Searcharxiv⌕ Search

arXiv subjects

Shuchang Tao

Publications and source records attributed to Shuchang Tao.

At least 19 recordsLinked to original sources
arXiv2026

RLCSD: Reinforcement Learning with Contrastive On-Policy Self-Distillation

On-policy self-distillation (OPSD) provides dense, token-level supervision for reasoning models by aligning a model's own distribution with that under privileged context, typically a verified solution. However, we show that the resulting distributional gap concentrates on style tokens rather than task-bearing ones, as the hinted model tends to produce shorter, more direct outputs. We term this pathology \emph{privilege-induced style drift}, which can destabilize training and shorten responses. To address this, we propose \textbf{RLCSD} (Reinforcement Learning with Contrastive on-policy Self-Distillation), which mitigates this drift by contrasting the teacher-student gap under a correct hint against that under a wrong hint, suppressing style shifts induced by hints regardless of correctness and yielding a signal more concentrated on task-bearing tokens. Experiments on Qwen3 (1.7B/4B/8B) and Olmo-3-7B-Think across mathematical and logical reasoning show that RLCSD consistently outperforms GRPO and prior OPSD methods, with additional results on agentic tasks supporting broader applicability. We further show that the contrastive principle is general: it plugs into existing OPSD methods to improve them, and its underlying insight extends to broader cross-model on-policy distillation.

cs.LG↗
arXiv2026

Omni-SafetyBench: A Benchmark for Safety Evaluation of Audio-Visual Large Language Models

Omni-modal Large Language Models (OLLMs) that integrate visual, auditory, and textual processing face severe safety risks. They exhibit fragile defenses against audio-visual joint harmful inputs and demonstrate inconsistent safety performance across different modalities, enabling simple modality-switching jailbreaks. However, existing safety benchmarks fail to comprehensively assess these risks due to the absence of audio-visual joint samples, limited modality coverage, and lack of parallel test cases for cross-modal consistency evaluation. To address these gaps, we introduce Omni-SafetyBench, the first comprehensive parallel benchmark for OLLM safety evaluation, featuring 23,328 test instances across 24 modality variations derived from 972 seed samples. Recognizing that complex inputs pose comprehension challenges and that cross-modal consistency is critical for OLLM safety, we propose tailored metrics: a Safety-score based on Conditional Attack Success Rate (C-ASR) and Conditional Refusal Rate (C-RR), and a Cross-Modal Safety Consistency score (CMSC-score). Evaluating 11 state-of-the-art OLLMs reveals severe vulnerabilities: only 3 models exceed 0.6 in both metrics, with safety degrading sharply for audio-visual inputs. Furthermore, evaluation of existing safety alignment methods on Omni-SafetyBench identifies fundamental challenges in OLLM safety alignment, highlighting urgent needs for enhanced research in this domain.

cs.CL↗
arXiv2026

AgentJet: A Distributed Swarm Training Framework for Agentic Reinforcement Learning

Training reinforcement learning (RL) policies for large language model (LLM) agents requires optimizing multi-turn trajectories that interact with external environments. Existing training frameworks struggle with runtime failures, single-model constraints, incompatible task environments, and redundant context. We present AgentJet, a distributed swarm training framework based on a decoupled multi-node architecture. AgentJet treats the server--client topology as configurable: swarm servers host trainable models and perform optimization on GPU clusters, while detachable swarm clients execute arbitrary agents and communicate through OpenAI-compatible APIs. Reconfiguring this topology supports heterogeneous multi-model RL, mixed-task training with isolated runtimes, fault-tolerant execution, and live code iteration through hot-swappable clients. AgentJet also introduces context tracking with timeline merging, reducing actor-update time by 6.25x on AppWorld. The same detachable-client design supports an automated research system that conducts long-horizon, multi-day RL studies on large-scale clusters with limited human intervention. AgentJet is open-source and compatible with agent systems that issue standard LLM inference requests.

cs.AI↗
arXiv2026

d-TreeRPO: Towards More Reliable Policy Optimization for Diffusion Language Models

Reinforcement learning (RL) is pivotal for enhancing the reasoning capabilities of diffusion large language models (dLLMs). However, existing dLLM policy optimization methods suffer from two critical reliability bottlenecks: (1) reward sparsity, arising from coarse or unverifiable signals that impede accurate advantage calculation; and (2) their probability estimates do not account for the gap to the unbiased expectation over all decoding orders, which are intractable to compute. To mitigate these issues, we propose d-TreeRPO, a reliable RL framework for dLLMs that leverages tree-structured rollouts and bottom-up advantage computation based on verifiable outcome rewards to provide fine-grained and verifiable step-wise reward signals. Furthermore, we provide a theoretical proof demonstrating that increasing prediction confidence effectively minimizes the gap between unbiased expected prediction probabilities and its single-step forward pass estimate. Guided by this analysis, we introduce a time-scheduled self-distillation loss during training that enhances prediction confidence in later training stages, thereby enabling more accurate probability estimation and better performance. Experiments demonstrate that d-TreeRPO outperforms existing baselines and achieves significant improvements across multiple reasoning benchmarks. Specifically, it achieves +86.2% on Sudoku, +51.6% on Countdown, +4.5% on GSM8K, and +5.3% on Math500 compared to the base model.

cs.CL↗
arXiv2026

Incentivizing Strong Reasoning from Weak Supervision

Large language models (LLMs) have demonstrated impressive performance on reasoning-intensive tasks, but enhancing their reasoning abilities typically relies on either reinforcement learning (RL) with verifiable signals or supervised fine-tuning (SFT) with high-quality long chain-of-thought (CoT) demonstrations, both of which are expensive. In this paper, we study a novel problem of incentivizing the reasoning capacity of LLMs without expensive high-quality demonstrations and reinforcement learning. We investigate whether the reasoning capabilities of LLMs can be effectively incentivized via supervision from significantly weaker models. We further analyze when and why such weak supervision succeeds in eliciting reasoning abilities in stronger models. Our findings show that supervision from significantly weaker reasoners can substantially improve student reasoning performance, recovering close to 94% of the gains of expensive RL at a fraction of the cost. Experiments across diverse benchmarks and model architectures demonstrate that weak reasoners can effectively incentivize reasoning in stronger student models, consistently improving performance across a wide range of reasoning tasks. Our results suggest that this simple weak-to-strong paradigm is a promising and generalizable alternative to costly methods for incentivizing strong reasoning capabilities at inference-time in LLMs. The code is publicly available at https://github.com/yuanyige/w2sr.

cs.CL↗
arXiv2026

Inference-time Alignment in Continuous Space

Aligning large language models with human feedback at inference time has received increasing attention due to its flexibility. Existing methods rely on generating multiple responses from the base policy for search using a reward model, which can be considered as searching in a discrete response space. However, these methods struggle to explore informative candidates when the base policy is weak or the candidate set is small, resulting in limited effectiveness. In this paper, to address this problem, we propose Simple Energy Adaptation ($\textbf{SEA}$), a simple yet effective algorithm for inference-time alignment. In contrast to expensive search over the discrete space, SEA directly adapts original responses from the base policy toward the optimal one via gradient-based sampling in continuous latent space. Specifically, SEA formulates inference as an iterative optimization procedure on an energy function over actions in the continuous space defined by the optimal policy, enabling simple and effective alignment. For instance, despite its simplicity, SEA outperforms the second-best baseline with a relative improvement of up to $ \textbf{77.51%}$ on AdvBench and $\textbf{16.36%}$ on MATH. Our code is publicly available at https://github.com/yuanyige/sea

cs.CL↗
arXiv2025

CuES: A Curiosity-driven and Environment-grounded Synthesis Framework for Agentic RL

Large language model based agents are increasingly deployed in complex, tool augmented environments. While reinforcement learning provides a principled mechanism for such agents to improve through interaction, its effectiveness critically depends on the availability of structured training tasks. In many realistic settings, however, no such tasks exist a challenge we term task scarcity, which has become a key bottleneck for scaling agentic RL. Existing approaches typically assume predefined task collections, an assumption that fails in novel environments where tool semantics and affordances are initially unknown. To address this limitation, we formalize the problem of Task Generation for Agentic RL, where an agent must learn within a given environment that lacks predefined tasks. We propose CuES, a Curiosity driven and Environment grounded Synthesis framework that autonomously generates diverse, executable, and meaningful tasks directly from the environment structure and affordances, without relying on handcrafted seeds or external corpora. CuES drives exploration through intrinsic curiosity, abstracts interaction patterns into reusable task schemas, and refines them through lightweight top down guidance and memory based quality control. Across three representative environments, AppWorld, BFCL, and WebShop, CuES produces task distributions that match or surpass manually curated datasets in both diversity and executability, yielding substantial downstream policy improvements. These results demonstrate that curiosity driven, environment grounded task generation provides a scalable foundation for agents that not only learn how to act, but also learn what to learn. The code is available at https://github.com/modelscope/AgentEvolver/tree/main/research/CuES.

cs.AI↗
arXiv2025

AgentEvolver: Towards Efficient Self-Evolving Agent System

Autonomous agents powered by large language models (LLMs) have the potential to significantly enhance human productivity by reasoning, using tools, and executing complex tasks in diverse environments. However, current approaches to developing such agents remain costly and inefficient, as they typically require manually constructed task datasets and reinforcement learning (RL) pipelines with extensive random exploration. These limitations lead to prohibitively high data-construction costs, low exploration efficiency, and poor sample utilization. To address these challenges, we present AgentEvolver, a self-evolving agent system that leverages the semantic understanding and reasoning capabilities of LLMs to drive autonomous agent learning. AgentEvolver introduces three synergistic mechanisms: (i) self-questioning, which enables curiosity-driven task generation in novel environments, reducing dependence on handcrafted datasets; (ii) self-navigating, which improves exploration efficiency through experience reuse and hybrid policy guidance; and (iii) self-attributing, which enhances sample efficiency by assigning differentiated rewards to trajectory states and actions based on their contribution. By integrating these mechanisms into a unified framework, AgentEvolver enables scalable, cost-effective, and continual improvement of agent capabilities. Preliminary experiments indicate that AgentEvolver achieves more efficient exploration, better sample utilization, and faster adaptation compared to traditional RL-based baselines.

cs.LG↗
arXiv2025

On the Diminishing Returns of Complex Robust RAG Training in the Era of Powerful LLMs

Retrieval-augmented generation (RAG) systems traditionally employ sophisticated training strategies to enhance robustness against retrieval noise. In this work, we investigate a critical question: does the benefit of these complex robust training methods diminish as language models become more powerful? Through systematic evaluation across multiple model scales and question-answering datasets, our analysis reveals a consistent trend: \emph{the marginal robustness benefit of sophisticated training strategies decreases substantially as model capacity increases.} While smaller models show significant performance improvements from complex document selection and adversarial objectives, more capable models achieve comparable or even superior performance with simpler training approaches. Further investigation demonstrates that stronger models naturally exhibit better confidence calibration, cross-dataset generalization capability, and more effective attention patterns, even under simple training regimes. These findings suggest that as foundation models evolve, the engineering effort invested in complex robust training may yield diminishing returns, indicating that simplified RAG pipelines could suffice for powerful models while maintaining competitive performance.

cs.CL↗
arXiv2025

ToolCoder: A Systematic Code-Empowered Tool Learning Framework for Large Language Models

Tool learning has emerged as a crucial capability for large language models (LLMs) to solve complex real-world tasks through interaction with external tools. Existing approaches face significant challenges, including reliance on hand-crafted prompts, difficulty in multi-step planning, and lack of precise error diagnosis and reflection mechanisms. We propose ToolCoder, a novel framework that reformulates tool learning as a code generation task. Inspired by software engineering principles, ToolCoder transforms natural language queries into structured Python function scaffold and systematically breaks down tasks with descriptive comments, enabling LLMs to leverage coding paradigms for complex reasoning and planning. It then generates and executes function implementations to obtain final responses. Additionally, ToolCoder stores successfully executed functions in a repository to promote code reuse, while leveraging error traceback mechanisms for systematic debugging, optimizing both execution efficiency and robustness. Experiments demonstrate that ToolCoder achieves superior performance in task completion accuracy and execution reliability compared to existing approaches, establishing the effectiveness of code-centric approaches in tool learning.

cs.CL↗
arXiv2025

Enhancing Tool Learning in Large Language Models with Hierarchical Error Checklists

Large language models (LLMs) have significantly advanced natural language processing, particularly through the integration of external tools and APIs. However, their effectiveness is frequently hampered by parameter mis-filling during tool calling. In this paper, we propose the Hierarchical Tool Error Checklist (HiTEC) framework to systematically diagnose and mitigate tool-calling errors without relying on extensive real-world interactions. HiTEC introduces a two-tiered approach: a global error checklist that identifies common, cross-tool issues, and a local error checklist that targets tool-specific and contextual failures. Building on this structure, we propose two deployments: HiTEC-In Context Learning (HiTEC-ICL) and HiTEC-Kahneman-Tversky Optimization (HiTEC-KTO). HiTEC-ICL embeds the global checklist in the initial prompts and leverages a two-round conversational interaction to dynamically refine parameter handling, while HiTEC-KTO generates high-quality negative examples to drive fine-tuning via preference-based optimization. Extensive experiments across five public datasets demonstrate that our framework significantly improves parameter-filling accuracy and tool-calling success rates compared to baseline methods.

cs.CL↗
arXiv2025

Robust Recommender System: A Survey and Future Directions

With the rapid growth of information, recommender systems have become integral for providing personalized suggestions and overcoming information overload. However, their practical deployment often encounters ``dirty'' data, where noise or malicious information can lead to abnormal recommendations. Research on improving recommender systems' robustness against such dirty data has thus gained significant attention. This survey provides a comprehensive review of recent work on recommender systems' robustness. We first present a taxonomy to organize current techniques for withstanding malicious attacks and natural noise. We then explore state-of-the-art methods in each category, including fraudster detection, adversarial training, certifiable robust training for defending against malicious attacks, and regularization, purification, self-supervised learning for defending against malicious attacks. Additionally, we summarize evaluation metrics and commonly used datasets for assessing robustness. We discuss robustness across varying recommendation scenarios and its interplay with other properties like accuracy, interpretability, privacy, and fairness. Finally, we delve into open issues and future research directions in this emerging field. Our goal is to provide readers with a comprehensive understanding of robust recommender systems and to identify key pathways for future research and development. To facilitate ongoing exploration, we maintain a continuously updated GitHub repository with related research: https://github.com/Kaike-Zhang/Robust-Recommender-System.

cs.IR↗
arXiv2024

When to Trust LLMs: Aligning Confidence with Response Quality

Despite the success of large language models (LLMs) in natural language generation, much evidence shows that LLMs may produce incorrect or nonsensical text. This limitation highlights the importance of discerning when to trust LLMs, especially in safety-critical domains. Existing methods often express reliability by confidence level, however, their effectiveness is limited by the lack of objective guidance. To address this, we propose CONfidence-Quality-ORDer-preserving alignment approach (CONQORD), which leverages reinforcement learning guided by a tailored dual-component reward function. This function integrates quality reward and order-preserving alignment reward functions. Specifically, the order-preserving reward incentivizes the model to verbalize greater confidence for responses of higher quality to align the order of confidence and quality. Experiments demonstrate that CONQORD significantly improves the alignment performance between confidence and response accuracy, without causing over-cautious. Furthermore, the aligned confidence provided by CONQORD informs when to trust LLMs, and acts as a determinant for initiating the retrieval process of external knowledge. Aligning confidence with response quality ensures more transparent and reliable responses, providing better trustworthiness.

cs.CL↗
arXiv2024

Accelerating the Surrogate Retraining for Poisoning Attacks against Recommender Systems

Recent studies have demonstrated the vulnerability of recommender systems to data poisoning attacks, where adversaries inject carefully crafted fake user interactions into the training data of recommenders to promote target items. Current attack methods involve iteratively retraining a surrogate recommender on the poisoned data with the latest fake users to optimize the attack. However, this repetitive retraining is highly time-consuming, hindering the efficient assessment and optimization of fake users. To mitigate this computational bottleneck and develop a more effective attack in an affordable time, we analyze the retraining process and find that a change in the representation of one user/item will cause a cascading effect through the user-item interaction graph. Under theoretical guidance, we introduce \emph{Gradient Passing} (GP), a novel technique that explicitly passes gradients between interacted user-item pairs during backpropagation, thereby approximating the cascading effect and accelerating retraining. With just a single update, GP can achieve effects comparable to multiple original training iterations. Under the same number of retraining epochs, GP enables a closer approximation of the surrogate recommender to the victim. This more accurate approximation provides better guidance for optimizing fake users, ultimately leading to enhanced data poisoning attacks. Extensive experiments on real-world datasets demonstrate the efficiency and effectiveness of our proposed GP.

cs.IR↗
arXiv2024

IDEA: Invariant Defense for Graph Adversarial Robustness

Despite the success of graph neural networks (GNNs), their vulnerability to adversarial attacks poses tremendous challenges for practical applications. Existing defense methods suffer from severe performance decline under unseen attacks, due to either limited observed adversarial examples or pre-defined heuristics. To address these limitations, we analyze the causalities in graph adversarial attacks and conclude that causal features are key to achieve graph adversarial robustness, owing to their determinedness for labels and invariance across attacks. To learn these causal features, we innovatively propose an Invariant causal DEfense method against adversarial Attacks (IDEA). We derive node-based and structure-based invariance objectives from an information-theoretic perspective. IDEA ensures strong predictability for labels and invariant predictability across attacks, which is provably a causally invariant defense across various attacks. Extensive experiments demonstrate that IDEA attains state-of-the-art defense performance under all five attacks on all five datasets. The implementation of IDEA is available at https://anonymous.4open.science/r/IDEA.

cs.LG↗
arXiv2023

Adversarial Camouflage for Node Injection Attack on Graphs

Node injection attacks on Graph Neural Networks (GNNs) have received increasing attention recently, due to their ability to degrade GNN performance with high attack success rates. However, our study indicates that these attacks often fail in practical scenarios, since defense/detection methods can easily identify and remove the injected nodes. To address this, we devote to camouflage node injection attack, making injected nodes appear normal and imperceptible to defense/detection methods. Unfortunately, the non-Euclidean structure of graph data and the lack of intuitive prior present great challenges to the formalization, implementation, and evaluation of camouflage. In this paper, we first propose and define camouflage as distribution similarity between ego networks of injected nodes and normal nodes. Then for implementation, we propose an adversarial CAmouflage framework for Node injection Attack, namely CANA, to improve attack performance under defense/detection methods in practical scenarios. A novel camouflage metric is further designed under the guide of distribution similarity. Extensive experiments demonstrate that CANA can significantly improve the attack performance under defense/detection methods with higher camouflage or imperceptibility. This work urges us to raise awareness of the security vulnerabilities of GNNs in practical applications.

cs.LG↗
arXiv2023

Graph Adversarial Immunization for Certifiable Robustness

Despite achieving great success, graph neural networks (GNNs) are vulnerable to adversarial attacks. Existing defenses focus on developing adversarial training or model modification. In this paper, we propose and formulate graph adversarial immunization, i.e., vaccinating part of graph structure to improve certifiable robustness of graph against any admissible adversarial attack. We first propose edge-level immunization to vaccinate node pairs. Unfortunately, such edge-level immunization cannot defend against emerging node injection attacks, since it only immunizes existing node pairs. To this end, we further propose node-level immunization. To avoid computationally intensive combinatorial optimization associated with adversarial immunization, we develop AdvImmune-Edge and AdvImmune-Node algorithms to effectively obtain the immune node pairs or nodes. Extensive experiments demonstrate the superiority of AdvImmune methods. In particular, AdvImmune-Node remarkably improves the ratio of robust nodes by 79%, 294%, and 100%, after immunizing only 5% of nodes. Furthermore, AdvImmune methods show excellent defensive performance against various attacks, outperforming state-of-the-art defenses. To the best of our knowledge, this is the first attempt to improve certifiable robustness from graph data perspective without losing performance on clean graphs, providing new insights into graph adversarial learning.

cs.LG↗
arXiv2023

Popularity Debiasing from Exposure to Interaction in Collaborative Filtering

Recommender systems often suffer from popularity bias, where popular items are overly recommended while sacrificing unpopular items. Existing researches generally focus on ensuring the number of recommendations exposure of each item is equal or proportional, using inverse propensity weighting, causal intervention, or adversarial training. However, increasing the exposure of unpopular items may not bring more clicks or interactions, resulting in skewed benefits and failing in achieving real reasonable popularity debiasing. In this paper, we propose a new criterion for popularity debiasing, i.e., in an unbiased recommender system, both popular and unpopular items should receive Interactions Proportional to the number of users who Like it, namely IPL criterion. Under the guidance of the criterion, we then propose a debiasing framework with IPL regularization term which is theoretically shown to achieve a win-win situation of both popularity debiasing and recommendation performance. Experiments conducted on four public datasets demonstrate that when equipping two representative collaborative filtering models with our framework, the popularity bias is effectively alleviated while maintaining the recommendation performance.

cs.IR↗