Chunyang Li

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.

Shangqing Tu, Zheyuan Zhang, Jifan Yu et al. · tsinghua

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.

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}

Eduardo Pérez-Pellitero, Sibi Catley-Chandar, Richard Shaw et al.

This paper reviews the challenge on constrained high dynamic range (HDR) imaging that was part of the New Trends in Image Restoration and Enhancement (NTIRE) workshop, held in conjunction with CVPR 2022. This manuscript focuses on the competition set-up, datasets, the proposed methods and their results. The challenge aims at estimating an HDR image from multiple respective low dynamic range (LDR) observations, which might suffer from under- or over-exposed regions and different sources of noise. The challenge is composed of two tracks with an emphasis on fidelity and complexity constraints: In Track 1, participants are asked to optimize objective fidelity scores while imposing a low-complexity constraint (i.e. solutions can not exceed a given number of operations). In Track 2, participants are asked to minimize the complexity of their solutions while imposing a constraint on fidelity scores (i.e. solutions are required to obtain a higher fidelity score than the prescribed baseline). Both tracks use the same data and metrics: Fidelity is measured by means of PSNR with respect to a ground-truth HDR image (computed both directly and with a canonical tonemapping operation), while complexity metrics include the number of Multiply-Accumulate (MAC) operations and runtime (in seconds).

Wei Fan, Yining Zhou, Mufan Zhang et al.

While large language models (LLMs) augmented with agentic search capabilities show promise for legal reasoning, they overlook a fundamental constraint that applicable law must match the temporal context of each case, as retroactive application of statutes violates core legal principles and leads to erroneous conclusions. Our observations reveal that current legal LLMs suffer from temporal bias anchored to their training cutoff, while search agents rarely incorporate temporal constraints into queries, and that web search alone cannot provide the precise statute and precedent citations that legal reasoning demands. To address these challenges, we propose LegalSearch-R1, an end-to-end reinforcement learning framework that pairs local statute RAG for precise article matching with online web search for broader legal knowledge, trained on temporally-indexed data spanning multiple amendment periods to enforce temporal consistency. Extensive experiments on our benchmark covering 13 legal tasks demonstrate that our 7B-parameter agent outperforms state-of-the-art deep research frameworks and specialized legal LLMs by 12.9% to 29.8%, surpasses baselines by 57.7% to 80.3% on temporal consistency, and exhibits robust out-of-domain generalization. The code and data are available at https://github.com/AlexFanw/LegalSearch-R1.

Jiaxuan Liu, Yunkang Cao, Yufeng Chen et al.

The deployment of zero-shot anomaly detection (AD) in embodied industrial inspection is severely bottlenecked by its reliance on passive, fixed-viewpoint 2D imagery. Such formulations inherently fail to accommodate the active, dynamic observations required in real-world environments. To break this limitation, we introduce Real-to-Twin Anomaly Detection, a novel task that evaluates physical observations directly against geometrically matched CAD Digital Twins. To tackle this new task, we propose AVATAR, a framework designed to learn robust semantic alignment between Real and Digital Twins. By bridging benign Sim2Real domain gaps using only defect-free pairs, AVATAR effectively transforms CAD priors into dynamic, anomaly-free references. This elegant formulation enables the model to localize diverse anomalies in a zero-shot manner as unalignable deviations, eliminating the need for defect annotations. Extensive experiments demonstrate that AVATAR substantially outperforms adapted state-of-the-art baselines, exhibiting exceptional robustness to severe viewpoint variations. The code and dataset will be made publicly available.

Weiqi Wang, Tianqing Fang, Chunyang Li et al. · tencent-ai

The sequential process of conceptualization and instantiation is essential to generalizable commonsense reasoning as it allows the application of existing knowledge to unfamiliar scenarios. However, existing works tend to undervalue the step of instantiation and heavily rely on pre-built concept taxonomies and human annotations to collect both types of knowledge, resulting in a lack of instantiated knowledge to complete reasoning, high cost, and limited scalability. To tackle these challenges, we introduce CANDLE, a distillation framework that iteratively performs contextualized conceptualization and instantiation over commonsense knowledge bases by instructing large language models to generate both types of knowledge with critic filtering. By applying CANDLE to ATOMIC, we construct a comprehensive knowledge base comprising six million conceptualizations and instantiated commonsense knowledge triples. Both types of knowledge are firmly rooted in the original ATOMIC dataset, and intrinsic evaluations demonstrate their exceptional quality and diversity. Empirical results indicate that distilling CANDLE on student models provides benefits across four downstream tasks. Our code, data, and models are publicly available at https://github.com/HKUST-KnowComp/CANDLE.

Jiayu Liu, Rui Wang, Qing Zong et al.

Accurately assessing model confidence is essential for deploying large language models (LLMs) in mission-critical factual domains. While retrieval-augmented generation (RAG) is widely adopted to improve grounding, confidence calibration in RAG settings remains poorly understood. We conduct a systematic study across four benchmarks, revealing that LLMs exhibit poor calibration performance due to noisy retrieved contexts. Specifically, contradictory or irrelevant evidence tends to inflate the model's false certainty, leading to severe overconfidence. To address this, we propose NAACL Rules (Noise-AwAre Confidence CaLibration Rules) to provide a principled foundation for resolving overconfidence under noise. We further design NAACL, a noise-aware calibration framework that synthesizes supervision from about 2K HotpotQA examples guided by these rules. By performing supervised fine-tuning (SFT) with this data, NAACL equips models with intrinsic noise awareness without relying on stronger teacher models. Empirical results show that NAACL yields substantial gains, improving ECE scores by 10.9% in-domain and 8.0% out-of-domain. By bridging the gap between retrieval noise and verbal calibration, NAACL paves the way for both accurate and epistemically reliable LLMs.

Chunyang Li, Weiqi Wang, Tianshi Zheng et al.

Inductive reasoning, a cornerstone of human cognition, enables generalization from limited data but hasn't yet been fully achieved by large language models (LLMs). While modern LLMs excel at reasoning tasks, their ability to maintain stable and consistent rule abstraction under imperfect observations remains underexplored. To fill this gap, in this work, we introduce Robust Rule Induction, a task that evaluates LLMs' capability in inferring rules from data that are fused with noisy examples. To address this task, we further propose Sample-steered Rule Refinement (SRR), a method enhancing reasoning stability via observation diversification and execution-guided feedback. Experiments across arithmetic, cryptography, and list functions reveal: (1) SRR outperforms other methods with minimal performance degradation under noise; (2) Despite slight accuracy variation, LLMs exhibit instability under noise (e.g., 0% accuracy change with only 70% consistent score); (3) Counterfactual task gaps highlight LLMs' reliance on memorized patterns over genuine abstraction. Our findings challenge LLMs' reasoning robustness, revealing susceptibility to hypothesis drift and pattern overfitting, while providing empirical evidence critical for developing human-like inductive systems. Code and data are available at https://github.com/HKUST-KnowComp/Robust-Rule-Induction.

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.

Haochen Shi, Tianshi Zheng, Weiqi Wang et al.

Large Language Model (LLM) routing is a pivotal technique for navigating a diverse landscape of LLMs, aiming to select the best-performing LLMs tailored to the domains of user queries, while managing computational resources. However, current routing approaches often face limitations in scalability when dealing with a large pool of specialized LLMs, or in their adaptability to extending model scope and evolving capability domains. To overcome those challenges, we propose InferenceDynamics, a flexible and scalable multi-dimensional routing framework by modeling the capability and knowledge of models. We operate it on our comprehensive dataset RouteMix, and demonstrate its effectiveness and generalizability in group-level routing using modern benchmarks including MMLU-Pro, GPQA, BigGenBench, and LiveBench, showcasing its ability to identify and leverage top-performing models for given tasks, leading to superior outcomes with efficient resource utilization. The broader adoption of Inference Dynamics can empower users to harness the full specialized potential of the LLM ecosystem, and our code will be made publicly available to encourage further research.

Chunyang Li, Yilun Zheng, Xinting Huang et al. · tencent-ai

The paradigm of LLM-as-a-judge is emerging as a scalable and efficient alternative to human evaluation, demonstrating strong performance on well-defined tasks. However, its reliability in open-ended tasks with dynamic environments and complex interactions remains unexplored. To bridge the gap, we introduce WebDevJudge, a systematic benchmark for assessing LLM-as-a-judge performance in web development, with support for both non-interactive evaluation based on static observations and continuous interactive evaluation with a dynamic web environment. WebDevJudge comprises human preference labels over paired web implementations, annotated with structured and query-grounded rubrics to ensure high-quality ground truth. Using this benchmark, we comprehensively evaluate various evaluators, including LLMs, MLLMs, and agentic workflows. We systematically investigate the impact of different paradigms and guidance mechanisms. Our experiments reveal a significant gap between LLM judges and human experts. In-depth analysis indicates this gap stems from fundamental model limitations, including failures in recognizing functional equivalence, verifying task feasibility, and mitigating bias. Overall, WebDevJudge presents a significant challenge to LLM-as-a-judge, offering insights to guide future research toward developing more reliable and capable automated evaluators for complicated scenarios. Code and data are available at https://github.com/lcy2723/WebDevJudge.

Tianshi Zheng, Jiayang Cheng, Chunyang Li et al.

Modern large language models (LLMs) employ diverse logical inference mechanisms for reasoning, making the strategic optimization of these approaches critical for advancing their capabilities. This paper systematically investigate the comparative dynamics of inductive (System 1) versus abductive/deductive (System 2) inference in LLMs. We utilize a controlled analogical reasoning environment, varying modality (textual, visual, symbolic), difficulty, and task format (MCQ / free-text). Our analysis reveals System 2 pipelines generally excel, particularly in visual/symbolic modalities and harder tasks, while System 1 is competitive for textual and easier problems. Crucially, task format significantly influences their relative advantage, with System 1 sometimes outperforming System 2 in free-text rule-execution. These core findings generalize to broader in-context learning. Furthermore, we demonstrate that advanced System 2 strategies like hypothesis selection and iterative refinement can substantially scale LLM reasoning. This study offers foundational insights and actionable guidelines for strategically deploying logical inference to enhance LLM reasoning. Resources are available at https://github.com/HKUST-KnowComp/LogiDynamics.

Tianshi Zheng, Yixiang Chen, Chengxi Li et al.

Chain-of-Thought (CoT) prompting has been widely recognized for its ability to enhance reasoning capabilities in large language models (LLMs). However, our study reveals a surprising contradiction to this prevailing perspective within the fundamental domain of pattern-based in-context learning (ICL). Through extensive experiments involving 16 state-of-the-art LLMs and nine diverse pattern-based ICL datasets, we demonstrate that CoT and its reasoning variants consistently underperform direct answering across varying model scales and benchmark complexities. To systematically investigate this unexpected phenomenon, we designed extensive experiments to validate several hypothetical explanations. Our analysis uncovers a fundamental hybrid mechanism of explicit-implicit reasoning driving CoT's performance in pattern-based ICL: while explicit reasoning falters due to LLMs' struggles to infer underlying patterns from demonstrations, implicit reasoning-disrupted by the increased contextual distance of CoT rationales-often compensates, delivering correct answers despite flawed rationales. This hybrid mechanism explains CoT's relative underperformance, as noise from weak explicit inference undermines the process, even as implicit mechanisms partially salvage outcomes. Notably, even long-CoT reasoning models, which excel in abstract and symbolic reasoning, fail to fully overcome these limitations despite higher computational costs. Our findings challenge existing assumptions regarding the universal efficacy of CoT, yielding novel insights into its limitations and guiding future research toward more nuanced and effective reasoning methodologies for LLMs.

Jiaxin Bai, Wei Fan, Qi Hu et al.

We present AutoSchemaKG, a framework for fully autonomous knowledge graph construction that eliminates the need for predefined schemas. Our system leverages large language models to simultaneously extract knowledge triples and induce comprehensive schemas directly from text, modeling both entities and events while employing conceptualization to organize instances into semantic categories. Processing over 50 million documents, we construct ATLAS (Automated Triple Linking And Schema induction), a family of knowledge graphs with 900+ million nodes and 5.9 billion edges. This approach outperforms state-of-the-art baselines on multi-hop QA tasks and enhances LLM factuality. Notably, our schema induction achieves 92\% semantic alignment with human-crafted schemas with zero manual intervention, demonstrating that billion-scale knowledge graphs with dynamically induced schemas can effectively complement parametric knowledge in large language models.

Wei Fan, Tianshi Zheng, Yiran Hu et al.

Legal rules encompass not only codified statutes but also implicit adjudicatory principles derived from precedents that contain discretionary norms, social morality, and policy. While computational legal research has advanced in applying established rules to cases, inducing legal rules from judicial decisions remains understudied, constrained by limitations in model inference efficacy and symbolic reasoning capability. The advent of Large Language Models (LLMs) offers unprecedented opportunities for automating the extraction of such latent principles, yet progress is stymied by the absence of formal task definitions, benchmark datasets, and methodologies. To address this gap, we formalize Legal Rule Induction (LRI) as the task of deriving concise, generalizable doctrinal rules from sets of analogous precedents, distilling their shared preconditions, normative behaviors, and legal consequences. We introduce the first LRI benchmark, comprising 5,121 case sets (38,088 Chinese cases in total) for model tuning and 216 expert-annotated gold test sets. Experimental results reveal that: 1) State-of-the-art LLMs struggle with over-generalization and hallucination; 2) Training on our dataset markedly enhances LLMs capabilities in capturing nuanced rule patterns across similar cases.

Baixuan Xu, Chunyang Li, Weiqi Wang et al.

Designing effective collaboration structure for multi-agent LLM systems to enhance collective reasoning is crucial yet remains under-explored. In this paper, we systematically investigate how collaborative reasoning performance is affected by three key design dimensions: (1) Expertise-Domain Alignment, (2) Collaboration Paradigm (structured workflow vs. diversity-driven integration), and (3) System Scale. Our findings reveal that expertise alignment benefits are highly domain-contingent, proving most effective for contextual reasoning tasks. Furthermore, collaboration focused on integrating diverse knowledge consistently outperforms rigid task decomposition. Finally, we empirically explore the impact of scaling the multi-agent system with expertise specialization and study the computational trade off, highlighting the need for more efficient communication protocol design. This work provides concrete guidelines for configuring specialized multi-agent system and identifies critical architectural trade-offs and bottlenecks for scalable multi-agent reasoning. The code will be made available upon acceptance.

Rongwei Lu, Yutong Jiang, Jinrui Zhang et al.

Gradient compression can effectively alleviate communication bottlenecks in Federated Learning (FL). Contemporary state-of-the-art sparse compressors, such as Top-$k$, exhibit high computational complexity, up to $\mathcal{O}(d\log_2{k})$, where $d$ is the number of model parameters. The hard-threshold compressor, which simply transmits elements with absolute values higher than a fixed threshold, is thus proposed to reduce the complexity to $\mathcal{O}(d)$. However, the hard-threshold compression causes accuracy degradation in FL, where the datasets are non-IID and the stepsize $γ$ is decreasing for model convergence. The decaying stepsize reduces the updates and causes the compression ratio of the hard-threshold compression to drop rapidly to an aggressive ratio. At or below this ratio, the model accuracy has been observed to degrade severely. To address this, we propose $γ$-FedHT, a stepsize-aware low-cost compressor with Error-Feedback to guarantee convergence. Given that the traditional theoretical framework of FL does not consider Error-Feedback, we introduce the fundamental conversation of Error-Feedback. We prove that $γ$-FedHT has the convergence rate of $\mathcal{O}(\frac{1}{T})$ ($T$ representing total training iterations) under $μ$-strongly convex cases and $\mathcal{O}(\frac{1}{\sqrt{T}})$ under non-convex cases, \textit{same as FedAVG}. Extensive experiments demonstrate that $γ$-FedHT improves accuracy by up to $7.42\%$ over Top-$k$ under equal communication traffic on various non-IID image datasets.

Qing Zong, Jiayu Liu, Tianshi Zheng et al.

Accurate confidence calibration in Large Language Models (LLMs) is critical for safe use in high-stakes domains, where clear verbalized confidence enhances user trust. Traditional methods that mimic reference confidence expressions often fail to capture the reasoning needed for accurate confidence assessment. We propose natural language critiques as a solution, ideally suited for confidence calibration, as precise gold confidence labels are hard to obtain and often require multiple generations. This paper studies how natural language critiques can enhance verbalized confidence, addressing: (1) What to critique: uncertainty (question-focused) or confidence (answer-specific)? Analysis shows confidence suits multiple-choice tasks, while uncertainty excels in open-ended scenarios. (2) How to critique: self-critique or critique calibration training? We propose Self-Critique, enabling LLMs to critique and optimize their confidence beyond mere accuracy, and CritiCal, a novel Critique Calibration training method that leverages natural language critiques to improve confidence calibration, moving beyond direct numerical optimization. Experiments show that CritiCal significantly outperforms Self-Critique and other competitive baselines, even surpassing its teacher model, GPT-4o, in complex reasoning tasks. CritiCal also shows robust generalization in out-of-distribution settings, advancing LLM's reliability.

Rongwei Lu, Jingyan Jiang, Chunyang Li et al.

Distributed machine learning in high end-to-end latency and low, varying bandwidth network environments undergoes severe throughput degradation. Due to its low communication requirements, distributed SGD (D-SGD) remains the mainstream optimizer in such challenging networks, but it still suffers from significant throughput reduction. To mitigate these limitations, existing approaches typically employ gradient compression and delayed aggregation to alleviate low bandwidth and high latency, respectively. To address both challenges simultaneously, these strategies are often combined, introducing a complex three-way trade-off among compression ratio, staleness (delayed synchronization steps), and model convergence rate. To achieve the balance under varying bandwidth conditions, an adaptive policy is required to dynamically adjust these parameters. Unfortunately, existing works rely on static heuristic strategies due to the lack of theoretical guidance, which prevents them from achieving this goal. This study fills in this theoretical gap by introducing a new theoretical tool, decomposing the joint optimization problem into a traditional convergence rate analysis with multiple analyzable noise terms. We are the first to reveal that staleness exponentially amplifies the negative impact of gradient compression on training performance, filling a critical gap in understanding how compressed and delayed gradients affect training. Furthermore, by integrating the convergence rate with a network-aware time minimization condition, we propose DeCo-SGD, which dynamically adjusts the compression ratio and staleness based on the real-time network condition and training task. DeCo-SGD achieves up to 5.07 and 1.37 speed-ups over D-SGD and static strategy in high-latency and low, varying bandwidth networks, respectively.

Xiaotian Zhang, Chunyang Li, Yi Zong et al.

Large Language Models(LLMs) have demonstrated remarkable performance across various natural language processing tasks; however, how to comprehensively and accurately assess their performance becomes an urgent issue to be addressed. This paper introduces GAOKAO-Bench, an intuitive benchmark that employs questions from the Chinese GAOKAO examination as test samples, including both subjective and objective questions. To align with human examination methods, we design a method based on zero-shot settings to evaluate the performance of LLMs. With human evaluation, we obtain the converted total score of LLMs, including GPT-4, ChatGPT and ERNIE-Bot.Our findings reveal that LLMs have achieved competitive scores in Chinese GAOKAO examination, while they exhibit significant performance disparities across various subjects. We also use LLMs to grade the subjective questions, and find that model scores achieve a moderate level of consistency with human scores. In conclusion, this research contributes a robust evaluation benchmark for future large language models and offers valuable insights into the advantages and limitations of such models.