Hi, my name is Wenhao Wu (吴文浩). I am a third-year master's student at Nanjing University, supervised by Prof. Zhi Wang (王志). My research interests focus on Large Language Model Agents, Retrieval-Augmented Generation, Meta Reinforcement Learning, and Embodied AI.
MemFuse: Multi-Source Memory Fusion from Fragmented Observations
Chao Li, Yuanfa Li, Wenhao Wu, Xule Liu, Zhi Wang, Kun Shao
Introduces MemFuseBench for evaluating temporal reasoning and cross-source evidence fusion, and a dual-layer causal fusion graph that preserves source provenance while organizing fragmented observations into retrievable memories.
From Conflict to Consensus: Boosting Medical Reasoning via Multi-Round Agentic RAG
Wenhao Wu, Zhentao Tang, Yafu Li, Shixiong Kai, Mingxuan Yuan, Zhenhong Sun, Chunlin Chen, Zhi Wang
Proposes a conflict-guided multi-round retrieval framework for medical reasoning. The method uses conflicts among sampled responses as retrieval signals, generates precise queries, and reranks historical reasoning trajectories with sequence-level entropy or a lightweight BERT verifier.
Text-to-3D Policy: Fine-Grained Language-Behavior Alignment for Unseen Specification Generalization
Xinhao Yang*, Wenhao Wu*, Ning Lv, Yanshen Ding, Zhenhong Sun, Daoyi Dong, Chunlin Chen, Zhi Wang
Introduces T3DP, a text-to-3D policy framework that aligns language tokens with behavioral segments to preserve fine-grained specification geometry and improve generalization to unseen robotic manipulation behaviors.
HarnessPAI: An Evolving Harness for Physical AI
Xin Wang*, Wenhao Wu*, Menghao Zhang*, Zhi Wang, Kun Shao, Jian Luan, Yang Li, Qing Li, Shangding Gu, Huichi Zhou, Shuqing Shi, Fei Ni, Shuo Lu, Weicheng Meng, Kang Li, Jin Wu, Kang Zhao, Shangmin Guo, Gen Li, Yongqiang Tang, Zhizhong Zhang, Yuan Xie, Heng Qu
Introduces a model- and embodiment-agnostic harness for Physical AI in which the harness and action model co-evolve through rollout feedback: action primitives are organized into executable programs, failures are distilled into reusable skills, and the system improves without retraining the underlying action model.
Look Before You Leap: Distilling Tree Search into Action Evaluation for Frozen VLA Models
Xinyi Xie, Zican Hu, Zhanyu Liu, Yicheng Dong, Wenhao Wu, Haoran Li, Chunlin Chen, Zhi Wang*, Pichao Wang
Proposes SVA (Search, Value, and Act), which keeps the VLA backbone frozen and equips it with long-term consequence awareness. Monte-Carlo tree search explores the policy's output distribution in simulation and is distilled into a lightweight Q-value evaluator that selects the best of N candidate actions at deployment without simulator access.
Mixture-of-Experts Meets In-Context Reinforcement Learning
Wenhao Wu, Fuhong Liu, Haoru Li, Zican Hu, Daoyi Dong, Chunlin Chen, Zhi Wang
Introduces a dual-MoE architecture for in-context RL. Token-wise MoE routes state, action, and reward information to specialized experts, while Task-wise MoE uses contrastive task representations to reduce gradient conflict in multi-task learning.
Text-to-Decision Agent: Offline Meta-Reinforcement Learning from Natural Language Supervision
Shilin Zhang*, Zican Hu*, Wenhao Wu*, Xinyi Xie, Jianxiang Tang, Chunlin Chen, Daoyi Dong, Yu Cheng, Zhenhong Sun, Zhi Wang
Aligns trajectory-derived decision embeddings with natural-language supervision. The framework trains a trajectory world model, performs contrastive alignment with LoRA-tuned language models, and conditions scalable policies on aligned text embeddings for zero-shot generalization.
Meta-DT: Offline Meta-RL as Conditional Sequence Modeling with World Model Disentanglement
Zhi Wang, Li Zhang, Wenhao Wu, Yuanheng Zhu, Dongbin Zhao, Chunlin Chen
Formulates offline meta-RL as conditional sequence modeling with world-model disentanglement. A context-aware world model learns task representations, while self-guided and complementary prompts support generalization to new tasks.
CAR-bench Challenge — Open Track
IJCAI-ECAI 2026 Competition Track · Darwin Agent team
Wenhao Wu*, Menghao Zhang*, Xin Wang*, Zhi Wang, Kun Shao, Jian Luan
Developed TRACE, a self-evolving Skill Bank for consistent and limit-aware LLM agents. Our team won both the Rank Award and Innovation Award, ranking 1st in the Open Track on the official hidden-set evaluation.
