermore, to demonstrate the benefits of an all-in-one dataset, we preliminarily explore a potential application, future event prediction, with LLMs. MAVEN-Arg and codes can be obtained from https://github.com/THU-KEG/MAVEN-Argument.

Hao Peng, Xiaozhi Wang, Shengding Hu et al. · tsinghua

Conceptual knowledge is fundamental to human cognition and knowledge bases. However, existing knowledge probing works only focus on evaluating factual knowledge of pre-trained language models (PLMs) and ignore conceptual knowledge. Since conceptual knowledge often appears as implicit commonsense behind texts, designing probes for conceptual knowledge is hard. Inspired by knowledge representation schemata, we comprehensively evaluate conceptual knowledge of PLMs by designing three tasks to probe whether PLMs organize entities by conceptual similarities, learn conceptual properties, and conceptualize entities in contexts, respectively. For the tasks, we collect and annotate 24k data instances covering 393 concepts, which is COPEN, a COnceptual knowledge Probing bENchmark. Extensive experiments on different sizes and types of PLMs show that existing PLMs systematically lack conceptual knowledge and suffer from various spurious correlations. We believe this is a critical bottleneck for realizing human-like cognition in PLMs. COPEN and our codes are publicly released at https://github.com/THU-KEG/COPEN.

Hao Peng, Xiaozhi Wang, Feng Yao et al. · tsinghua

Event extraction (EE) is a crucial task aiming at extracting events from texts, which includes two subtasks: event detection (ED) and event argument extraction (EAE). In this paper, we check the reliability of EE evaluations and identify three major pitfalls: (1) The data preprocessing discrepancy makes the evaluation results on the same dataset not directly comparable, but the data preprocessing details are not widely noted and specified in papers. (2) The output space discrepancy of different model paradigms makes different-paradigm EE models lack grounds for comparison and also leads to unclear mapping issues between predictions and annotations. (3) The absence of pipeline evaluation of many EAE-only works makes them hard to be directly compared with EE works and may not well reflect the model performance in real-world pipeline scenarios. We demonstrate the significant influence of these pitfalls through comprehensive meta-analyses of recent papers and empirical experiments. To avoid these pitfalls, we suggest a series of remedies, including specifying data preprocessing, standardizing outputs, and providing pipeline evaluation results. To help implement these remedies, we develop a consistent evaluation framework OMNIEVENT, which can be obtained from https://github.com/THU-KEG/OmniEvent.

Hao Peng, Xiaozhi Wang, Feng Yao et al. · tsinghua

Event understanding aims at understanding the content and relationship of events within texts, which covers multiple complicated information extraction tasks: event detection, event argument extraction, and event relation extraction. To facilitate related research and application, we present an event understanding toolkit OmniEvent, which features three desiderata: (1) Comprehensive. OmniEvent supports mainstream modeling paradigms of all the event understanding tasks and the processing of 15 widely-used English and Chinese datasets. (2) Fair. OmniEvent carefully handles the inconspicuous evaluation pitfalls reported in Peng et al. (2023), which ensures fair comparisons between different models. (3) Easy-to-use. OmniEvent is designed to be easily used by users with varying needs. We provide off-the-shelf models that can be directly deployed as web services. The modular framework also enables users to easily implement and evaluate new event understanding models with OmniEvent. The toolkit (https://github.com/THU-KEG/OmniEvent) is publicly released along with the demonstration website and video (https://omnievent.xlore.cn/).

Shangqing Tu, Chunyang Li, Jifan Yu et al. · tsinghua

ChatGPT has achieved great success and can be considered to have acquired an infrastructural status. There are abundant works for evaluating ChatGPT on benchmarks. However, existing benchmarks encounter two challenges: (1) Disregard for periodical evaluation and (2) Lack of fine-grained features. In this paper, we construct ChatLog, an ever-updating dataset with large-scale records of diverse long-form ChatGPT responses for 21 NLP benchmarks from March, 2023 to now. We conduct a comprehensive performance evaluation to find that most capabilities of ChatGPT improve over time except for some abilities, and there exists a step-wise evolving pattern of ChatGPT. We further analyze the inherent characteristics of ChatGPT by extracting the knowledge and linguistic features. We find some stable features that stay unchanged and apply them on the detection of ChatGPT-generated texts to improve the robustness of cross-version detection. We will continuously maintain our project at \url{https://github.com/THU-KEG/ChatLog/}.

Chunyang Li, Hao Peng, Xiaozhi Wang et al. · tsinghua

Event Factuality Detection (EFD) task determines the factuality of textual events, i.e., classifying whether an event is a fact, possibility, or impossibility, which is essential for faithfully understanding and utilizing event knowledge. However, due to the lack of high-quality large-scale data, event factuality detection is under-explored in event understanding research, which limits the development of EFD community. To address these issues and provide faithful event understanding, we introduce MAVEN-Fact, a large-scale and high-quality EFD dataset based on the MAVEN dataset. MAVEN-Fact includes factuality annotations of 112,276 events, making it the largest EFD dataset. Extensive experiments demonstrate that MAVEN-Fact is challenging for both conventional fine-tuned models and large language models (LLMs). Thanks to the comprehensive annotations of event arguments and relations in MAVEN, MAVEN-Fact also supports some further analyses and we find that adopting event arguments and relations helps in event factuality detection for fine-tuned models but does not benefit LLMs. Furthermore, we preliminarily study an application case of event factuality detection and find it helps in mitigating event-related hallucination in LLMs. Our dataset and codes can be obtained from \url{https://github.com/lcy2723/MAVEN-FACT}

Sha Yuan, Hanyu Zhao, Shuai Zhao et al. · bytedance, pku

With the rapid development of deep learning, training Big Models (BMs) for multiple downstream tasks becomes a popular paradigm. Researchers have achieved various outcomes in the construction of BMs and the BM application in many fields. At present, there is a lack of research work that sorts out the overall progress of BMs and guides the follow-up research. In this paper, we cover not only the BM technologies themselves but also the prerequisites for BM training and applications with BMs, dividing the BM review into four parts: Resource, Models, Key Technologies and Application. We introduce 16 specific BM-related topics in those four parts, they are Data, Knowledge, Computing System, Parallel Training System, Language Model, Vision Model, Multi-modal Model, Theory&Interpretability, Commonsense Reasoning, Reliability&Security, Governance, Evaluation, Machine Translation, Text Generation, Dialogue and Protein Research. In each topic, we summarize clearly the current studies and propose some future research directions. At the end of this paper, we conclude the further development of BMs in a more general view.

Yushi Bai, Jiahao Ying, Yixin Cao et al. · tsinghua

Numerous benchmarks have been established to assess the performance of foundation models on open-ended question answering, which serves as a comprehensive test of a model's ability to understand and generate language in a manner similar to humans. Most of these works focus on proposing new datasets, however, we see two main issues within previous benchmarking pipelines, namely testing leakage and evaluation automation. In this paper, we propose a novel benchmarking framework, Language-Model-as-an-Examiner, where the LM serves as a knowledgeable examiner that formulates questions based on its knowledge and evaluates responses in a reference-free manner. Our framework allows for effortless extensibility as various LMs can be adopted as the examiner, and the questions can be constantly updated given more diverse trigger topics. For a more comprehensive and equitable evaluation, we devise three strategies: (1) We instruct the LM examiner to generate questions across a multitude of domains to probe for a broad acquisition, and raise follow-up questions to engage in a more in-depth assessment. (2) Upon evaluation, the examiner combines both scoring and ranking measurements, providing a reliable result as it aligns closely with human annotations. (3) We additionally propose a decentralized Peer-examination method to address the biases in a single examiner. Our data and benchmarking results are available at: http://lmexam.xlore.cn.

Chaojun Xiao, Zhengyan Zhang, Chenyang Song et al. · tencent-ai, tsinghua

Advancements in LLMs have recently unveiled challenges tied to computational efficiency and continual scalability due to their requirements of huge parameters, making the applications and evolution of these models on devices with limited computation resources and scenarios requiring various abilities increasingly cumbersome. Inspired by modularity within the human brain, there is a growing tendency to decompose LLMs into numerous functional modules, allowing for inference with part of modules and dynamic assembly of modules to tackle complex tasks, such as mixture-of-experts. To highlight the inherent efficiency and composability of the modular approach, we coin the term brick to represent each functional module, designating the modularized structure as configurable foundation models. In this paper, we offer a comprehensive overview and investigation of the construction, utilization, and limitation of configurable foundation models. We first formalize modules into emergent bricks - functional neuron partitions that emerge during the pre-training phase, and customized bricks - bricks constructed via additional post-training to improve the capabilities and knowledge of LLMs. Based on diverse functional bricks, we further present four brick-oriented operations: retrieval and routing, merging, updating, and growing. These operations allow for dynamic configuration of LLMs based on instructions to handle complex tasks. To verify our perspective, we conduct an empirical analysis on widely-used LLMs. We find that the FFN layers follow modular patterns with functional specialization of neurons and functional neuron partitions. Finally, we highlight several open issues and directions for future research. Overall, this paper aims to offer a fresh modular perspective on existing LLM research and inspire the future creation of more efficient and scalable foundational models.

Hao Peng, Xiaozhi Wang, Jianhui Chen et al. · tsinghua

In-context learning (ICL) has become the default method for using large language models (LLMs), making the exploration of its limitations and understanding the underlying causes crucial. In this paper, we find that ICL falls short of handling specification-heavy tasks, which are tasks with complicated and extensive task specifications, requiring several hours for ordinary humans to master, such as traditional information extraction tasks. The performance of ICL on these tasks mostly cannot reach half of the state-of-the-art results. To explore the reasons behind this failure, we conduct comprehensive experiments on 18 specification-heavy tasks with various LLMs and identify three primary reasons: inability to specifically understand context, misalignment in task schema comprehension with humans, and inadequate long-text understanding ability. Furthermore, we demonstrate that through fine-tuning, LLMs can achieve decent performance on these tasks, indicating that the failure of ICL is not an inherent flaw of LLMs, but rather a drawback of existing alignment methods that renders LLMs incapable of handling complicated specification-heavy tasks via ICL. To substantiate this, we perform dedicated instruction tuning on LLMs for these tasks and observe a notable improvement. We hope the analyses in this paper could facilitate advancements in alignment methods enabling LLMs to meet more sophisticated human demands.

Haotian Xia, Hao Peng, Yunjia Qi et al.

Story generation aims to automatically produce coherent, structured, and engaging narratives. Although large language models (LLMs) have significantly advanced text generation, stories generated by LLMs still diverge from human-authored works regarding complex narrative structure and human-aligned preferences. A key reason is the absence of effective modeling of human story preferences, which are inherently subjective and under-explored. In this work, we systematically evaluate the modeling of human story preferences and introduce StoryRMB, the first benchmark for assessing reward models on story preferences. StoryRMB contains $1,133$ high-quality, human-verified instances, each consisting of a prompt, one chosen story, and three rejected stories. We find existing reward models struggle to select human-preferred stories, with the best model achieving only $66.3\%$ accuracy. To address this limitation, we construct roughly $100,000$ high-quality story preference pairs across diverse domains and develop StoryReward, an advanced reward model for story preference trained on this dataset. StoryReward achieves state-of-the-art (SoTA) performance on StoryRMB, outperforming much larger models. We also adopt StoryReward in downstream test-time scaling applications for best-of-n (BoN) story selection and find that it generally chooses stories better aligned with human preferences. We will release our dataset, model, and code to facilitate future research. Related code and data are available at https://github.com/THU-KEG/StoryReward.

Ming Li, Yusheng Su, Hsiu-Yuan Huang et al. · tsinghua

Humans no doubt use language to communicate about their emotional experiences, but does language in turn help humans understand emotions, or is language just a vehicle of communication? This study used a form of artificial intelligence (AI) known as large language models (LLMs) to assess whether language-based representations of emotion causally contribute to the AI's ability to generate inferences about the emotional meaning of novel situations. Fourteen attributes of human emotion concept representation were found to be represented by the LLM's distinct artificial neuron populations. By manipulating these attribute-related neurons, we in turn demonstrated the role of emotion concept knowledge in generative emotion inference. The attribute-specific performance deterioration was related to the importance of different attributes in human mental space. Our findings provide a proof-in-concept that even a LLM can learn about emotions in the absence of sensory-motor representations and highlight the contribution of language-derived emotion-concept knowledge for emotion inference.

Yong Guan, Hao Peng, Xiaozhi Wang et al. · tsinghua

Future event prediction (FEP) is a long-standing and crucial task in the world, as understanding the evolution of events enables early risk identification, informed decision-making, and strategic planning. Existing work typically treats event prediction as classification tasks and confines the outcomes of future events to a fixed scope, such as yes/no questions, candidate set, and taxonomy, which is difficult to include all possible outcomes of future events. In this paper, we introduce OpenEP (an Open-Ended Future Event Prediction task), which generates flexible and diverse predictions aligned with real-world scenarios. This is mainly reflected in two aspects: firstly, the predictive questions are diverse, covering different stages of event development and perspectives; secondly, the outcomes are flexible, without constraints on scope or format. To facilitate the study of this task, we construct OpenEPBench, an open-ended future event prediction dataset. For question construction, we pose questions from seven perspectives, including location, time, event development, event outcome, event impact, event response, and other, to facilitate an in-depth analysis and understanding of the comprehensive evolution of events. For outcome construction, we collect free-form text containing the outcomes as ground truth to provide semantically complete and detail-enriched outcomes. Furthermore, we propose StkFEP, a stakeholder-enhanced future event prediction framework, that incorporates event characteristics for open-ended settings. Our method extracts stakeholders involved in events to extend questions to gather diverse information. We also collect historically events that are relevant and similar to the question to reveal potential evolutionary patterns. Experiment results indicate that accurately predicting future events in open-ended settings is challenging for existing LLMs.

Ji Qi, Kaixuan Ji, Xiaozhi Wang et al. · tsinghua

Open Information Extraction (OIE) aims to extract objective structured knowledge from natural texts, which has attracted growing attention to build dedicated models with human experience. As the large language models (LLMs) have exhibited remarkable in-context learning capabilities, a question arises as to whether the task of OIE can be effectively tackled with this paradigm? In this paper, we explore solving the OIE problem by constructing an appropriate reasoning environment for LLMs. Specifically, we first propose a method to effectively estimate the discrepancy of syntactic distribution between a LLM and test samples, which can serve as correlation evidence for preparing positive demonstrations. Upon the evidence, we introduce a simple yet effective mechanism to establish the reasoning environment for LLMs on specific tasks. Without bells and whistles, experimental results on the standard CaRB benchmark demonstrate that our $6$-shot approach outperforms state-of-the-art supervised method, achieving an $55.3$ $F_1$ score. Further experiments on TACRED and ACE05 show that our method can naturally generalize to other information extraction tasks, resulting in improvements of $5.7$ and $6.8$ $F_1$ scores, respectively.

Hao Peng, Yunjia Qi, Xiaozhi Wang et al.

Reward models (RMs) are crucial for the training of large language models (LLMs), yet they typically rely on large-scale human-annotated preference pairs. With the widespread deployment of LLMs, in-the-wild interactions have emerged as a rich source of implicit reward signals. This raises the question: Can we develop reward models directly from in-the-wild interactions? In this work, we explore this possibility by adopting WildChat as an interaction source and proposing a pipeline to extract reliable human feedback, yielding 186k high-quality instances for training WildReward via ordinal regression directly on user feedback without preference pairs. Extensive experiments demonstrate that WildReward achieves comparable or even superior performance compared to conventional reward models, with improved calibration and cross-sample consistency. We also observe that WildReward benefits directly from user diversity, where more users yield stronger reward models. Finally, we apply WildReward to online DPO training and observe significant improvements across various tasks. Code and data are released at https://github.com/THU-KEG/WildReward.

Haowen Pan, Yixin Cao, Xiaozhi Wang et al.

Understanding the internal mechanisms by which multi-modal large language models (LLMs) interpret different modalities and integrate cross-modal representations is becoming increasingly critical for continuous improvements in both academia and industry. In this paper, we propose a novel method to identify key neurons for interpretability -- how multi-modal LLMs bridge visual and textual concepts for captioning. Our method improves conventional works upon efficiency and applied range by removing needs of costly gradient computation. Based on those identified neurons, we further design a multi-modal knowledge editing method, beneficial to mitigate sensitive words or hallucination. For rationale of our design, we provide theoretical assumption. For empirical evaluation, we have conducted extensive quantitative and qualitative experiments. The results not only validate the effectiveness of our methods, but also offer insightful findings that highlight three key properties of multi-modal neurons: sensitivity, specificity and causal-effect, to shed light for future research.

Yi Jing, Zao Dai, Jinwu Hu et al.

Model internals encode rich information about how a large language model (LLM) processes its training data; however, post-training data engineering largely relies on external signals and ignores rich intrinsic signals lying in model internals. We propose SAERL, a data engineering framework for LLM reinforcement learning (RL). It models three intrinsic data properties: diversity, difficulty, and quality, using model internals extracted with Sparse Autoencoder (SAE), an advanced mechanistic interpretability tool. Each property grounds a concrete data engineering operation: SAE-space clustering with moderate batch mixing for batch diversity control, a difficulty proxy for easy-to-hard curriculum ordering, and a quality probe for data filtering. SAERL improves average accuracy by 3.00% over vanilla GRPO and reaches target accuracy with 20% fewer training steps on Qwen2.5-Math-1.5B, with consistent gains across model scales and RL algorithms. Experiments show that SAE transfers effectively across model families and scales, serving as a lightweight and reusable data engineering tool. These results demonstrate that model internals are a powerful and practical source of signals for post-training data engineering.

Yunjia Qi, Hao Peng, Xintong Shi et al.

Instruction following aims to align Large Language Models (LLMs) with human intent by specifying explicit constraints on how tasks should be performed. However, we reveal a counterintuitive phenomenon: instruction following can paradoxically interfere with LLMs' task-solving capability. We propose a metric, SUSTAINSCORE, to quantify the interference of instruction following with task solving. It measures task performance drop after inserting into the instruction a self-evident constraint, which is naturally met by the original successful model output and extracted from it. Experiments on current LLMs in mathematics, multi-hop QA, and code generation show that adding the self-evident constraints leads to substantial performance drops, even for advanced models such as Claude-Sonnet-4.5. We validate the generality of the interference across constraint types and scales. Furthermore, we identify common failure patterns, and by investigating the mechanisms of interference, we observe that failed cases allocate significantly more attention to constraints compared to successful ones. Finally, we use SUSTAINSCORE to conduct an initial investigation into how distinct post-training paradigms affect the interference, presenting empirical observations on current alignment strategies. We will release our code and data to facilitate further research

Yixiu Zhao, Xiaozhi Wang, Zijun Yao et al.

Large language models (LLMs) exhibit remarkable capabilities across a wide range of tasks, yet their internal mechanisms remain largely opaque. In this paper, we introduce a simple, lightweight, and broadly applicable method with a focus on isolating neurons that encode specific skills. Building upon prior work that identified "skill neurons" via soft prompt training on classification tasks, our approach extends the analysis to complex scenarios involving multiple skills. We correlate neuron activations with auxiliary metrics -- such as external labels and the model's own confidence score -- thereby uncovering interpretable and task-specific behaviors without the need for manual token aggregation. We empirically validate our method on tasks spanning open-ended text generation and natural language inference, demonstrating its ability to detect neurons that not only drive known skills but also reveal previously unidentified shortcuts in arithmetic reasoning on BigBench.

Hao Peng, Yunjia Qi, Xiaozhi Wang et al. · tsinghua

Reward models (RMs) are crucial for the training and inference-time scaling up of large language models (LLMs). However, existing reward models primarily focus on human preferences, neglecting verifiable correctness signals which have shown strong potential in training LLMs. In this paper, we propose agentic reward modeling, a reward system that combines reward models with verifiable correctness signals from different aspects to provide reliable rewards. We empirically implement a reward agent, named RewardAgent, that combines human preference rewards with two verifiable signals: factuality and instruction following, to provide more reliable rewards. We conduct comprehensive experiments on existing reward model benchmarks and inference time best-of-n searches on real-world downstream tasks. RewardAgent significantly outperforms vanilla reward models, demonstrating its effectiveness. We further construct training preference pairs using RewardAgent and train an LLM with the DPO objective, achieving superior performance on various NLP benchmarks compared to conventional reward models. Our codes are publicly released to facilitate further research (https://github.com/THU-KEG/Agentic-Reward-Modeling).

Yunjia Qi, Hao Peng, Xiaozhi Wang et al. · tsinghua

Large language models (LLMs) usually fall short on information extraction (IE) tasks and struggle to follow the complex instructions of IE tasks. This primarily arises from LLMs not being aligned with humans, as mainstream alignment datasets typically do not include IE data. In this paper, we introduce ADELIE (Aligning large language moDELs on Information Extraction), an aligned LLM that effectively solves various IE tasks, including closed IE, open IE, and on-demand IE. We first collect and construct a high-quality alignment corpus IEInstruct for IE. Then we train ADELIE_SFT using instruction tuning on IEInstruct. We further train ADELIE_SFT with direct preference optimization (DPO) objective, resulting in ADELIE_DPO. Extensive experiments on various held-out IE datasets demonstrate that our models (ADELIE_SFT and ADELIE_DPO) achieve state-of-the-art (SoTA) performance among open-source models. We further explore the general capabilities of ADELIE, and experimental results reveal that their general capabilities do not exhibit a noticeable decline. We will release the code, data, and models to facilitate further research.

Hao Peng, Xiaozhi Wang, Chunyang Li et al.

Knowledge editing aims at updating knowledge of large language models (LLMs) to prevent them from becoming outdated. Existing work edits LLMs at the level of factual knowledge triplets. However, natural knowledge updates in the real world come from the occurrences of new events rather than direct changes in factual triplets. In this paper, we propose a new task setting: event-level knowledge editing, which directly edits new events into LLMs and improves over conventional triplet-level editing on (1) Efficiency. A single event edit leads to updates in multiple entailed knowledge triplets. (2) Completeness. Beyond updating factual knowledge, event-level editing also requires considering the event influences and updating LLMs' knowledge about future trends. We construct a high-quality event-level editing benchmark ELKEN, consisting of 1,515 event edits, 6,449 questions about factual knowledge, and 10,150 questions about future tendencies. We systematically evaluate the performance of various knowledge editing methods and LLMs on this benchmark. We find that ELKEN poses significant challenges to existing knowledge editing approaches. Our codes and dataset are publicly released to facilitate further research.

Hao Peng, Yunjia Qi, Xiaozhi Wang et al. · tsinghua

Reinforcement learning with verifiable rewards (RLVR) has become a key technique for enhancing large language models (LLMs), with verification engineering playing a central role. However, best practices for RL in instruction following remain underexplored. In this work, we explore the verification challenge in RL for instruction following and propose VerIF, a verification method that combines rule-based code verification with LLM-based verification from a large reasoning model (e.g., QwQ-32B). To support this approach, we construct a high-quality instruction-following dataset, VerInstruct, containing approximately 22,000 instances with associated verification signals. We apply RL training with VerIF to two models, achieving significant improvements across several representative instruction-following benchmarks. The trained models reach state-of-the-art performance among models of comparable size and generalize well to unseen constraints. We further observe that their general capabilities remain unaffected, suggesting that RL with VerIF can be integrated into existing RL recipes to enhance overall model performance. We have released our datasets, codes, and models to facilitate future research at https://github.com/THU-KEG/VerIF.

Yunjia Qi, Hao Peng, Xiaozhi Wang et al. · tsinghua

Large Language Models (LLMs) have demonstrated advanced capabilities in real-world agentic applications. Growing research efforts aim to develop LLM-based agents to address practical demands, introducing a new challenge: agentic scenarios often involve lengthy instructions with complex constraints, such as extended system prompts and detailed tool specifications. While adherence to such instructions is crucial for agentic applications, whether LLMs can reliably follow them remains underexplored. In this paper, we introduce AgentIF, the first benchmark for systematically evaluating LLM instruction following ability in agentic scenarios. AgentIF features three key characteristics: (1) Realistic, constructed from 50 real-world agentic applications. (2) Long, averaging 1,723 words with a maximum of 15,630 words. (3) Complex, averaging 11.9 constraints per instruction, covering diverse constraint types, such as tool specifications and condition constraints. To construct AgentIF, we collect 707 human-annotated instructions across 50 agentic tasks from industrial application agents and open-source agentic systems. For each instruction, we annotate the associated constraints and corresponding evaluation metrics, including code-based evaluation, LLM-based evaluation, and hybrid code-LLM evaluation. We use AgentIF to systematically evaluate existing advanced LLMs. We observe that current models generally perform poorly, especially in handling complex constraint structures and tool specifications. We further conduct error analysis and analytical experiments on instruction length and meta constraints, providing some findings about the failure modes of existing LLMs. We have released the code and data to facilitate future research.

Yunjia Qi, Hao Peng, Xiaozhi Wang et al. · tsinghua

Large language models (LLMs) struggle to follow instructions with complex constraints in format, length, etc. Following the conventional instruction-tuning practice, previous works conduct post-training on complex instruction-response pairs generated by feeding complex instructions to advanced LLMs. However, even advanced LLMs cannot follow complex instructions well, thus limiting the quality of generated data. In this work, we find that existing datasets inherently contain implicit complex constraints and propose a novel data generation technique, constraint back-translation. Specifically, we take the high-quality instruction-response pairs in existing datasets and only adopt advanced LLMs to add complex constraints already met by the responses to the instructions, which naturally reduces costs and data noise. In the experiments, we adopt Llama3-70B-Instruct to back-translate constraints and create a high-quality complex instruction-response dataset, named CRAB. We present that post-training on CRAB improves multiple backbone LLMs' complex instruction-following ability, evaluated on extensive instruction-following benchmarks. We further find that constraint back-translation also serves as a useful auxiliary training objective in post-training. Our code, data, and models will be released to facilitate future research.

Jianhui Chen, Xiaozhi Wang, Zijun Yao et al.

Large language models (LLMs) excel in various capabilities but pose safety risks such as generating harmful content and misinformation, even after safety alignment. In this paper, we explore the inner mechanisms of safety alignment through the lens of mechanistic interpretability, focusing on identifying and analyzing safety neurons within LLMs that are responsible for safety behaviors. We propose inference-time activation contrasting to locate these neurons and dynamic activation patching to evaluate their causal effects on model safety. Experiments on multiple prevalent LLMs demonstrate that we can consistently identify about $5\%$ safety neurons, and by only patching their activations we can restore over $90\%$ of the safety performance across various red-teaming benchmarks without influencing general ability. The finding of safety neurons also helps explain the ''alignment tax'' phenomenon by revealing that the key neurons for model safety and helpfulness significantly overlap, yet they require different activation patterns for the same neurons. Furthermore, we demonstrate an application of our findings in safeguarding LLMs by detecting unsafe outputs before generation. The source code is available at https://github.com/THU-KEG/SafetyNeuron.

Shangqing Tu, Yuanchun Wang, Jifan Yu et al.

Large language models have achieved remarkable success on general NLP tasks, but they may fall short for domain-specific problems. Recently, various Retrieval-Augmented Large Language Models (RALLMs) are proposed to address this shortcoming. However, existing evaluation tools only provide a few baselines and evaluate them on various domains without mining the depth of domain knowledge. In this paper, we address the challenges of evaluating RALLMs by introducing the R-Eval toolkit, a Python toolkit designed to streamline the evaluation of different RAG workflows in conjunction with LLMs. Our toolkit, which supports popular built-in RAG workflows and allows for the incorporation of customized testing data on the specific domain, is designed to be user-friendly, modular, and extensible. We conduct an evaluation of 21 RALLMs across three task levels and two representative domains, revealing significant variations in the effectiveness of RALLMs across different tasks and domains. Our analysis emphasizes the importance of considering both task and domain requirements when choosing a RAG workflow and LLM combination. We are committed to continuously maintaining our platform at https://github.com/THU-KEG/R-Eval to facilitate both the industry and the researchers.

Zhengyan Zhang, Zhiyuan Zeng, Yankai Lin et al.

This work examines the presence of modularity in pre-trained Transformers, a feature commonly found in human brains and thought to be vital for general intelligence. In analogy to human brains, we consider two main characteristics of modularity: (1) functional specialization of neurons: we evaluate whether each neuron is mainly specialized in a certain function, and find that the answer is yes. (2) function-based neuron grouping: we explore finding a structure that groups neurons into modules by function, and each module works for its corresponding function. Given the enormous amount of possible structures, we focus on Mixture-of-Experts as a promising candidate, which partitions neurons into experts and usually activates different experts for different inputs. Experimental results show that there are functional experts, where clustered are the neurons specialized in a certain function. Moreover, perturbing the activations of functional experts significantly affects the corresponding function. Finally, we study how modularity emerges during pre-training, and find that the modular structure is stabilized at the early stage, which is faster than neuron stabilization. It suggests that Transformers first construct the modular structure and then learn fine-grained neuron functions. Our code and data are available at https://github.com/THUNLP/modularity-analysis.

Ji Qi, Chuchun Zhang, Xiaozhi Wang et al.

The robustness to distribution changes ensures that NLP models can be successfully applied in the realistic world, especially for information extraction tasks. However, most prior evaluation benchmarks have been devoted to validating pairwise matching correctness, ignoring the crucial measurement of robustness. In this paper, we present the first benchmark that simulates the evaluation of open information extraction models in the real world, where the syntactic and expressive distributions under the same knowledge meaning may drift variously. We design and annotate a large-scale testbed in which each example is a knowledge-invariant clique that consists of sentences with structured knowledge of the same meaning but with different syntactic and expressive forms. By further elaborating the robustness metric, a model is judged to be robust if its performance is consistently accurate on the overall cliques. We perform experiments on typical models published in the last decade as well as a popular large language model, the results show that the existing successful models exhibit a frustrating degradation, with a maximum drop of 23.43 F1 score. Our resources and code are available at https://github.com/qijimrc/ROBUST.

Yusheng Su, Xiaozhi Wang, Yujia Qin et al.

Prompt tuning (PT) is a promising parameter-efficient method to utilize extremely large pre-trained language models (PLMs), which can achieve comparable performance to full-parameter fine-tuning by only tuning a few soft prompts. However, PT requires much more training time than fine-tuning. Intuitively, knowledge transfer can help to improve the efficiency. To explore whether we can improve PT via prompt transfer, we empirically investigate the transferability of soft prompts across different downstream tasks and PLMs in this work. We find that (1) in zero-shot setting, trained soft prompts can effectively transfer to similar tasks on the same PLM and also to other PLMs with a cross-model projector trained on similar tasks; (2) when used as initialization, trained soft prompts of similar tasks and projected prompts of other PLMs can significantly accelerate training and also improve the performance of PT. Moreover, to explore what decides prompt transferability, we investigate various transferability indicators and find that the overlapping rate of activated neurons strongly reflects the transferability, which suggests how the prompts stimulate PLMs is essential. Our findings show that prompt transfer is promising for improving PT, and further research shall focus more on prompts' stimulation to PLMs. The source code can be obtained from https://github.com/thunlp/Prompt-Transferability.

Chenglei Si, Zhengyan Zhang, Yingfa Chen et al.

Tokenization is fundamental to pretrained language models (PLMs). Existing tokenization methods for Chinese PLMs typically treat each character as an indivisible token. However, they ignore the unique feature of the Chinese writing system where additional linguistic information exists below the character level, i.e., at the sub-character level. To utilize such information, we propose sub-character (SubChar for short) tokenization. Specifically, we first encode the input text by converting each Chinese character into a short sequence based on its glyph or pronunciation, and then construct the vocabulary based on the encoded text with sub-word segmentation. Experimental results show that SubChar tokenizers have two main advantages over existing tokenizers: 1) They can tokenize inputs into much shorter sequences, thus improving the computational efficiency. 2) Pronunciation-based SubChar tokenizers can encode Chinese homophones into the same transliteration sequences and produce the same tokenization output, hence being robust to homophone typos. At the same time, models trained with SubChar tokenizers perform competitively on downstream tasks. We release our code and models at https://github.com/thunlp/SubCharTokenization to facilitate future work.

Ziqi Wang, Xiaozhi Wang, Xu Han et al.

Event extraction (EE) has considerably benefited from pre-trained language models (PLMs) by fine-tuning. However, existing pre-training methods have not involved modeling event characteristics, resulting in the developed EE models cannot take full advantage of large-scale unsupervised data. To this end, we propose CLEVE, a contrastive pre-training framework for EE to better learn event knowledge from large unsupervised data and their semantic structures (e.g. AMR) obtained with automatic parsers. CLEVE contains a text encoder to learn event semantics and a graph encoder to learn event structures respectively. Specifically, the text encoder learns event semantic representations by self-supervised contrastive learning to represent the words of the same events closer than those unrelated words; the graph encoder learns event structure representations by graph contrastive pre-training on parsed event-related semantic structures. The two complementary representations then work together to improve both the conventional supervised EE and the unsupervised "liberal" EE, which requires jointly extracting events and discovering event schemata without any annotated data. Experiments on ACE 2005 and MAVEN datasets show that CLEVE achieves significant improvements, especially in the challenging unsupervised setting. The source code and pre-trained checkpoints can be obtained from https://github.com/THU-KEG/CLEVE.

Zhengyan Zhang, Xu Han, Hao Zhou et al.

Pre-trained Language Models (PLMs) have proven to be beneficial for various downstream NLP tasks. Recently, GPT-3, with 175 billion parameters and 570GB training data, drew a lot of attention due to the capacity of few-shot (even zero-shot) learning. However, applying GPT-3 to address Chinese NLP tasks is still challenging, as the training corpus of GPT-3 is primarily English, and the parameters are not publicly available. In this technical report, we release the Chinese Pre-trained Language Model (CPM) with generative pre-training on large-scale Chinese training data. To the best of our knowledge, CPM, with 2.6 billion parameters and 100GB Chinese training data, is the largest Chinese pre-trained language model, which could facilitate several downstream Chinese NLP tasks, such as conversation, essay generation, cloze test, and language understanding. Extensive experiments demonstrate that CPM achieves strong performance on many NLP tasks in the settings of few-shot (even zero-shot) learning. The code and parameters are available at https://github.com/TsinghuaAI/CPM-Generate.

Xiaozhi Wang, Ziqi Wang, Xu Han et al.

Event detection (ED), which means identifying event trigger words and classifying event types, is the first and most fundamental step for extracting event knowledge from plain text. Most existing datasets exhibit the following issues that limit further development of ED: (1) Data scarcity. Existing small-scale datasets are not sufficient for training and stably benchmarking increasingly sophisticated modern neural methods. (2) Low coverage. Limited event types of existing datasets cannot well cover general-domain events, which restricts the applications of ED models. To alleviate these problems, we present a MAssive eVENt detection dataset (MAVEN), which contains 4,480 Wikipedia documents, 118,732 event mention instances, and 168 event types. MAVEN alleviates the data scarcity problem and covers much more general event types. We reproduce the recent state-of-the-art ED models and conduct a thorough evaluation on MAVEN. The experimental results show that existing ED methods cannot achieve promising results on MAVEN as on the small datasets, which suggests that ED in the real world remains a challenging task and requires further research efforts. We also discuss further directions for general domain ED with empirical analyses. The source code and dataset can be obtained from https://github.com/THU-KEG/MAVEN-dataset.

Yuxian Gu, Zhengyan Zhang, Xiaozhi Wang et al.

Recently, pre-trained language models mostly follow the pre-train-then-fine-tuning paradigm and have achieved great performance on various downstream tasks. However, since the pre-training stage is typically task-agnostic and the fine-tuning stage usually suffers from insufficient supervised data, the models cannot always well capture the domain-specific and task-specific patterns. In this paper, we propose a three-stage framework by adding a task-guided pre-training stage with selective masking between general pre-training and fine-tuning. In this stage, the model is trained by masked language modeling on in-domain unsupervised data to learn domain-specific patterns and we propose a novel selective masking strategy to learn task-specific patterns. Specifically, we design a method to measure the importance of each token in sequences and selectively mask the important tokens. Experimental results on two sentiment analysis tasks show that our method can achieve comparable or even better performance with less than 50% of computation cost, which indicates our method is both effective and efficient. The source code of this paper can be obtained from https://github.com/thunlp/SelectiveMasking.

Xiaozhi Wang, Tianyu Gao, Zhaocheng Zhu et al.

Pre-trained language representation models (PLMs) cannot well capture factual knowledge from text. In contrast, knowledge embedding (KE) methods can effectively represent the relational facts in knowledge graphs (KGs) with informative entity embeddings, but conventional KE models cannot take full advantage of the abundant textual information. In this paper, we propose a unified model for Knowledge Embedding and Pre-trained LanguagE Representation (KEPLER), which can not only better integrate factual knowledge into PLMs but also produce effective text-enhanced KE with the strong PLMs. In KEPLER, we encode textual entity descriptions with a PLM as their embeddings, and then jointly optimize the KE and language modeling objectives. Experimental results show that KEPLER achieves state-of-the-art performances on various NLP tasks, and also works remarkably well as an inductive KE model on KG link prediction. Furthermore, for pre-training and evaluating KEPLER, we construct Wikidata5M, a large-scale KG dataset with aligned entity descriptions, and benchmark state-of-the-art KE methods on it. It shall serve as a new KE benchmark and facilitate the research on large KG, inductive KE, and KG with text. The source code can be obtained from https://github.com/THU-KEG/KEPLER.

Yushi Bai, Shangqing Tu, Jiajie Zhang et al. · tsinghua

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. The project is available at https://longbench2.github.io.

Yixin Cao, Shibo Hong, Xinze Li et al.

Large Language Models (LLMs) are advancing at an amazing speed and have become indispensable across academia, industry, and daily applications. To keep pace with the status quo, this survey probes the core challenges that the rise of LLMs poses for evaluation. We identify and analyze two pivotal transitions: (i) from task-specific to capability-based evaluation, which reorganizes benchmarks around core competencies such as knowledge, reasoning, instruction following, multi-modal understanding, and safety; and (ii) from manual to automated evaluation, encompassing dynamic dataset curation and "LLM-as-a-judge" scoring. Yet, even with these transitions, a crucial obstacle persists: the evaluation generalization issue. Bounded test sets cannot scale alongside models whose abilities grow seemingly without limit. We will dissect this issue, along with the core challenges of the above two transitions, from the perspectives of methods, datasets, evaluators, and metrics. Due to the fast evolving of this field, we will maintain a living GitHub repository (links are in each section) to crowd-source updates and corrections, and warmly invite contributors and collaborators.

Haowen Pan, Xiaozhi Wang, Yixin Cao et al.

Knowledge editing aims to update outdated information in Large Language Models (LLMs). A representative line of study is locate-then-edit methods, which typically employ causal tracing to identify the modules responsible for recalling factual knowledge about entities. However, we find these methods are often sensitive only to changes in the subject entity, leaving them less effective at adapting to changes in relations. This limitation results in poor editing locality, which can lead to the persistence of irrelevant or inaccurate facts, ultimately compromising the reliability of LLMs. We believe this issue arises from the insufficient precision of knowledge localization. To address this, we propose a Fine-grained Neuron-level Knowledge Editing (FiNE) method that enhances editing locality without affecting overall success rates. By precisely identifying and modifying specific neurons within feed-forward networks, FiNE significantly improves knowledge localization and editing. Quantitative experiments demonstrate that FiNE efficiently achieves better overall performance compared to existing techniques, providing new insights into the localization and modification of knowledge within LLMs.

Haotian Xia, Hao Peng, Yunjia Qi et al. · tsinghua

Long story generation remains a challenge for existing large language models (LLMs), primarily due to two main factors: (1) discourse coherence, which requires plot consistency, logical coherence, and completeness in the long-form generation, and (2) narrative complexity, which requires an interwoven and engaging narrative. To address these challenges, we propose StoryWriter, a multi-agent story generation framework, which consists of three main modules: (1) outline agent, which generates event-based outlines containing rich event plots, character, and event-event relationships. (2) planning agent, which further details events and plans which events should be written in each chapter to maintain an interwoven and engaging story. (3) writing agent, which dynamically compresses the story history based on the current event to generate and reflect new plots, ensuring the coherence of the generated story. We conduct both human and automated evaluation, and StoryWriter significantly outperforms existing story generation baselines in both story quality and length. Furthermore, we use StoryWriter to generate a dataset, which contains about $6,000$ high-quality long stories, with an average length of $8,000$ words. We train the model Llama3.1-8B and GLM4-9B using supervised fine-tuning on LongStory and develop StoryWriter_GLM and StoryWriter_GLM, which demonstrates advanced performance in long story generation.

Yong Guan, Xiaozhi Wang, Lei Hou et al. · tsinghua

Event relation extraction (ERE) is a critical and fundamental challenge for natural language processing. Existing work mainly focuses on directly modeling the entire document, which cannot effectively handle long-range dependencies and information redundancy. To address these issues, we propose a cluster-aware compression method for improving event relation extraction (TacoERE), which explores a compression-then-extraction paradigm. Specifically, we first introduce document clustering for modeling event dependencies. It splits the document into intra- and inter-clusters, where intra-clusters aim to enhance the relations within the same cluster, while inter-clusters attempt to model the related events at arbitrary distances. Secondly, we utilize cluster summarization to simplify and highlight important text content of clusters for mitigating information redundancy and event distance. We have conducted extensive experiments on both pre-trained language models, such as RoBERTa, and large language models, such as ChatGPT and GPT-4, on three ERE datasets, i.e., MAVEN-ERE, EventStoryLine and HiEve. Experimental results demonstrate that TacoERE is an effective method for ERE.

Yong Guan, Dingxiao Liu, Jinchen Ma et al. · tsinghua

Generative document retrieval, an emerging paradigm in information retrieval, learns to build connections between documents and identifiers within a single model, garnering significant attention. However, there are still two challenges: (1) neglecting inner-content correlation during document representation; (2) lacking explicit semantic structure during identifier construction. Nonetheless, events have enriched relations and well-defined taxonomy, which could facilitate addressing the above two challenges. Inspired by this, we propose Event GDR, an event-centric generative document retrieval model, integrating event knowledge into this task. Specifically, we utilize an exchange-then-reflection method based on multi-agents for event knowledge extraction. For document representation, we employ events and relations to model the document to guarantee the comprehensiveness and inner-content correlation. For identifier construction, we map the events to well-defined event taxonomy to construct the identifiers with explicit semantic structure. Our method achieves significant improvement over the baselines on two datasets, and also hopes to provide insights for future research.