Publications

2026

NeurIPS 2026
Progress-Aware Distillation for Mitigating Stagnation in Small Language Model Agents
Yuanpu Cao, Saket Sathe, Hanyu Wang, Ziyi Yin, Fenglong Ma, Jinghui Chen
Advances in Neural Information Processing Systems (NeurIPS 2026)
We propose Progress-Aware Distillation (ProD), which detects and penalizes unproductive loops in student agents during distillation from larger models, helping small language model agents keep making progress on complex tasks.
Preprint
DecisionRobust: Benchmarking Robustness for JEV Style Decision Models Across Text and Vision
Yifan Lan, Zhongye Liu, Yurui Chang, Hanyu Wang, Yongkang Du, Lu Lin, Jinghui Chen
We introduce DecisionRobust, a benchmark for evaluating the adversarial robustness of decision-making models across both textual and visual inputs.
Preprint
ARGAdaptive Reference Guidance
Hanyu Wang, Nakul Agarwal, Hossein Nourkhiz Mahjoub, Ehsan Moradi Pari, Makoto Fukushima, Jinghui Chen, Vaishnav Tadiparthi
We study how much reference-prefix guidance to use during rollouts in RLVR, derive theoretical bounds on the induced KL divergence, and propose Adaptive Reference Guidance, which improves mathematical reasoning performance.
NeurIPS 2026 Workshop
Skill Transfer: Adapting Agent Skills from Strong to Weak Agents
Yifan Lan, Hanyu Wang, Lu Lin, Jinghui Chen
NeurIPS 2026 Workshop on Towards Test-Time Continual Learning Agents (TTCL)
We adapt agent skills authored for strong models so that weaker agents can execute them well, by abstracting strong and weak trajectories into typed execution structures, localizing their divergence with graph edit distance, and iteratively patching the skill.
Preprint
SkillGrad: Optimizing Agent Skills Like Gradient Descent
Hanyu Wang, Yifan Lan, Bochuan Cao, Lu Lin, Jinghui Chen
We optimize agent skills in a gradient-descent-like loop, iteratively refining skill documents from execution feedback.
Preprint
The Illusion of Reasoning: Exposing Evasive Data Contamination in LLMs via Zero-CoT Truncation
Yifan Lan, Yuanpu Cao, Hanyu Wang, Lu Lin, Jinghui Chen
We show that chain-of-thought reasoning can mask memorization, and propose the Zero-CoT Probe (ZCP), a black-box method that truncates the reasoning process and compares performance against an isomorphically perturbed reference set to detect both direct and evasive data contamination.
Preprint
PreFlect: From Retrospective to Prospective Reflection in Large Language Model Agents
Hanyu Wang, Yuanpu Cao, Lu Lin, Jinghui Chen
We develop a framework that shifts reflection from reactive, post-hoc correction to proactive, pre-execution foresight, anchoring a planning-phase reflection loop with a distilled taxonomy of historical errors.

2025

ICML 2025
TruthFlow: Truthful LLM Generation via Representation Flow CorrectionTruthFlow: Truthful LLM Generation via Representation Flow Correction
Hanyu Wang, Bochuan Cao, Yuanpu Cao, Jinghui Chen
Proceedings of the 42nd International Conference on Machine Learning (ICML 2025)
We propose TruthFlow, a representation intervention framework that mitigates LLM hallucination by generating query-specific correction vectors via flow matching, enhanced by a truth-related subspace projection using SVD.

2024

ICML 2024 Workshop
DiffusionMemorization vs. Generation
Hanyu Wang, Yujin Han, Difan Zou
ICML 2024 Workshop on Foundation Models in the Wild
We address the paradox in diffusion models where the theoretical optimal score suggests memorization while trained models generalize, attributing generalization to the superior geometric smoothness of the trained score function.