My research focuses on mid- and post-training for coding agents and the statistical foundations of large language models, including reinforcement learning, reward modeling, and evaluation from noisy feedback.

• Just Reuse the Rollout Log-Probs: Stable Mixture-of-Experts Reinforcement Learning with a Vocabulary-KL Clip
  Zihan Dong, Shuo Yang, Chiyu Ma, Haozhe Li, Yang Zhou, Zihang Meng, Hao Zhou, Linjun Zhang, and Guoyin Wang.

  Submitted to ICLR 2027.

  VoKL stabilizes mixture-of-experts reinforcement learning by reusing rollout log-probabilities and applying a vocabulary-KL trust region to address the training–inference mismatch.

• Evaluating LLMs When They Do Not Know the Answer: Statistical Evaluation of Mathematical Reasoning via Comparative Signals
  Zihan Dong, Zhixian Zhang, Yang Zhou, Can Jin, Ruijia Wu, Linjun Zhang.

  ICML 2026.

  Combines labeled outcomes with pairwise reasoning comparisons in a semiparametric estimator to improve the precision and reliability of LLM evaluation.

  Paper / Code

• Alternating Reinforcement Learning for Rubric-Based Reward Modeling in Non-Verifiable LLM Post-Training
  Ran Xu, Tianci Liu, Zihan Dong, Tony Yu, Ilgee Hong, Carl Yang, Linjun Zhang, Tuo Zhao, Haoyu Wang.

  ICML 2026.

  Rubric-ARM alternates reinforcement learning updates between a rubric generator and a judge to improve reward modeling for open-ended tasks.

  Paper / Code

• AutoResearchClaw: Self-Reinforcing Autonomous Research with Human-AI Collaboration
  Jiaqi Liu, Shi Qiu, Mairui Li, …, Zihan Dong, et al.

  Submitted to ICLR 2027.

  An iterative autonomous research system that combines multi-agent debate, recovery from failed experiments, cross-run learning, and targeted human collaboration.

  Paper / Code

• Labels or Preferences? Budget-Constrained Learning with Human Judgments over AI-Generated Outputs
  Zihan Dong, Xiaotian Hou, Ruijia Wu, and Linjun Zhang.

  Preprint, 2026.

  A semiparametric active learning framework that allocates a fixed annotation budget between labels and preferences while targeting statistically efficient estimation.

  Paper

• Contrastive Network Representation Learning
  Zihan Dong, Xin Zhou, Ryumei Nakada, Lexin Li, and Linjun Zhang.

  Preprint, 2025.

  Learns network edge representations through contrastive learning and adaptive masking, with minimax-optimal guarantees and applications to brain connectivity analysis.

  Paper

• RIMS: Preference Optimization via Smoothed Multi-pair Aggregation for Small-Scale LLM Retrieval-Augmented Generation
  Pei Tian, Zihan Dong, Tianci Liu, Linjun Zhang, and Haoyu Wang.

  COLM 2026.

  RIMS improves small-model retrieval-augmented generation by generating its own preference data and smoothly aggregating multiple preference pairs during optimization.

  Paper / Code

• Personalization as Inverse Planning: Learning Latent Design Intents for Agentic Slide Generation via Structural Denoising
  Tianci Liu, Zihan Dong, Linjun Zhang, Haoyu Wang, Jing Gao, Emre Kiciman, Ranveer Chandra, and Wei-Ting Chen.

  ECCV 2026.

  SPIRE learns latent slide-design preferences through inverse planning, training collaborative agents to reconstruct structurally corrupted slides with reinforcement learning.

  Paper

• Conformal Path Reasoning: Trustworthy Knowledge Graph Question Answering via Path-Level Calibration
  Shuhang Lin, Chuhao Zhou, Xiao Lin, Zihan Dong, Kuan Lu, Zhencan Peng, Jie Yin, Dimitris N. Metaxas.

  ICML 2026.

  CPR combines query-level conformal calibration with learned path scores to produce compact answer sets with coverage guarantees for knowledge graph question answering.

  Paper / Code

• MMedAgent-RL: Optimizing Multi-Agent Collaboration for Multimodal Medical Reasoning
  Peng Xia, Jinglu Wang, Yibo Peng, Kaide Zeng, Zihan Dong, Xian Wu, Xiangru Tang, Hongtu Zhu, Yun Li, Linjun Zhang, Shujie Liu, Yan Lu, and Huaxiu Yao.

  ICLR 2026.

  MMedAgent-RL trains triage and attending agents with reinforcement learning to route cases to specialists and integrate their judgments for multimodal medical reasoning.

  Paper

• AceSearcher: Bootstrapping Reasoning and Search for LLMs via Reinforced Self-Play
  Ran Xu, Yuchen Zhuang, Zihan Dong, et al.

  NeurIPS 2025 Spotlight.

  AceSearcher trains an LLM to alternate between query decomposition and answer generation through supervised fine-tuning and reinforced self-play for multi-hop search.

  Paper / Code

• Mitigating Heterogeneous Token Overfitting in LLM Knowledge Editing
  Tianci Liu, Ruirui Li, Zihan Dong, Hui Liu, Xianfeng Tang, Qingyu Yin, Linjun Zhang, Haoyu Wang, and Jing Gao.

  ICML 2025.

  OVERTONE adaptively smooths token-level training targets to reduce heterogeneous overfitting and preserve reasoning about newly edited LLM knowledge.

  Paper

• Hybrid-Policy Self-Editing for Composable Unstructured Knowledge Editing
  Tianci Liu, Zihan Dong, Tianchun Li, Yi-Chung Chen, Qiming Cao, Xingchen Wang, Shiyang Wang, Zichen Miao, Linjun Zhang, Haoyu Wang, and Jing Gao.

  Submitted to ICLR 2027.

  Hybrid-policy self-distillation helps edited language models apply newly inserted knowledge to factual questions and multi-hop reasoning without external supervision.

  Paper

• DARE: Difficulty-Adaptive Reinforcement Learning with Co-Evolved Difficulty Estimation
  Yang Zhou, Can Jin, Zihan Dong, Zhepeng Wang, Yanting Yang, Shiyu Zhao, Lei Li, Runxue Bao, Yaochen Xie, and Dimitris N. Metaxas.

  Submitted to ICLR 2027.

  DARE co-evolves difficulty estimation with the policy and adapts data selection, training strategies, and computation to improve reasoning across easy and hard tasks.

  Paper / Code

• Evidence Over Plans: Online Trajectory Verification for Skill Distillation
  Yang Zhou, Zihan Dong, et al.

  Submitted to ICLR 2027.

  Distills agent skills from environment-verified execution trajectories and uses a trajectory-level diagnostic to keep those skills grounded in observed evidence.

  Paper / Code

• RUBRIC-ARROW: Alternating Pointwise Rubric Reward Modeling for LLM Post-training in Non-verifiable Domains
  Haoxiang Jiang, Zihan Dong, Tianci Liu, Wanying Wang, Ran Xu, Tony Yu, Linjun Zhang, and Haoyu Wang.

  Preprint, 2026.

  RUBRIC-ARROW alternates training a rubric generator and a rubric-conditioned judge from pairwise preferences to provide pointwise rewards for non-verifiable tasks.

  Paper

• AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
  Can Jin, Yang Zhou, Qixin Zhang, Hongwu Peng, Di Zhang, Zihan Dong, Marco Pavone, Ligong Han, Zhang-Wei Hong, Tong Che, Dimitris N. Metaxas.

  Preprint, 2026.

  AIRL-S learns dense process rewards from reference trajectories through adversarial inverse reinforcement learning, using the same reward model for policy training and test-time search.

  Paper

• LUCID-SAE: Learning Unified Vision-Language Sparse Codes for Interpretable Concept Discovery
  Difei Gu, Yunhe Gao, Gerasimos Chatzoudis, Zihan Dong, et al.

  Preprint, 2026.

  LUCID-SAE learns a shared sparse dictionary for image patches and text tokens, aligning interpretable concepts across modalities through optimal transport without labels.

  Paper