Yi Hu, Jiaqi Gu, Ruxin Wang, Zijun Yao, Hao Peng, Xiaobao Wu, Jianhui Chen, Muhan Zhang, Liangming Pan. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

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.

Daniel Zhang-Li, Joy Jia Yin Lim, Binglin Liu, Shangqing Tu, Zijun Yao, Hao Peng, Jifan Yu, Haoxuan Li, Zhanxin Hao, Ye He, Zekun Li, Jiangyi Wang, Lei Hou, Bin Xu, Xin Cong, Zhiyuan Liu, Huiqin Liu, Yu Zhang, Juanzi Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zijun Yao, Weijian Qi, Liangming Pan, Shulin Cao, Linmei Hu, Liu Weichuan, Lei Hou, Juanzi Li. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Knowledge distillation (KD) aims to transfer knowledge from a large teacher model to a smaller student model.Previous work applying KD in the field of large language models (LLMs) typically focused on the post-training phase, where the student LLM learns directly from instructions and corresponding responses generated by the teacher model.In this paper, we extend KD to the pre-training phase of LLMs, named pre-training distillation (PD).We first conduct a preliminary experiment using GLM-4-9B as the teacher LLM to distill a 1.9B parameter student LLM, validating the effectiveness of PD.Considering the key impact factors of distillation, we systematically explore the design space of pre-training distillation across four aspects: logits processing, loss selection, scaling law, and offline or online logits.We conduct extensive experiments to explore the design space of pre-training distillation and find better configurations and interesting conclusions, such as larger student LLMs generally benefiting more from pre-training distillation, while a larger teacher LLM does not necessarily guarantee better results.We hope our exploration of the design space will inform future practices in pre-training distillation.

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.

Specialized entity linking (EL) models are well-trained at mapping mentions to unique knowledge base (KB) entities according to a given context. However, specialized EL models struggle to disambiguate long-tail entities due to their limited training data. Meanwhile, extensively pre-trained large language models (LLMs) possess broader knowledge of uncommon entities. Yet, with a lack of specialized EL training, LLMs frequently fail to generate accurate KB entity names, limiting their standalone effectiveness in EL. With the observation that LLMs are more adept at context generation instead of EL execution, we introduce LLM-Augmented Entity Linking (LLMAEL), the first framework to enhance specialized EL models with LLM data augmentation. LLMAEL leverages off-the-shelf, tuning-free LLMs as context augmenters, generating entity descriptions to serve as additional input for specialized EL models. Experiments show that LLMAEL sets new state-of-the-art results across 6 widely adopted EL benchmarks: compared to prior methods that integrate tuning-free LLMs into EL, LLMAEL achieves an absolute 8.9% gain in EL accuracy. We release our code and datasets.

Large language models (LLMs) demonstrate exceptional performance on tasks requiring complex linguistic abilities, such as reference disambiguation and metaphor recognition/generation.Although LLMs possess impressive capabilities, their internal mechanisms for processing and representing linguistic knowledge remain largely opaque.Prior research on linguistic mechanisms is limited by coarse granularity, limited analysis scale, and narrow focus.In this study, we propose LINGUALENS, a systematic and comprehensive framework for analyzing the linguistic mechanisms of large language models, based on Sparse Auto-Encoders (SAEs).We extract a broad set of Chinese and English linguistic features across four dimensions-morphology, syntax, semantics, and pragmatics.By employing counterfactual methods, we construct a large-scale counterfactual dataset of linguistic features for mechanism analysis.Our findings reveal intrinsic representations of linguistic knowledge in LLMs, uncover patterns of cross-layer and cross-lingual distribution, and demonstrate the potential to control model outputs.This work provides a systematic suite of resources and methods for studying linguistic mechanisms, offers strong evidence that LLMs possess genuine linguistic knowledge, and lays the foundation for more interpretable and controllable language modeling in future research.

Reward models are critical in techniques like Reinforcement Learning from Human Feedback (RLHF) and Inference Scaling Laws, where they guide language model alignment and select optimal responses. Despite their importance, existing reward model benchmarks often evaluate models by asking them to distinguish between responses generated by models of varying power. However, this approach fails to assess reward models on subtle but critical content changes and variations in style, resulting in a low correlation with policy model performance. To this end, we introduce RM-Bench, a novel benchmark designed to evaluate reward models based on their sensitivity to subtle content differences and resistance to style biases. Extensive experiments demonstrate that RM-Bench strongly correlates with policy model performance, making it a reliable reference for selecting reward models to align language models effectively. We evaluate nearly 40 reward models on RM-Bench. Our results reveal that even state-of-the-art models achieve an average performance of only 46.6%, which falls short of random-level accuracy (50%) when faced with style bias interference. These findings highlight the significant room for improvement in current reward models.

Large language models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks. However, existing approaches mainly rely on imitation learning and struggle to achieve effective test-time scaling. While reinforcement learning (RL) holds promise for enabling self-exploration, recent attempts yield modest improvements in complex reasoning. In this paper, we present T1 to scale RL by encouraging exploration and understand inference scaling. We first initialize the LLM using synthesized chain-of-thought data that integrates trial-and-error and self-verification. To scale RL training, we promote increased sampling diversity through over-sampling. We demonstrate that T1 with open LLMs as its base exhibits inference scaling behavior and achieves superior performance on challenging math reasoning benchmarks. More importantly, we present a simple strategy to examine inference scaling, where increased inference budgets directly lead to T1’s better performance without any additional verification. The model weights and training data are publicly available at https://github.com/THUDM/T1.

Despite the outstanding capabilities of large language models (LLMs), knowledge-intensive reasoning still remains a challenging task due to LLMs' limitations in compositional reasoning and the hallucination problem. A prevalent solution is to employ chain-of-thought (CoT) with retrieval-augmented generation (RAG), which first formulates a reasoning plan by decomposing complex questions into simpler sub-questions, and then applies iterative RAG at each sub-question. However, prior works exhibit two crucial problems: inadequate reasoning planning and poor incorporation of heterogeneous knowledge. In this paper, we introduce AtomR, a framework for LLMs to conduct accurate heterogeneous knowledge reasoning at the atomic level. Inspired by how knowledge graph query languages model compositional reasoning through combining predefined operations, we propose three atomic knowledge operators, a unified set of operators for LLMs to retrieve and manipulate knowledge from heterogeneous sources. First, in the reasoning planning stage, AtomR decomposes a complex question into a reasoning tree where each leaf node corresponds to an atomic knowledge operator, achieving question decomposition that is highly fine-grained and orthogonal. Subsequently, in the reasoning execution stage, AtomR executes each atomic knowledge operator, which flexibly selects, retrieves, and operates atomic level knowledge from heterogeneous sources. We also introduce BlendQA, a challenging benchmark specially tailored for heterogeneous knowledge reasoning. Experiments on three single-source and two multi-source datasets show that AtomR outperforms state-of-the-art baselines by a large margin, with absolute F1 score improvements of 9.4% on 2WikiMultihop and 9.5% on BlendQA. We release our code and data https://github.com/THU-KEG/AtomR.git.

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.

Recent work suggests that large language models (LLMs) can perform multi-hop reasoning implicitly---producing correct answers without explicitly verbalizing intermediate steps---but the underlying mechanisms remain poorly understood. In this paper, we study how such implicit reasoning emerges by training transformers from scratch in a controlled symbolic environment. Our analysis reveals a three-stage developmental trajectory: early memorization, followed by in-distribution generalization, and eventually cross-distribution generalization. We find that training with atomic triples is not necessary but accelerates learning, and that second-hop generalization relies on query-level exposure to specific compositional structures. To interpret these behaviors, we introduce two diagnostic tools: cross-query semantic patching, which identifies semantically reusable intermediate representations, and a cosine-based representational lens, which reveals that successful reasoning correlates with the cosine-base clustering in hidden space. This clustering phenomenon in turn provides a coherent explanation for the behavioral dynamics observed across training, linking representational structure to reasoning capability. These findings provide new insights into the interpretability of implicit multi-hop reasoning in LLMs, helping to clarify how complex reasoning processes unfold internally and offering pathways to enhance the transparency of such models.

Xiaokang Zhang, Zijun Yao, Jing Zhang, Kaifeng Yun, Jifan Yu, Juanzi Li, Jie Tang. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Modern Large Language Models (LLMs) have showcased remarkable prowess in various tasks necessitating sophisticated cognitive behaviors. Nevertheless, a paradoxical performance discrepancy is observed, where these models underperform in seemingly elementary tasks like relation extraction and event extraction due to two issues in conventional evaluation. (1) The imprecision of existing evaluation metrics that struggle to effectively gauge semantic consistency between model outputs and ground truth, and (2) The inherent incompleteness of evaluation benchmarks, primarily due to restrictive human annotation schemas, resulting in underestimated LLM performances. Inspired by the principles in subjective question correction, we propose a new evaluation method, SQC-Score. This method innovatively utilizes LLMs, fine-tuned through subjective question correction data, to refine matching between model outputs and golden labels. Additionally, by incorporating a Natural Language Inference (NLI) model, SQC-Score enriches golden labels, addressing benchmark incompleteness by acknowledging correct yet previously omitted answers. Results on three information extraction tasks show that SQC-Score is more preferred by human annotators than the baseline metrics. Utilizing SQC-Score, we conduct a comprehensive evaluation of the state-of-the-art LLMs and provide insights for future research for information extraction. Dataset and associated codes can be accessed at our GitHub repository.

Providing knowledge documents for large language models (LLMs) has emerged as a promising solution to update the static knowledge inherent in their parameters. However, knowledge in the document may conflict with the memory of LLMs due to outdated or incorrect knowledge in the LLMs’ parameters. This leads to the necessity of examining the capability of LLMs to assimilate supplemental external knowledge that conflicts with their memory. While previous studies have explained to what extent LLMs extract conflicting knowledge from the provided text, they neglect the necessity to <b>reason</b> with conflicting knowledge. Furthermore, there lack a detailed analysis on strategies to enable LLMs to resolve conflicting knowledge via prompting, decoding strategy, and supervised fine-tuning. To address these limitations, we construct a new dataset, dubbed KNOT, for knowledge conflict resolution examination in the form of question answering. KNOT facilitates in-depth analysis by dividing reasoning with conflicting knowledge into three levels: (1) Direct Extraction, which directly extracts conflicting knowledge to answer questions. (2) Explicit Reasoning, which reasons with conflicting knowledge when the reasoning path is explicitly provided in the question. (3) Implicit Reasoning, where reasoning with conflicting knowledge requires LLMs to infer the reasoning path independently to answer questions. We also conduct extensive experiments on KNOT to establish empirical guidelines for LLMs to utilize conflicting knowledge in complex circumstances. Dataset and associated codes can be accessed at our <a href=https://github.com/THU-KEG/KNOT>GitHub repository</a> .

Empowered by the large-scale pretrained language models, existing dialogue systems have demonstrated impressive performance conducting fluent and natural-sounding conversations. However, they are still plagued by the <b>hallucination</b> problem, causing unpredictable factual errors in the generated responses. Recently, knowledge-grounded dialogue generation models, that intentionally invoke external knowledge resources to more informative responses, are also proven to be effective in reducing hallucination. Following the idea of getting high-quality knowledge, a few efforts have achieved pretty good performance on this issue. As some inevitable knowledge noises may also lead to hallucinations, it is emergent to investigate the reason and future directions for building noise-tolerant methods in KGD tasks. In this paper, we analyze the causal story behind this problem with counterfactual reasoning methods. Based on the causal effect analysis, we propose a possible solution for alleviating the hallucination in KGD by exploiting the dialogue-knowledge interaction. Experimental results of our example implementation show that this method can reduce hallucination without disrupting other dialogue performance, while keeping adaptive to different generation models. We hope our efforts can support and call for more attention to developing lightweight techniques towards robust and trusty dialogue systems.

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 ability modeling, we mimic human cognition to form a four-level taxonomy of knowledge-related abilities, covering 19 tasks. (2) For 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 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 21 open-source and commercial LLMs and obtain some intriguing findings. The KoLA dataset will be updated every three months to provide timely references for developing LLMs and knowledge-related systems.

Graph Neural Networks (GNNs) have already been widely used in various graph mining tasks. However, recent works reveal that the learned weights (channels) in well-trained GNNs are highly redundant, which inevitably limits the performance of GNNs. Instead of removing these redundant channels for efficiency consideration, we aim to reactivate them to enlarge the representation capacity of GNNs for effective graph learning. In this paper, we propose to substitute these redundant channels with other informative channels to achieve this goal. We introduce a novel GNN learning framework named AKE-GNN, which performs the Adaptive Knowledge Exchange strategy among multiple graph views generated by graph augmentations. AKE-GNN first trains multiple GNNs each corresponding to one graph view to obtain informative channels. Then, AKE-GNN iteratively exchanges redundant channels in the weight parameter matrix of one GNN with informative channels of another GNN in a layer-wise manner. Additionally, existing GNNs can be seamlessly incorporated into our framework. AKE-GNN achieves superior performance compared with various baselines across a suite of experiments on node classification, link prediction, and graph classification. In particular, we conduct a series of experiments on 15 public benchmark datasets, 8 popular GNN models, and 3 graph tasks and show that AKE-GNN consistently outperforms existing popular GNN models and even their ensembles. Extensive ablation studies and analyses on knowledge exchange methods validate the effectiveness of AKE-GNN.

Teaching assistants have played essential roles in the long history of education. However, few MOOC platforms are providing human or virtual teaching assistants to support learning for massive online students due to the complexity of real-world online education scenarios and the lack of training data. In this paper, we present a virtual MOOC teaching assistant, LittleMu with minimum labeled training data, to provide question answering and chit-chat services. Consisting of two interactive modules of heterogeneous retrieval and language model prompting, LittleMu first integrates structural, semi- and unstructured knowledge sources to support accurate answers for a wide range of questions. Then, we design delicate demonstrations named "Chain of Teach" prompts to exploit the large-scale pre-trained model to handle complex uncollected questions. Except for question answering, we develop other educational services such as knowledge-grounded chit-chat. We test the system's performance via both offline evaluation and online deployment. Since May 2020, our LittleMu system has served over 80,000 users with over 300,000 queries from over 500 courses on XuetangX MOOC platform, which continuously contributes to a more convenient and fair education. Our code, services, and dataset will be available at https://github.com/THU-KEG/VTA.

We present GLM-Dialog, a large-scale language model (LLM) with 10B parameters capable of knowledge-grounded conversation in Chinese using a search engine to access the Internet knowledge. GLM-Dialog offers a series of applicable techniques for exploiting various external knowledge including both helpful and noisy knowledge, enabling the creation of robust knowledge-grounded dialogue LLMs with limited proper datasets. To evaluate the GLM-Dialog more fairly, we also propose a novel evaluation method to allow humans to converse with multiple deployed bots simultaneously and compare their performance implicitly instead of explicitly rating using multidimensional metrics. Comprehensive evaluations from automatic to human perspective demonstrate the advantages of GLM-Dialog comparing with existing open source Chinese dialogue models. We release both the model checkpoint and source code, and also deploy it as a WeChat application to interact with users. We offer our evaluation platform online in an effort to prompt the development of open source models and reliable dialogue evaluation systems. All the source code is available on Github.

Student modeling, the task of inferring a student's learning characteristics through their interactions with coursework, is a fundamental issue in intelligent education. Although the recent attempts from knowledge tracing and cognitive diagnosis propose several promising directions for improving the usability and effectiveness of current models, the existing public datasets are still insufficient to meet the need for these potential solutions due to their ignorance of complete exercising contexts, fine-grained concepts, and cognitive labels. In this paper, we present MoocRadar, a fine-grained, multi-aspect knowledge repository consisting of 2,513 exercise questions, 5,600 knowledge concepts, and over 12 million behavioral records. Specifically, we propose a framework to guarantee a high-quality and comprehensive annotation of fine-grained concepts and cognitive labels. The statistical and experimental results indicate that our dataset provides the basis for the future improvements of existing methods. Moreover, to support the convenient usage for researchers, we release a set of tools for data querying, model adaption, and even the extension of our repository, which are now available at https://github.com/THU-KEG/MOOC-Radar.

Shulin Cao, Jiaxin Shi, Zijun Yao, Xin Lv, Jifan Yu, Lei Hou, Juanzi Li, Zhiyuan Liu, Jinghui Xiao. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Zijun Yao, Chengjiang Li, Tiansi Dong, Xin Lv, Jifan Yu, Lei Hou, Juanzi Li, Yichi Zhang, Zelin Dai. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

The prosperity of massive open online courses provides fodder for plentiful research efforts on adaptive learning. However, current open-access educational datasets are still far from sufficient to meet the need for various topics of adaptive learning. Existing released datasets often cover only small-scale data, lack fine-grained knowledge concepts. They are even difficult to curate and supplement due to platform limitations. In this work, we construct MOOCCubeX, a large, knowledge-centered repository consisting of 4,216 courses, 230,263 videos, 358,265 exercises, 637,572 fine-grained concepts and over 296 million behavioral data of 3,330,294 students, for supporting the research topics on adaptive learning in MOOCs. Licensed by XuetangX, one of the largest MOOC websites in China, we obtain abundant and diverse course resources and student behavioral data and are permitted to make subsequent periodic updates. We propose a framework to accomplish data processing, weakly supervised fine-grained concept graph mining, and data curation to improve usability and richness. Based on the fine-grained concepts, we re-organize the data from the knowledge perspective and acquire more external learning resources from the web. Our repository is now available at https://github.com/THU-KEG/MOOCCubeX.