Xuanming Zhang

Jia Li, Ge Li, Xuanming Zhang et al. · pku

How to evaluate Large Language Models (LLMs) in code generation remains an open question. Existing benchmarks have two limitations - data leakage and lack of domain-specific evaluation. The former hurts the fairness of benchmarks, and the latter hinders practitioners from selecting superior LLMs for specific programming domains. To address these two limitations, we propose a new benchmark - EvoCodeBench, which has the following advances: (1) Evolving data. EvoCodeBench will be dynamically updated every period (e.g., 6 months) to avoid data leakage. This paper releases the first version - EvoCodeBench-2403, containing 275 samples from 25 repositories. (2) A domain taxonomy and domain labels. Based on the statistics of open-source communities, we design a programming domain taxonomy consisting of 10 popular domains. Based on the taxonomy, we annotate each sample in EvoCodeBench with a domain label. (3) Domain-specific evaluations. Besides the Pass@k, we compute the Domain-Specific Improvement (DSI) and define LLMs' comfort and strange domains. These evaluations help practitioners select superior LLMs in specific domains and discover the shortcomings of existing LLMs. We evaluate 8 popular LLMs (e.g., gpt-4, DeepSeek Coder) on EvoCodeBench and summarize some insights. EvoCodeBench reveals the actual abilities of these LLMs in real-world repositories. For example, the highest Pass@1 of gpt-4 on EvoCodeBench-2403 is only 20.74%. Besides, we evaluate LLMs in different domains and discover their comfort and strange domains. For example, gpt-4 performs best in most domains but falls behind others in the Internet domain. StarCoder 2-15B unexpectedly performs well in the Database domain and even outperforms 33B LLMs. EvoCodeBench has been released.

Jia Li, Ge Li, Yunfei Zhao et al. · pku

How to evaluate Large Language Models (LLMs) in code generation is an open question. Many benchmarks have been proposed but are inconsistent with practical software projects, e.g., unreal program distributions, insufficient dependencies, and small-scale project contexts. Thus, the capabilities of LLMs in practical projects are still unclear. In this paper, we propose a new benchmark named DevEval, aligned with Developers' experiences in practical projects. DevEval is collected through a rigorous pipeline, containing 2,690 samples from 119 practical projects and covering 10 domains. Compared to previous benchmarks, DevEval aligns to practical projects in multiple dimensions, e.g., real program distributions, sufficient dependencies, and enough-scale project contexts. We assess five popular LLMs on DevEval (e.g., gpt-4, gpt-3.5-turbo, CodeLLaMa, and StarCoder) and reveal their actual abilities in code generation. For instance, the highest Pass@1 of gpt-3.5-turbo only is 42 in our experiments. We also discuss the challenges and future directions of code generation in practical projects. We open-source DevEval and hope it can facilitate the development of code generation in practical projects.

Xuanming Zhang, Yuxuan Chen, Samuel Yeh et al.

Human social interactions depend on the ability to infer others' unspoken intentions, emotions, and beliefs-a cognitive skill grounded in the psychological concept of Theory of Mind (ToM). While large language models (LLMs) excel in semantic understanding tasks, they struggle with the ambiguity and contextual nuance inherent in human communication. To bridge this gap, we introduce MetaMind, a multi-agent framework inspired by psychological theories of metacognition, designed to emulate human-like social reasoning. MetaMind decomposes social understanding into three collaborative stages: (1) a Theory-of-Mind Agent generates hypotheses about user mental states (e.g., intent, emotion), (2) a Moral Agent refines these hypotheses using cultural norms and ethical constraints, and (3) a Response Agent generates contextually appropriate responses while validating alignment with inferred intent. Our framework achieves state-of-the-art performance across three challenging benchmarks, with 35.7% improvement in real-world social scenarios and 6.2% gain in ToM reasoning. Notably, it enables LLMs to match human-level performance on key ToM tasks for the first time. Ablation studies confirm the necessity of all components, which showcase the framework's ability to balance contextual plausibility, social appropriateness, and user adaptation. This work advances AI systems toward human-like social intelligence, with applications in empathetic dialogue and culturally sensitive interactions. Code is available at https://github.com/XMZhangAI/MetaMind.

Xuanming Zhang, Yuxuan Chen, Yiming Zheng et al.

In real world software development, improper or missing exception handling can severely impact the robustness and reliability of code. Exception handling mechanisms require developers to detect, capture, and manage exceptions according to high standards, but many developers struggle with these tasks, leading to fragile code. This problem is particularly evident in open-source projects and impacts the overall quality of the software ecosystem. To address this challenge, we explore the use of large language models (LLMs) to improve exception handling in code. Through extensive analysis, we identify three key issues: Insensitive Detection of Fragile Code, Inaccurate Capture of Exception Block, and Distorted Handling Solution. These problems are widespread across real world repositories, suggesting that robust exception handling practices are often overlooked or mishandled. In response, we propose Seeker, a multi-agent framework inspired by expert developer strategies for exception handling. Seeker uses agents: Scanner, Detector, Predator, Ranker, and Handler to assist LLMs in detecting, capturing, and resolving exceptions more effectively. Our work is the first systematic study on leveraging LLMs to enhance exception handling practices in real development scenarios, providing valuable insights for future improvements in code reliability.

Yang Xu, Xuanming Zhang, Samuel Yeh et al.

Deception is a pervasive feature of human communication and an emerging concern in large language models (LLMs). While recent studies document instances of LLM deception under pressure, most evaluations remain confined to single-turn prompts and fail to capture the long-horizon interactions in which deceptive strategies typically unfold. We introduce the first simulation framework for probing and evaluating deception in LLMs under extended sequences of interdependent tasks and dynamic contextual pressures. Our framework instantiates a multi-agent system: a performer agent tasked with completing tasks and a supervisor agent that evaluates progress, provides feedback, and maintains evolving states of trust. An independent deception auditor then reviews full trajectories to identify when and how deception occurs. We conduct extensive experiments across 11 frontier models, spanning both closed- and open-source systems, and find that deception is model-dependent, increases with event pressure, and consistently erodes supervisor trust. Qualitative analyses further reveal distinct strategies of concealment, equivocation, and falsification. Our findings establish deception as an emergent risk in long-horizon interactions and provide a foundation for evaluating future LLMs in real-world, trust-sensitive contexts.

Xuanming Zhang, Shwan Ashrafi, Aziza Mirsaidova et al.

We study the reasoning behavior of large language models (LLMs) under limited computation budgets. In such settings, producing useful partial solutions quickly is often more practical than exhaustive reasoning, which incurs high inference costs. Many real-world tasks, such as trip planning, require models to deliver the best possible output within a fixed reasoning budget. We introduce an anytime reasoning framework and the Anytime Index, a metric that quantifies how effectively solution quality improves as reasoning tokens increase. To further enhance efficiency, we propose an inference-time self-improvement method using LLM-synthesized preference data, where models learn from their own reasoning comparisons to produce better intermediate solutions. Experiments on NaturalPlan (Trip), AIME, and GPQA datasets show consistent gains across Grok-3, GPT-oss, GPT-4.1/4o, and LLaMA models, improving both reasoning quality and efficiency under budget constraints.

Jia Li, Ge Li, Xuanming Zhang et al. · pku

How to evaluate Large Language Models (LLMs) in code generation is an open question. Existing benchmarks demonstrate poor alignment with real-world code repositories and are insufficient to evaluate the coding abilities of LLMs. This paper proposes a new benchmark - EvoCodeBench to address the preceding problems, which has three primary advances. (1) EvoCodeBench aligns with real-world repositories in multiple dimensions, e.g., code distributions and dependency distributions. (2) EvoCodeBench offers comprehensive annotations (e.g., requirements, reference code, and reference dependencies), and robust evaluation metrics (e.g., Pass@k and Recall@k). (3) EvoCodeBench is an evolving benchmark to avoid data leakage. We build an automatic pipeline to update EvoCodeBench from the latest repositories. We release the first version - EvoCodeBench-2403, containing 275 samples from 25 real-world repositories. Based on EvoCodeBench, we propose repository-level code generation and evaluate 10 popular LLMs (e.g., gpt-4, gpt-3.5, DeepSeek Coder, StarCoder 2, CodeLLaMa, Gemma, and Qwen 1.5). Our experiments reveal the coding abilities of these LLMs in real-world repositories. For example, the highest Pass@1 of gpt-4 only is 20.73% in our experiments. We also analyze failed cases and summarize the shortcomings of existing LLMs in EvoCodeBench. We release EvoCodeBench, all prompts, and LLMs' completions for further community analysis.

Maximillian Chen, Xuanming Zhang, Michael Peng et al.

The rise of Internet of Things (IoT) devices in the physical world necessitates voice-based interfaces capable of handling complex user experiences. While modern Large Language Models (LLMs) already demonstrate strong tool-usage capabilities, modeling real-world IoT devices presents a difficult, understudied challenge which combines modeling spatiotemporal constraints with speech inputs, dynamic state tracking, and mixed-initiative interaction patterns. We introduce MIST (the Multimodal Interactive Speech-based Tool-calling Dataset), a synthetic multi-turn, voice-driven code generation task that operates over IoT devices. We find that there is a significant gap between open- and closed-weight multimodal LLMs on MIST, and that even frontier closed-weight LLMs have substantial headroom. We release MIST and an extensible data generation framework to build related datasets in order to facilitate research on mixed-initiative voice assistants which reason about physical world constraints.

Sahand Sabour, June M. Liu, Siyang Liu et al.

Artificial Intelligence (AI) systems are increasingly intertwined with daily life, assisting users in executing various tasks and providing guidance on decision-making. This integration introduces risks of AI-driven manipulation, where such systems may exploit users' cognitive biases and emotional vulnerabilities to steer them toward harmful outcomes. Through a randomized controlled trial with 233 participants, we examined human susceptibility to such manipulation in financial (e.g., purchases) and emotional (e.g., conflict resolution) decision-making contexts. Participants interacted with one of three AI agents: a neutral agent (NA) optimizing for user benefit without explicit influence, a manipulative agent (MA) designed to covertly influence beliefs and behaviors, or a strategy-enhanced manipulative agent (SEMA) employing explicit psychological tactics to reach its hidden objectives. By analyzing participants' decision patterns and shifts in their preference ratings post-interaction, we found significant susceptibility to AI-driven manipulation. Particularly, across both decision-making domains, participants interacting with the manipulative agents shifted toward harmful options at substantially higher rates (financial, MA: 62.3%, SEMA: 59.6%; emotional, MA: 42.3%, SEMA: 41.5%) compared to the NA group (financial, 35.8%; emotional, 12.8%). Notably, our findings reveal that even subtle manipulative objectives (MA) can be as effective as employing explicit psychological strategies (SEMA) in swaying human decision-making. By revealing the potential for covert AI influence, this study highlights a critical vulnerability in human-AI interactions, emphasizing the need for ethical safeguards and regulatory frameworks to ensure responsible deployment of AI technologies and protect human autonomy.

Jinfeng Zhou, Yuxuan Chen, Yihan Shi et al.

LLMs exhibit promising Social Intelligence (SI) in modeling human behavior, raising the need to evaluate LLMs' SI and their discrepancy with humans. SI equips humans with interpersonal abilities to behave wisely in navigating social interactions to achieve social goals. This presents an operational evaluation paradigm: outcome-oriented goal achievement evaluation and process-oriented interpersonal ability evaluation, which existing work fails to address. To this end, we propose SocialEval, a script-based bilingual SI benchmark, integrating outcome- and process-oriented evaluation by manually crafting narrative scripts. Each script is structured as a world tree that contains plot lines driven by interpersonal ability, providing a comprehensive view of how LLMs navigate social interactions. Experiments show that LLMs fall behind humans on both SI evaluations, exhibit prosociality, and prefer more positive social behaviors, even if they lead to goal failure. Analysis of LLMs' formed representation space and neuronal activations reveals that LLMs have developed ability-specific functional partitions akin to the human brain.

Xuanming Zhang, Yuxuan Chen, Samuel Yeh et al.

Large language models (LLMs) excel at complex reasoning but can still exhibit harmful behaviors. Current alignment strategies typically embed safety into model weights, making these controls implicit, static, and difficult to modify. This paper introduces Cognition-of-Thought (CooT), a novel decoding-time framework that equips LLMs with an explicit cognitive self-monitoring loop. CooT couples a standard text Generator with a cognitive Perceiver that continuously monitors the unfolding sequence. The Perceiver uses a structured, precedence-based hierarchy of principles (e.g., safety over obedience) to detect potential misalignments as they arise. When violations are flagged, CooT intervenes by rolling back the generation to the point of error and regenerating under injected guidance that combines universal social priors with context-specific warnings. CooT thus transforms alignment from a fixed property into an explicit, dynamic, and auditable process active during inference, allowing for flexible policy updates without retraining the model. Extensive experiments across multiple benchmarks and model families confirm that CooT consistently improves safety and social reasoning performance.

Lecheng Wang, Xianjie Shi, Ge Li et al. · pku

Auto-regressive language models (LMs) have been widely used to generate data in data-scarce domains to train new LMs, compensating for the scarcity of real-world data. Previous work experimentally found that LMs collapse when trained on recursively generated data. This paper presents a theoretical proof: once a corpus (such as a subset of the World Wide Web) begins to incorporate generated data and no new real-world data is added to the corpus, then no matter how small the amount of data each LM generates and contributes to the corpus, LM collapse is inevitable after sufficient time. This finding suggests that attempts to mitigate collapse by limiting the quantity of synthetic data in the corpus are fundamentally insufficient. Instead, avoiding collapse hinges on ensuring the quality of synthetic data.

Xuanming Zhang

Large Language Models (LLMs) exhibit a troubling duality, capable of both remarkable generalization and brittle, verbatim memorization of their training data. This unpredictability undermines their reliability in high-stakes applications. In this work, we propose a unified framework to understand, identify, and control these distinct reasoning modes. First, we introduce a theoretical model based on the Information Bottleneck (IB) principle, formalizing generalization as the learning of a compressed, task-relevant representation and memorization as a failure to compress. Building on this theory, we develop Dynamic Mode Steering (DMS), a novel inference-time algorithm which comprises two components: (1) a lightweight, causally-grounded linear probe that identifies the model's instantaneous reliance on memorization, and (2) a dynamic activation steering mechanism that nudges the model's computation towards pre-identified generalization circuits. We frame DMS as a form of adaptive, self-contrastive decoding. Experiments on reasoning and faithfulness tasks demonstrate that DMS significantly improves logical consistency and factual accuracy, thereby offering a principled approach to enhancing LLM reliability.

Li Siyan, Zhen Xu, Vethavikashini Chithrra Raghuram et al.

Asynchronous learning environments (ALEs) are widely adopted for formal and informal learning, but timely and personalized support is often limited. In this context, Virtual Teaching Assistants (VTAs) can potentially reduce the workload of instructors, but rigorous and pedagogically sound evaluation is essential. Existing assessments often rely on surface-level metrics and lack sufficient grounding in educational theories, making it difficult to meaningfully compare the pedagogical effectiveness of different VTA systems. To bridge this gap, we propose an evaluation framework rooted in learning sciences and tailored to asynchronous forum discussions, a common VTA deployment context in ALE. We construct classifiers using expert annotations of VTA responses on a diverse set of forum posts. We evaluate the effectiveness of our classifiers, identifying approaches that improve accuracy as well as challenges that hinder generalization. Our work establishes a foundation for theory-driven evaluation of VTA systems, paving the way for more pedagogically effective AI in education.

Tao Long, Xuanming Zhang, Sitong Wang et al.

Agentic workflows promise efficiency, but adoption hinges on whether people actually trust systems that act on their behalf. We present DoubleAgents, an agentic planning tool that embeds transparency and control through user intervention, value-reflecting policies, rich state visualizations, and uncertainty flagging for human coordination tasks. A built-in respondent simulation generates realistic scenarios, allowing users to rehearse, refine policies, and calibrate their reliance before live use. We evaluate DoubleAgents in a two-day lab study (n=10), two deployments (n=2), and a technical evaluation. Results show that participants initially hesitated to delegate but grew more reliant as they experienced transparency, control, and adaptive learning during simulated cases. Deployment results demonstrate DoubleAgents' real-world relevance and usefulness, showing that the effort required scaled appropriately with task complexity and contextual data. We contribute trust-by-design patterns and mechanisms for proactive AI -- consistency, controllability, and explainability -- along with simulation as a safe path to build and calibrate trust over time.

Xuanming Zhang

Industrial carbon emissions are a major driver of climate change, yet modeling these emissions is challenging due to multicollinearity among factors and complex interdependencies across sectors and time. We propose a novel graph-based deep learning framework DGL to analyze and forecast industrial CO_2 emissions, addressing high feature correlation and capturing industrial-temporal interdependencies. Unlike traditional regression or clustering methods, our approach leverages a Graph Neural Network (GNN) with attention mechanisms to model relationships between industries (or regions) and a temporal transformer to learn long-range patterns. We evaluate our framework on public global industry emissions dataset derived from EDGAR v8.0, spanning multiple countries and sectors. The proposed model achieves superior predictive performance - reducing error by over 15% compared to baseline deep models - while maintaining interpretability via attention weights and causal analysis. We believe that we are the first Graph-Temporal architecture that resolves multicollinearity by structurally encoding feature relationships, along with integration of causal inference to identify true drivers of emissions, improving transparency and fairness. We also stand a demonstration of policy relevance, showing how model insights can guide sector-specific decarbonization strategies aligned with sustainable development goals. Based on the above, we show high-emission "hotspots" and suggest equitable intervention plans, illustrating the potential of state-of-the-art AI graph learning to advance climate action, offering a powerful tool for policymakers and industry stakeholders to achieve carbon reduction targets.

Xuanming Zhang, Anthony Diaz, Zixun Chen et al.

Coherence in writing, an aspect that second-language (L2) English learners often struggle with, is crucial in assessing L2 English writing. Existing automated writing evaluation systems primarily use basic surface linguistic features to detect coherence in writing. However, little effort has been made to correct the detected incoherence, which could significantly benefit L2 language learners seeking to improve their writing. To bridge this gap, we introduce DECOR, a novel benchmark that includes expert annotations for detecting incoherence in L2 English writing, identifying the underlying reasons, and rewriting the incoherent sentences. To our knowledge, DECOR is the first coherence assessment dataset specifically designed for improving L2 English writing, featuring pairs of original incoherent sentences alongside their expert-rewritten counterparts. Additionally, we fine-tuned models to automatically detect and rewrite incoherence in student essays. We find that incorporating specific reasons for incoherence during fine-tuning consistently improves the quality of the rewrites, achieving a result that is favored in both automatic and human evaluations.

Kun Qian, Shunji Wan, Claudia Tang et al.

As large language models achieve impressive scores on traditional benchmarks, an increasing number of researchers are becoming concerned about benchmark data leakage during pre-training, commonly known as the data contamination problem. To ensure fair evaluation, recent benchmarks release only the training and validation sets, keeping the test set labels closed-source. They require anyone wishing to evaluate his language model to submit the model's predictions for centralized processing and then publish the model's result on their leaderboard. However, this submission process is inefficient and prevents effective error analysis. To address this issue, we propose to variabilize benchmarks and evaluate language models dynamically. Specifically, we extract variables from each test case and define a value range for each variable. For each evaluation, we sample new values from these value ranges to create unique test cases, thus ensuring a fresh evaluation each time. We applied this variable perturbation method to four datasets: GSM8K, ARC, CommonsenseQA, and TruthfulQA, which cover mathematical generation and multiple-choice tasks. Our experimental results demonstrate that this approach provides a more accurate assessment of the true capabilities of language models, effectively mitigating the contamination problem.

Xuanming Zhang, Zixun Chen, Zhou Yu

Lexical Substitution discovers appropriate substitutes for a given target word in a context sentence. However, the task fails to consider substitutes that are of equal or higher proficiency than the target, an aspect that could be beneficial for language learners looking to improve their writing. To bridge this gap, we propose a new task, language proficiency-oriented lexical substitution. We also introduce ProLex, a novel benchmark designed to assess systems' ability to generate not only appropriate substitutes but also substitutes that demonstrate better language proficiency. Besides the benchmark, we propose models that can automatically perform the new task. We show that our best model, a Llama2-13B model fine-tuned with task-specific synthetic data, outperforms ChatGPT by an average of 3.2% in F-score and achieves comparable results with GPT-4 on ProLex.

Xuanming Zhang

This study proposes a novel analytical framework that integrates DBSCAN clustering with the Elastic Net regression model to address multifactorial problems characterized by structural complexity and multicollinearity, exemplified by carbon emissions analysis. DBSCAN is employed for unsupervised learning to objectively cluster features, while the Elastic Net is utilized for high-dimensional feature selection and complexity control. The Elastic Net is specifically chosen for its ability to balance feature selection and regularization by combining L1 (lasso) and L2 (ridge) penalties, making it particularly suited for datasets with correlated predictors. Applying this framework to energy consumption data from 46 industries in China (2000-2019) resulted in the identification of 16 categories. Emission characteristics and drivers were quantitatively assessed for each category, demonstrating the framework's capacity to identify primary emission sources and provide actionable insights. This research underscores the global applicability of the framework for analyzing complex regional challenges, such as carbon emissions, and highlights its potential to identify opportunities for emission reduction.

Yihan Pan, Zhenghang Xu, Jin Guang et al.

We collaborate with a large teaching hospital in Shenzhen, China and build a high-fidelity simulation model for its ultrasound center to predict key performance metrics, including the distributions of queue length, waiting time and sojourn time, with high accuracy. The key challenge to build an accurate simulation model is to understanding the complicated patient routing at the ultrasound center. To address the issue, we propose a novel two-level routing component to the queueing network model. We apply machine learning tools to calibrate the key components of the queueing model from data with enhanced accuracy.