Mengdi Zhang

Xiao Liu, Shiyu Zhao, Kai Su et al. · tsinghua

Knowledge graph (KG) embeddings have been a mainstream approach for reasoning over incomplete KGs. However, limited by their inherently shallow and static architectures, they can hardly deal with the rising focus on complex logical queries, which comprise logical operators, imputed edges, multiple source entities, and unknown intermediate entities. In this work, we present the Knowledge Graph Transformer (kgTransformer) with masked pre-training and fine-tuning strategies. We design a KG triple transformation method to enable Transformer to handle KGs, which is further strengthened by the Mixture-of-Experts (MoE) sparse activation. We then formulate the complex logical queries as masked prediction and introduce a two-stage masked pre-training strategy to improve transferability and generalizability. Extensive experiments on two benchmarks demonstrate that kgTransformer can consistently outperform both KG embedding-based baselines and advanced encoders on nine in-domain and out-of-domain reasoning tasks. Additionally, kgTransformer can reason with explainability via providing the full reasoning paths to interpret given answers.

Yejie Wang, Keqing He, Dayuan Fu et al.

Recently, there has been a growing interest in studying how to construct better code instruction tuning data. However, we observe Code models trained with these datasets exhibit high performance on HumanEval but perform worse on other benchmarks such as LiveCodeBench. Upon further investigation, we find that many datasets suffer from severe data leakage. After cleaning up most of the leaked data, some well-known high-quality datasets perform poorly. This discovery reveals a new challenge: identifying which dataset genuinely qualify as high-quality code instruction data. To address this, we propose an efficient code data pruning strategy for selecting good samples. Our approach is based on three dimensions: instruction complexity, response quality, and instruction diversity. Based on our selected data, we present XCoder, a family of models finetuned from LLaMA3. Our experiments show XCoder achieves new state-of-the-art performance using fewer training data, which verify the effectiveness of our data strategy. Moreover, we perform a comprehensive analysis on the data composition and find existing code datasets have different characteristics according to their construction methods, which provide new insights for future code LLMs. Our models and dataset are released in https://github.com/banksy23/XCoder

Kai Zhang, Qi Liu, Zhenya Huang et al.

Cross-domain sentiment classification (CDSC) aims to use the transferable semantics learned from the source domain to predict the sentiment of reviews in the unlabeled target domain. Existing studies in this task attach more attention to the sequence modeling of sentences while largely ignoring the rich domain-invariant semantics embedded in graph structures (i.e., the part-of-speech tags and dependency relations). As an important aspect of exploring characteristics of language comprehension, adaptive graph representations have played an essential role in recent years. To this end, in the paper, we aim to explore the possibility of learning invariant semantic features from graph-like structures in CDSC. Specifically, we present Graph Adaptive Semantic Transfer (GAST) model, an adaptive syntactic graph embedding method that is able to learn domain-invariant semantics from both word sequences and syntactic graphs. More specifically, we first raise a POS-Transformer module to extract sequential semantic features from the word sequences as well as the part-of-speech tags. Then, we design a Hybrid Graph Attention (HGAT) module to generate syntax-based semantic features by considering the transferable dependency relations. Finally, we devise an Integrated aDaptive Strategy (IDS) to guide the joint learning process of both modules. Extensive experiments on four public datasets indicate that GAST achieves comparable effectiveness to a range of state-of-the-art models.

Yongqiang Han, Likang Wu, Hao Wang et al.

Sequential Recommendation is a widely studied paradigm for learning users' dynamic interests from historical interactions for predicting the next potential item. Although lots of research work has achieved remarkable progress, they are still plagued by the common issues: data sparsity of limited supervised signals and data noise of accidentally clicking. To this end, several works have attempted to address these issues, which ignored the complex association of items across several sequences. Along this line, with the aim of learning representative item embedding to alleviate this dilemma, we propose GUESR, from the view of graph contrastive learning. Specifically, we first construct the Global Item Relationship Graph (GIRG) from all interaction sequences and present the Bucket-Cluster Sampling (BCS) method to conduct the sub-graphs. Then, graph contrastive learning on this reduced graph is developed to enhance item representations with complex associations from the global view. We subsequently extend the CapsNet module with the elaborately introduced target-attention mechanism to derive users' dynamic preferences. Extensive experimental results have demonstrated our proposed GUESR could not only achieve significant improvements but also could be regarded as a general enhancement strategy to improve the performance in combination with other sequential recommendation methods.

Jin Jiang, Yuchen Yan, Yang Liu et al.

In this paper, we propose a new data synthesis method called \textbf{LogicPro}, which leverages LeetCode-style algorithm \underline{Pro}blems and their corresponding \underline{Pro}gram solutions to synthesize Complex \underline{Logic}al Reasoning data in text format. First, we synthesize complex reasoning problems through source algorithm problems and test cases. Then, standard answers and intermediate variable outputs are obtained for each problem based on standard python solutions and test cases. Finally, with the guidance of code intermediate variables, we synthesize the text reasoning process for each reasoning problems. Through this method, we can synthesize data that is difficult, scalable, effective, and comes with golden standard answers and high-quality reasoning processes. As a result, with our 540K synthesized dataset constructed solely from 2,360 algorithm problems, our approach \footnote{Code and data are publicly available at https://github.com/jiangjin1999/LogicPro} achieves significant improvements in multiple models for the datasets \textit{BBH$^{27}$}, \textit{LogicBench}, \textit{DROP}, \textit{AR-LSAT}, and \textit{GSM8K}, etc. outperforming a wide range of existing reasoning datasets.

Mengdi Zhang, Jun Sun

Given a discriminating neural network, the problem of fairness improvement is to systematically reduce discrimination without significantly scarifies its performance (i.e., accuracy). Multiple categories of fairness improving methods have been proposed for neural networks, including pre-processing, in-processing and post-processing. Our empirical study however shows that these methods are not always effective (e.g., they may improve fairness by paying the price of huge accuracy drop) or even not helpful (e.g., they may even worsen both fairness and accuracy). In this work, we propose an approach which adaptively chooses the fairness improving method based on causality analysis. That is, we choose the method based on how the neurons and attributes responsible for unfairness are distributed among the input attributes and the hidden neurons. Our experimental evaluation shows that our approach is effective (i.e., always identify the best fairness improving method) and efficient (i.e., with an average time overhead of 5 minutes).

Likang Wu, Junji Jiang, Hongke Zhao et al.

Recently, Zero-Shot Node Classification (ZNC) has been an emerging and crucial task in graph data analysis. This task aims to predict nodes from unseen classes which are unobserved in the training process. Existing work mainly utilizes Graph Neural Networks (GNNs) to associate features' prototypes and labels' semantics thus enabling knowledge transfer from seen to unseen classes. However, the multi-faceted semantic orientation in the feature-semantic alignment has been neglected by previous work, i.e. the content of a node usually covers diverse topics that are relevant to the semantics of multiple labels. It's necessary to separate and judge the semantic factors that tremendously affect the cognitive ability to improve the generality of models. To this end, we propose a Knowledge-Aware Multi-Faceted framework (KMF) that enhances the richness of label semantics via the extracted KG (Knowledge Graph)-based topics. And then the content of each node is reconstructed to a topic-level representation that offers multi-faceted and fine-grained semantic relevancy to different labels. Due to the particularity of the graph's instance (i.e., node) representation, a novel geometric constraint is developed to alleviate the problem of prototype drift caused by node information aggregation. Finally, we conduct extensive experiments on several public graph datasets and design an application of zero-shot cross-domain recommendation. The quantitative results demonstrate both the effectiveness and generalization of KMF with the comparison of state-of-the-art baselines.

Yuling Wang, Hao Xu, Yanhua Yu et al.

It is well established that graph neural networks (GNNs) can be interpreted and designed from the perspective of optimization objective. With this clear optimization objective, the deduced GNNs architecture has sound theoretical foundation, which is able to flexibly remedy the weakness of GNNs. However, this optimization objective is only proved for GNNs with single-relational graph. Can we infer a new type of GNNs for multi-relational graphs by extending this optimization objective, so as to simultaneously solve the issues in previous multi-relational GNNs, e.g., over-parameterization? In this paper, we propose a novel ensemble multi-relational GNNs by designing an ensemble multi-relational (EMR) optimization objective. This EMR optimization objective is able to derive an iterative updating rule, which can be formalized as an ensemble message passing (EnMP) layer with multi-relations. We further analyze the nice properties of EnMP layer, e.g., the relationship with multi-relational personalized PageRank. Finally, a new multi-relational GNNs which well alleviate the over-smoothing and over-parameterization issues are proposed. Extensive experiments conducted on four benchmark datasets well demonstrate the effectiveness of the proposed model.

Huixuan Chi, Hao Xu, Hao Fu et al.

Modeling the evolution of user preference is essential in recommender systems. Recently, dynamic graph-based methods have been studied and achieved SOTA for recommendation, majority of which focus on user's stable long-term preference. However, in real-world scenario, user's short-term preference evolves over time dynamically. Although there exists sequential methods that attempt to capture it, how to model the evolution of short-term preference with dynamic graph-based methods has not been well-addressed yet. In particular: 1) existing methods do not explicitly encode and capture the evolution of short-term preference as sequential methods do; 2) simply using last few interactions is not enough for modeling the changing trend. In this paper, we propose Long Short-Term Preference Modeling for Continuous-Time Sequential Recommendation (LSTSR) to capture the evolution of short-term preference under dynamic graph. Specifically, we explicitly encode short-term preference and optimize it via memory mechanism, which has three key operations: Message, Aggregate and Update. Our memory mechanism can not only store one-hop information, but also trigger with new interactions online. Extensive experiments conducted on five public datasets show that LSTSR consistently outperforms many state-of-the-art recommendation methods across various lines.

Mengdi Zhang, Jun Sun, Jingyi Wang et al.

Discrimination has been shown in many machine learning applications, which calls for sufficient fairness testing before their deployment in ethic-relevant domains such as face recognition, medical diagnosis and criminal sentence. Existing fairness testing approaches are mostly designed for identifying individual discrimination, i.e., discrimination against individuals. Yet, as another widely concerning type of discrimination, testing against group discrimination, mostly hidden, is much less studied. To address the gap, in this work, we propose TESTSGD, an interpretable testing approach which systematically identifies and measures hidden (which we call `subtle' group discrimination} of a neural network characterized by conditions over combinations of the sensitive features. Specifically, given a neural network, TESTSGDfirst automatically generates an interpretable rule set which categorizes the input space into two groups exposing the model's group discrimination. Alongside, TESTSGDalso provides an estimated group fairness score based on sampling the input space to measure the degree of the identified subtle group discrimination, which is guaranteed to be accurate up to an error bound. We evaluate TESTSGDon multiple neural network models trained on popular datasets including both structured data and text data. The experiment results show that TESTSGDis effective and efficient in identifying and measuring such subtle group discrimination that has never been revealed before. Furthermore, we show that the testing results of TESTSGDcan guide generation of new samples to mitigate such discrimination through retraining with negligible accuracy drop.

Yuanchen Bei, Hao Xu, Sheng Zhou et al.

Dynamic graph data mining has gained popularity in recent years due to the rich information contained in dynamic graphs and their widespread use in the real world. Despite the advances in dynamic graph neural networks (DGNNs), the rich information and diverse downstream tasks have posed significant difficulties for the practical application of DGNNs in industrial scenarios. To this end, in this paper, we propose to address them by pre-training and present the Contrastive Pre-Training Method for Dynamic Graph Neural Networks (CPDG). CPDG tackles the challenges of pre-training for DGNNs, including generalization capability and long-short term modeling capability, through a flexible structural-temporal subgraph sampler along with structural-temporal contrastive pre-training schemes. Extensive experiments conducted on both large-scale research and industrial dynamic graph datasets show that CPDG outperforms existing methods in dynamic graph pre-training for various downstream tasks under three transfer settings.

Yuchen Yan, Jin Jiang, Yang Liu et al.

Self-correction is a novel method that can stimulate the potential reasoning abilities of large language models (LLMs). It involves detecting and correcting errors during the inference process when LLMs solve reasoning problems. However, recent works do not regard self-correction as a spontaneous and intrinsic capability of LLMs. Instead, such correction is achieved through post-hoc generation, external knowledge introduction, multi-model collaboration, and similar techniques. In this paper, we propose a series of mathematical LLMs called S$^3$c-Math, which are able to perform Spontaneous Step-level Self-correction for Mathematical reasoning. This capability helps LLMs to recognize whether their ongoing inference tends to contain errors and simultaneously correct these errors to produce a more reliable response. We proposed a method, which employs a step-level sampling approach to construct step-wise self-correction data for achieving such ability. Additionally, we implement a training strategy that uses above constructed data to equip LLMs with spontaneous step-level self-correction capacities. Our data and methods have been demonstrated to be effective across various foundation LLMs, consistently showing significant progress in evaluations on GSM8K, MATH, and other mathematical benchmarks. To the best of our knowledge, we are the first to introduce the spontaneous step-level self-correction ability of LLMs in mathematical reasoning.

Yuxuan Sun, Yuze Zhao, Yufeng Wang et al.

Evaluating software engineering capabilities has become a core component of modern large language models (LLMs); however, the key bottleneck hindering further scaling lies not in the scarcity of high-quality solutions, but in the lack of high-quality test suites. Test suites are indispensable both for synthesizing program repair trajectories and for providing precise feedback signals in reinforcement learning. Unfortunately, due to the high cost and difficulty of annotation, high-quality test suites have long been hard to obtain, while those automatically generated by LLMs tend to be superficial and lack sufficient discriminative power. As a first step toward constructing high-quality test suites, we introduce SWE-Mutation, a benchmark for evaluating LLM-generated test suites. The benchmark characterizes test suites by introducing systematically mutated solutions that attempt to ``fool'' the test suites and pass validation. We further propose an agentic, language-agnostic framework for automatically generating complex mutants. Our benchmark consists of 2,636 mutated variants derived from 800 original instances and includes a multilingual subset spanning nine programming languages. Experiments on seven LLMs reveal that even DeepSeek-V3.1 achieves only 10.20% verification and 36.15% detection rates, highlighting the inadequacy of current LLMs. Additionally, our agentic mutation strategy enhances realism, reducing average detection rates from 71.04% to 39.81% compared to conventional methods. These findings expose persistent deficiencies in the ability of current LLMs to generate reliable and discriminative test suites.

Zhihao Li, Yao Du, Yang Liu et al.

Multi-modal Large Language Models (MLLMs) have advanced greatly in general tasks. However, they still face challenges in geometric reasoning, a task that requires synergistic integration of visual recognition proficiency and complex reasoning strength. Existing MLLMs prioritize optimizing the LLM backbone to enhance problem-solving capabilities, while rarely emphasizing improvements in discerning visual elements. However, we reveal that MLLMs suffer from severe visual perception deficiencies, including inaccurate geometric comprehension and severe visual hallucinations, which constrain their reasoning performance. To address this issue, we revisit geometric reasoning through a visual-centric lens that highlights the role of visual perception. To achieve this, we propose EAGLE, a novel coarse-to-fine visual enhancement framework that progressively leverages LLMs' guidance to improve perception proficiency. Specifically, given the substantial disparity between geometric diagrams and natural images, we first introduce Geometric Knowledge Injection. This process explores fundamental knowledge from diagram-caption data to enhance recognition capabilities and improve geometry-language alignments. Then, recognizing that different elements contribute unequally in the reasoning process, we introduce Geometric Knowledge Refinement. This stage leverages LLM-driven chain-of-thought solutions to guide the vision encoder in adaptively prioritizing key elements, fostering a synergistic interplay between visual comprehension and mathematical reasoning. Finally, we develop EAGLE, a geometry expert with strong perception and reasoning capabilities. Extensive experiments demonstrate its effectiveness on three popular benchmarks.

Changyi Xiao, Mengdi Zhang, Yixin Cao

Recent studies, including DeepSeek-R1 and Kimi-k1.5, have demonstrated that reinforcement learning with rule-based, binary-valued reward functions can significantly enhance the reasoning capabilities of large language models. These models primarily utilize REINFORCE-based policy optimization techniques, such as REINFORCE with baseline and group relative policy optimization (GRPO). However, a key limitation remains: current policy optimization methods either neglect reward normalization or employ static normalization strategies, which fail to adapt to the dynamic nature of policy updates during training. This may result in unstable gradient estimates and hinder training stability. To address this issue, we propose Beta Normalization Policy Optimization (BNPO), a novel policy optimization method that adaptively normalizes rewards using a Beta distribution with dynamically updated parameters. BNPO aligns the normalization with the changing policy distribution, enabling more precise and lower-variance gradient estimation, which in turn promotes stable training dynamics. We provide theoretical analysis demonstrating BNPO's variance-reducing properties and show that it generalizes both REINFORCE and GRPO under binary-valued reward settings. Furthermore, we introduce an advantage decomposition mechanism to extend BNPO's applicability to more complex reward systems. Experimental results confirm that BNPO achieves state-of-the-art performance among policy optimization methods on reasoning tasks. The code is available at https://github.com/changyi7231/BNPO.

Jin Jiang, Jianing Wang, Yuchen Yan et al.

Large Language Models (LLMs) have been shown to achieve breakthrough performance on complex logical reasoning tasks. Nevertheless, most existing research focuses on employing formal language to guide LLMs to derive reliable reasoning paths, while systematic evaluations of these capabilities are still limited. In this paper, we aim to conduct a comprehensive evaluation of LLMs across various logical reasoning problems utilizing formal languages. From the perspective of three dimensions, i.e., spectrum of LLMs, taxonomy of tasks, and format of trajectories, our key findings are: 1) Thinking models significantly outperform Instruct models, especially when formal language is employed; 2) All LLMs exhibit limitations in inductive reasoning capability, irrespective of whether they use a formal language; 3) Data with PoT format achieves the best generalization performance across other languages. Additionally, we also curate the formal-relative training data to further enhance the small language models, and the experimental results indicate that a simple rejected fine-tuning method can better enable LLMs to generalize across formal languages and achieve the best overall performance. Our codes and reports are available at https://github.com/jiangjin1999/FormalEval.

Xinyu Zhang, Changzhi Zhou, Linmei Hu et al.

Existing code large language models (LLMs) often rely on large-scale instruction data distilled from proprietary LLMs for fine-tuning, which typically incurs high costs. In this paper, we explore the potential of small-scale open-source LLMs (e.g., 7B) as synthesizers for high-quality code instruction data construction. We first observe that the data synthesis capability of small-scale LLMs can be enhanced by training on a few superior data synthesis samples from proprietary LLMs. Building on this, we propose a novel iterative self-distillation approach to bootstrap small-scale LLMs, transforming them into powerful synthesizers that reduce reliance on proprietary LLMs and minimize costs. Concretely, in each iteration, to obtain diverse and high-quality self-distilled data, we design multi-checkpoint sampling and multi-aspect scoring strategies for initial data selection. Furthermore, to identify the most influential samples, we introduce a gradient-based influence estimation method for final data filtering. Based on the code instruction datasets from the small-scale synthesizers, we develop SCoder, a family of code generation models fine-tuned from DeepSeek-Coder. SCoder models achieve state-of-the-art code generation capabilities, demonstrating the effectiveness of our method.

Jianing Wang, Jin Jiang, Yang Liu et al.

In this paper, we introduce a new \emph{process prejudge} strategy in LLM reasoning to demonstrate that bootstrapping with process prejudge allows the LLM to adaptively anticipate the errors encountered when advancing the subsequent reasoning steps, similar to people sometimes pausing to think about what mistakes may occur and how to avoid them, rather than relying solely on trial and error. Specifically, we define a prejudge node in the rationale, which represents a reasoning step, with at least one step that follows the prejudge node that has no paths toward the correct answer. To synthesize the prejudge reasoning process, we present an automated reasoning framework with a dynamic tree-searching strategy. This framework requires only one LLM to perform answer judging, response critiquing, prejudge generation, and thought completion. Furthermore, we develop a two-phase training mechanism with supervised fine-tuning (SFT) and reinforcement learning (RL) to further enhance the reasoning capabilities of LLMs. Experimental results from competition-level complex reasoning demonstrate that our method can teach the model to prejudge before thinking and significantly enhance the reasoning ability of LLMs. Code and data is released at https://github.com/wjn1996/Prejudge-Before-Think.

Yufang Liu, Yao Du, Tao Ji et al.

Recent research has increasingly focused on multimodal mathematical reasoning, particularly emphasizing the creation of relevant datasets and benchmarks. Despite this, the role of visual information in reasoning has been underexplored. Our findings show that existing multimodal mathematical models minimally leverage visual information, and model performance remains largely unaffected by changes to or removal of images in the dataset. We attribute this to the dominance of textual information and answer options that inadvertently guide the model to correct answers. To improve evaluation methods, we introduce the HC-M3D dataset, specifically designed to require image reliance for problem-solving and to challenge models with similar, yet distinct, images that change the correct answer. In testing leading models, their failure to detect these subtle visual differences suggests limitations in current visual perception capabilities. Additionally, we observe that the common approach of improving general VQA capabilities by combining various types of image encoders does not contribute to math reasoning performance. This finding also presents a challenge to enhancing visual reliance during math reasoning. Our benchmark and code would be available at \href{https://github.com/Yufang-Liu/visual_modality_role}{https://github.com/Yufang-Liu/visual\_modality\_role}.

Yuchen Yan, Yongliang Shen, Yang Liu et al.

Advanced reasoning in large language models has achieved remarkable performance on challenging tasks, but the prevailing long-context reasoning paradigm faces critical limitations: quadratic computational scaling with sequence length, reasoning constrained by maximum context boundaries, and performance degradation beyond pre-training context windows. Existing approaches primarily compress reasoning chains without addressing the fundamental scaling problem. To overcome these challenges, we introduce InftyThink, a paradigm that transforms monolithic reasoning into an iterative process with intermediate summarization. By interleaving short reasoning segments with concise progress summaries, our approach enables unbounded reasoning depth while maintaining bounded computational costs. This creates a characteristic sawtooth memory pattern that significantly reduces computational complexity compared to traditional approaches. Furthermore, we develop a methodology for reconstructing long-context reasoning datasets into our iterative format, transforming OpenR1-Math into 333K training instances. Experiments across multiple model architectures demonstrate that our approach reduces computational costs while improving performance, with Qwen2.5-Math-7B showing 3-13% improvements across MATH500, AIME24, and GPQA_diamond benchmarks. Our work challenges the assumed trade-off between reasoning depth and computational efficiency, providing a more scalable approach to complex reasoning without architectural modifications.

Yejie Wang, Keqing He, Guanting Dong et al.

Code Large Language Models (Code LLMs) have demonstrated outstanding performance in code-related tasks. Several instruction tuning approaches have been proposed to boost the code generation performance of pre-trained Code LLMs. In this paper, we introduce a diverse instruction model (DolphCoder) with self-evaluating for code generation. It learns diverse instruction targets and combines a code evaluation objective to enhance its code generation ability. Our model achieves superior performance on the HumanEval and MBPP benchmarks, demonstrating new insights for future code instruction tuning work. Our key findings are: (1) Augmenting more diverse responses with distinct reasoning paths increases the code capability of LLMs. (2) Improving one's ability to evaluate the correctness of code solutions also enhances their ability to create it.

Liang Wang, Shu Wu, Qiang Liu et al.

Next point-of-interest (POI) recommendation aims to predict a user's next destination based on sequential check-in history and a set of POI candidates. Graph neural networks (GNNs) have demonstrated a remarkable capability in this endeavor by exploiting the extensive global collaborative signals present among POIs. However, most of the existing graph-based approaches construct graph structures based on pre-defined heuristics, failing to consider inherent hierarchical structures of POI features such as geographical locations and visiting peaks, or suffering from noisy and incomplete structures in graphs. To address the aforementioned issues, this paper presents a novel Bi-level Graph Structure Learning (BiGSL) for next POI recommendation. BiGSL first learns a hierarchical graph structure to capture the fine-to-coarse connectivity between POIs and prototypes, and then uses a pairwise learning module to dynamically infer relationships between POI pairs and prototype pairs. Based on the learned bi-level graphs, our model then employs a multi-relational graph network that considers both POI- and prototype-level neighbors, resulting in improved POI representations. Our bi-level structure learning scheme is more robust to data noise and incompleteness, and improves the exploration ability for recommendation by alleviating sparsity issues. Experimental results on three real-world datasets demonstrate the superiority of our model over existing state-of-the-art methods, with a significant improvement in recommendation accuracy and exploration performance.

Xiancai Chen, Zhengwei Tao, Kechi Zhang et al. · pku

Large language models (LLMs) have shown significant advancements in code generation, but still face challenges on tasks beyond their basic capabilities. Recently, the notion of self-debugging has been proposed to boost the performance of code generation by leveraging execution feedback from tests. Despite its promise, the availability of high-quality tests in real-world scenarios is limited. In this context, self-debugging with self-generated tests is a promising solution but lacks a full exploration of its limitations and practical potential. Therefore, we investigate its efficacy on diverse programming problems. To deepen our understanding, we propose two distinct paradigms for the process: post-execution and in-execution self-debugging. Within the scope of self-contained Python programming tasks, we find that post-execution self-debugging struggles on basic problems but shows potential for improvement on competitive ones, due to the bias introduced by self-generated tests. On the other hand, in-execution self-debugging enables LLMs to mitigate the bias by solely leveraging intermediate states during execution, thereby enhancing code generation.

Yuchen Yan, Jin Jiang, Zhenbang Ren et al.

Large reasoning models such as OpenAI o1 and DeepSeek-R1 have achieved remarkable performance in the domain of reasoning. A key component of their training is the incorporation of verifiable rewards within reinforcement learning (RL). However, existing reward benchmarks do not evaluate reference-based reward systems, leaving researchers with limited understanding of the accuracy of verifiers used in RL. In this paper, we introduce two benchmarks, VerifyBench and VerifyBench-Hard, designed to assess the performance of reference-based reward systems. These benchmarks are constructed through meticulous data collection and curation, followed by careful human annotation to ensure high quality. Current models still show considerable room for improvement on both VerifyBench and VerifyBench-Hard, especially smaller-scale models. Furthermore, we conduct a thorough and comprehensive analysis of evaluation results, offering insights for understanding and developing reference-based reward systems. Our proposed benchmarks serve as effective tools for guiding the development of verifier accuracy and the reasoning capabilities of models trained via RL in reasoning tasks.

Changzhi Zhou, Xinyu Zhang, Dandan Song et al.

Code generation has attracted increasing attention with the rise of Large Language Models (LLMs). Many studies have developed powerful code LLMs by synthesizing code-related instruction data and applying supervised fine-tuning. However, these methods are limited by teacher model distillation and ignore the potential of iterative refinement by self-generated code. In this paper, we propose Adaptive Critique Refinement (ACR), which enables the model to refine itself by self-generated code and external critique, rather than directly imitating the code responses of the teacher model. Concretely, ACR includes a composite scoring system with LLM-as-a-Judge to evaluate the quality of code responses and a selective critique strategy with LLM-as-a-Critic to critique self-generated low-quality code responses. We develop the RefineCoder series by iteratively applying ACR, achieving continuous performance improvement on multiple code generation benchmarks. Compared to the baselines of the same size, our proposed RefineCoder series can achieve comparable or even superior performance using less data.

Yuchen Yan, Yongliang Shen, Yang Liu et al.

Mathematical reasoning represents a critical frontier in advancing large language models (LLMs). While step-by-step approaches have emerged as the dominant paradigm for mathematical problem-solving in LLMs, the quality of reasoning steps in training data fundamentally constrains the performance of the models. Recent studies has demonstrated that more detailed intermediate steps can enhance model performance, yet existing methods for step expansion either require more powerful external models or incur substantial computational costs. In this paper, we introduce MathFimer, a novel framework for mathematical reasoning step expansion inspired by the "Fill-in-the-middle" task from code completion. By decomposing solution chains into prefix-suffix pairs and training models to reconstruct missing intermediate steps, we develop a specialized model, MathFimer-7B, on our carefully curated NuminaMath-FIM dataset. We then apply these models to enhance existing mathematical reasoning datasets by inserting detailed intermediate steps into their solution chains, creating MathFimer-expanded versions. Through comprehensive experiments on multiple mathematical reasoning datasets, including MathInstruct, MetaMathQA and etc., we demonstrate that models trained on MathFimer-expanded data consistently outperform their counterparts trained on original data across various benchmarks such as GSM8K and MATH. Our approach offers a practical, scalable solution for enhancing mathematical reasoning capabilities in LLMs without relying on powerful external models or expensive inference procedures.

Mengdi Zhang, Kai Kiat Goh, Peixin Zhang et al.

Despite the success of Large Language Models (LLMs) across various fields, their potential to generate untruthful, biased and harmful responses poses significant risks, particularly in critical applications. This highlights the urgent need for systematic methods to detect and prevent such misbehavior. While existing approaches target specific issues such as harmful responses, this work introduces LLMScan, an innovative LLM monitoring technique based on causality analysis, offering a comprehensive solution. LLMScan systematically monitors the inner workings of an LLM through the lens of causal inference, operating on the premise that the LLM's `brain' behaves differently when misbehaving. By analyzing the causal contributions of the LLM's input tokens and transformer layers, LLMScan effectively detects misbehavior. Extensive experiments across various tasks and models reveal clear distinctions in the causal distributions between normal behavior and misbehavior, enabling the development of accurate, lightweight detectors for a variety of misbehavior detection tasks.

Yu Yuan, Lili Zhao, Wei Chen et al.

Capturing human learning behavior based on deep learning methods has become a major research focus in both psychology and intelligent systems. Recent approaches rely on controlled experiments or rule-based models to explore cognitive processes. However, they struggle to capture learning dynamics, track progress over time, or provide explainability. To address these challenges, we introduce LearnerAgent, a novel multi-agent framework based on Large Language Models (LLMs) to simulate a realistic teaching environment. To explore human-like learning dynamics, we construct learners with psychologically grounded profiles-such as Deep, Surface, and Lazy-as well as a persona-free General Learner to inspect the base LLM's default behavior. Through weekly knowledge acquisition, monthly strategic choices, periodic tests, and peer interaction, we can track the dynamic learning progress of individual learners over a full-year journey. Our findings are fourfold: 1) Longitudinal analysis reveals that only Deep Learner achieves sustained cognitive growth. Our specially designed "trap questions" effectively diagnose Surface Learner's shallow knowledge. 2) The behavioral and cognitive patterns of distinct learners align closely with their psychological profiles. 3) Learners' self-concept scores evolve realistically, with the General Learner developing surprisingly high self-efficacy despite its cognitive limitations. 4) Critically, the default profile of base LLM is a "diligent but brittle Surface Learner"-an agent that mimics the behaviors of a good student but lacks true, generalizable understanding. Extensive simulation experiments demonstrate that LearnerAgent aligns well with real scenarios, yielding more insightful findings about LLMs' behavior.

Kang Chen, Mengdi Zhang, Yixin Cao

This paper explores the challenges of test-time scaling of large language models (LLMs), regarding both the data and inference efficiency. We highlight the diversity of multi-lingual reasoning based on our pilot studies, and then introduce a novel approach, \(L^2\) multi-lingual unification learning with a decoding intervention strategy for further investigation. The basic idea of \(L^2\) is that the reasoning process varies across different languages, which may be mutually beneficial to enhance both model performance and efficiency. In specific, there are two types of multi-lingual data: the entire long chain-of-thought annotations in different languages and the step-wise mixture of languages. By further tuning based on them, we show that even small amounts of data can significantly improve reasoning capabilities. Our findings suggest that multilingual learning reduces both the required data and the number of inference tokens while maintaining a comparable performance. Furthermore, \(L^2\) is orthogonal to other data efficient methods. Thus, we also emphasize the importance of diverse data selection. The \(L^2\) method offers a promising solution to the challenges of data collection and test-time compute efficiency in LLMs.

Bin Hong, Jiayu Liu, Zhenya Huang et al.

Recent advances in Large Reasoning Models (LRMs) have demonstrated strong performance on complex tasks through long Chain-of-Thought (CoT) reasoning. However, their lengthy outputs increase computational costs and may lead to overthinking, raising challenges in balancing reasoning effectiveness and efficiency. Current methods for efficient reasoning often compromise reasoning quality or require extensive resources. This paper investigates efficient methods to reduce the generation length of LRMs. We analyze generation path distributions and filter generated trajectories through difficulty estimation. Subsequently, we analyze the convergence behaviors of the objectives of various preference optimization methods under a Bradley-Terry loss based framework. Based on the analysis, we propose Length Controlled Preference Optimization (LCPO) that directly balances the implicit reward related to NLL loss. LCPO can effectively learn length preference with limited data and training. Extensive experiments demonstrate that our approach significantly reduces the average output length by over 50\% across multiple benchmarks while maintaining the reasoning performance. Our work highlights the potential for computationally efficient approaches in guiding LRMs toward efficient reasoning.

Shibo Hong, Jiahao Ying, Haiyuan Liang et al.

Evaluating open-ended outputs of Multimodal Large Language Models has become a bottleneck as model capabilities, task diversity, and modality rapidly expand. Existing ``MLLM-as-a-Judge'' evaluators, though promising, remain constrained to specific tasks and aspects. In this paper, we argue that, on one hand, based on the interconnected nature of aspects, learning specific aspects can generalize to unseen aspects; on the other hand, jointly learning to assess multiple visual aspects and tasks may foster a synergistic effect. To this end, we propose UFEval, the first unified fine-grained evaluator with task and aspect generalization for four evaluation tasks -- Natural Language Generation, Image Understanding, Image Generation, and Interleaved Text-and-Image Generation. However, training such a unified evaluator is hindered by the lack of a large-scale, multi-modal, and aspect-level resource. To address this gap, we introduce FRABench, a comprehensive fine-grained evaluation dataset. Specifically, (1) We first construct a hierarchical aspect taxonomy encompassing 112 distinct aspects across the aforementioned four tasks. (2) Based on this taxonomy, we create FRABench, comprising 60.4k pairwise samples with 325k evaluation labels obtained from a combination of human and GPT-4o annotations. (3) Finally, leveraging FRABench, we develop UFEval, a unified fine-grained evaluator. Experiments show that learning on specific aspects enables UFEval to generalize to unseen aspects, and joint learning to assess diverse visual tasks and aspects can lead to substantial mutual benefits.

Kai Zhang, Kun Zhang, Mengdi Zhang et al.

Aspect-based sentiment analysis (ABSA) predicts sentiment polarity towards a specific aspect in the given sentence. While pre-trained language models such as BERT have achieved great success, incorporating dynamic semantic changes into ABSA remains challenging. To this end, in this paper, we propose to address this problem by Dynamic Re-weighting BERT (DR-BERT), a novel method designed to learn dynamic aspect-oriented semantics for ABSA. Specifically, we first take the Stack-BERT layers as a primary encoder to grasp the overall semantic of the sentence and then fine-tune it by incorporating a lightweight Dynamic Re-weighting Adapter (DRA). Note that the DRA can pay close attention to a small region of the sentences at each step and re-weigh the vitally important words for better aspect-aware sentiment understanding. Finally, experimental results on three benchmark datasets demonstrate the effectiveness and the rationality of our proposed model and provide good interpretable insights for future semantic modeling.

Likang Wu, Zhi Li, Hongke Zhao et al.

Existing representation learning methods in graph convolutional networks are mainly designed by describing the neighborhood of each node as a perceptual whole, while the implicit semantic associations behind highly complex interactions of graphs are largely unexploited. In this paper, we propose a Semantic Graph Convolutional Networks (SGCN) that explores the implicit semantics by learning latent semantic-paths in graphs. In previous work, there are explorations of graph semantics via meta-paths. However, these methods mainly rely on explicit heterogeneous information that is hard to be obtained in a large amount of graph-structured data. SGCN first breaks through this restriction via leveraging the semantic-paths dynamically and automatically during the node aggregating process. To evaluate our idea, we conduct sufficient experiments on several standard datasets, and the empirical results show the superior performance of our model.

Chenyang Lei, Xuhua Huang, Mengdi Zhang et al.

We present a novel formulation to removing reflection from polarized images in the wild. We first identify the misalignment issues of existing reflection removal datasets where the collected reflection-free images are not perfectly aligned with input mixed images due to glass refraction. Then we build a new dataset with more than 100 types of glass in which obtained transmission images are perfectly aligned with input mixed images. Second, capitalizing on the special relationship between reflection and polarized light, we propose a polarized reflection removal model with a two-stage architecture. In addition, we design a novel perceptual NCC loss that can improve the performance of reflection removal and general image decomposition tasks. We conduct extensive experiments, and results suggest that our model outperforms state-of-the-art methods on reflection removal.

Hongwei Wang, Fuzheng Zhang, Mengdi Zhang et al.

Knowledge graphs capture structured information and relations between a set of entities or items. As such knowledge graphs represent an attractive source of information that could help improve recommender systems. However, existing approaches in this domain rely on manual feature engineering and do not allow for an end-to-end training. Here we propose Knowledge-aware Graph Neural Networks with Label Smoothness regularization (KGNN-LS) to provide better recommendations. Conceptually, our approach computes user-specific item embeddings by first applying a trainable function that identifies important knowledge graph relationships for a given user. This way we transform the knowledge graph into a user-specific weighted graph and then apply a graph neural network to compute personalized item embeddings. To provide better inductive bias, we rely on label smoothness assumption, which posits that adjacent items in the knowledge graph are likely to have similar user relevance labels/scores. Label smoothness provides regularization over the edge weights and we prove that it is equivalent to a label propagation scheme on a graph. We also develop an efficient implementation that shows strong scalability with respect to the knowledge graph size. Experiments on four datasets show that our method outperforms state of the art baselines. KGNN-LS also achieves strong performance in cold-start scenarios where user-item interactions are sparse.

Zhe Liu, Anbang Xu, Mengdi Zhang et al.

One problem that every presenter faces when delivering a public discourse is how to hold the listeners' attentions or to keep them involved. Therefore, many studies in conversation analysis work on this issue and suggest qualitatively con-structions that can effectively lead to audience's applause. To investigate these proposals quantitatively, in this study we an-alyze the transcripts of 2,135 TED Talks, with a particular fo-cus on the rhetorical devices that are used by the presenters for applause elicitation. Through conducting regression anal-ysis, we identify and interpret 24 rhetorical devices as triggers of audience applauding. We further build models that can rec-ognize applause-evoking sentences and conclude this work with potential implications.