Shuo He

LongCat-DeepResearch Technical Report

Meituan Longcat Team

Paper   Code   Blog

We present LongCat-DeepResearch, a multi-agent deep research system that separates global planning from parallel section-level investigation and targeted revision, enabling comprehensive, evidence-grounded reports while achieving strong results across several deep research benchmarks.

Negative-Only Policy Optimization for One-Sided Verifiable Rewards

Jiacheng Xu, Shuo He, Fuxiang Zhang, Chaojie Wang, Bo An

Neurips 2026   Paper

We show that standard RLVR becomes unreliable when verifiers can confidently detect failures but cannot confirm correctness, and we address this with NOPO by suppressing only trusted negative samples while treating non-falsified candidates as unlabeled, improving code, logic, and math performance under imperfect verification.

Dr. MAS: Stable Reinforcement Learning for Multi-Agent LLM Systems

Lang Feng, Longtao Zheng, Shuo He, Fuxiang Zhang, Bo An

Neurips 2026   Paper  Code

We show that multi-agent LLM RL becomes unstable when we use one global normalization for all agents, and we fix it with Dr. MAS by normalizing each agent’s advantages using its own reward statistics, which stabilizes training and improves results on multi-agent math and search tasks.

Beyond Next-Observation Prediction: Agent-Authored World Modeling for Sequential Decision Making

Guangfeng Cai, Kaibing Yang, Shuo He, Yu Li, Shengtian Yang, Jiaqi Lv, Lei Feng

EMNLP 2026   Paper

We show that next-observation prediction may overlook the dynamics most relevant to an agent’s decisions, and we address this with AAWM by letting the agent identify its own information needs and synthesize relevant evidence across trajectories into decision-aware training targets, improving world modeling across multiple environments and training settings.

Beyond Trajectory-Level Attribution: Graph-Based Credit Assignment for Agentic Reinforcement Learning

Xin Cheng, Shuo He, Lang Feng, Haiyang Xu, Ming Yan, Lei Feng^†, Bo An

ICML 2026   Paper  Code

We show that existing group-based RL methods for LLM agents rely on coarse trajectory-level credit assignment, which fails to identify valuable intermediate decisions hidden inside unsuccessful trajectories. Hence, we propose GraphGPO, which aggregates rollouts into a unified state-transition graph and leverages graph-level global information to estimate each state's distance to the goal, enabling more faithful step-level credit assignment and improved agent learning.

Phase-Aware Mixture of Experts for Agentic Reinforcement Learning

Shengtian Yang, Yu Li, Shuo He, Yewen Li, Qingpeng Cai, Peng Jiang, Lei Feng^†

ICML 2026   Paper  Code

We show that conventional MoE policies in agentic RL suffer from phase fragmentation because token-level routing scatters temporally coherent reasoning stages across different experts, preventing effective specialization. Hence, we propose PA-MoE, which learns latent task phases with a lightweight phase router and assigns phase-consistent segments to dedicated experts, enabling stronger expert specialization and improved performance on complex agentic tasks.

Hierarchy-of-Groups Policy Optimization for Long-Horizon Agentic Tasks

Shuo He, Lang Feng, Qi Wei, Xin Cheng, Lei Feng^†, Bo An

ICLR 2026   Paper   Code

We show that step-wise group-based RL for long-horizon LLM agents can suffer from historical context inconsistency, which severely biases stepwise advantage estimates and harms optimization. Hence, we propose HGPO, which forms multiple hierarchical groups per step based on historical-context consistency and adaptively aggregates their advantage estimates to get a better bias-variance tradeoff without extra models or rollouts.

Online Causal Kalman Filtering for Stable and Effective Policy Optimization

Shuo He, Lang Feng, Xin Cheng, Lei Feng^†, Bo An

Preprint   Paper  Code

We make LLM policy training more stable by using an online causal Kalman filter to smooth noisy token-level importance ratios across a sequence, which gives steadier updates and better results on challenging math reasoning tasks.

Solving Puzzles? Jailbreaking Multimodal Large Language Models! (reported to Anthropic's Safeguards Team)

Shuo He, Qi Wei, Zhifang Zhang, Feng Liu, Lei Feng, Bo An

OpenReview   Paper  

We show that multimodal LLM safety can be bypassed even in high-risk settings when malicious intent is hidden inside cross-modal visual puzzles, making the attack more covert than prior jailbreaks. Hence, we present the first vision-prompt-based puzzle jailbreak pipeline for MLLMs—combining textual taxonomy generation, visual object decomposition, and puzzle construction—and build PuzzleV-JailBench, covering 144 weapon and 54 chemical cases, to demonstrate that both open-source and production MLLMs can still be induced to produce dangerous outputs.

Defending Multimodal Backdoored Models by Repulsive Visual Prompt Tuning

Zhifang Zhang, Shuo He^†, Haobo Wang, Bingquan Shen, Lei Feng^†

NeurIPS 2025   Paper   Code

We defend backdoored multimodal models like CLIP by using Repulsive Visual Prompt Tuning (RVPT), which tunes only small visual prompts on a few clean samples and uses a feature-repelling loss to make the model ignore trigger features, reducing attack success from 89.70% to 2.76% and generalizing across datasets.

BDetCLIP: Multimodal Prompting Contrastive Test-Time Backdoor Detection

Yuwei Niu^*, Shuo He^*, Qi Wei, Feng Liu, Lei Feng

ICML 2025   Paper   Code

We propose BDetCLIP, a fast test-time method that detects backdoored CLIP inputs by using contrastive prompts and checking that backdoored images stay almost unchanged when we change the class text.

A Closer Look at Backdoor Attacks on CLIP

Shuo He, Zhifang Zhang, Feng Liu, Roy Ka-Wei Lee, Bo An, Lei Feng

ICML 2025   Paper   Code

We study how backdoor attacks change CLIP by breaking image features into patches, attention heads, and MLPs, we find different attacks infect different parts of the model, and we use these findings to detect and repair infected parts (or filter suspicious samples) at inference time.

Representation Surgery in Model Merging with Probabilistic Modeling

Qi Wei, Shuo He, Jiahan Zhang, Lei Feng, Bo An

ICML 2025   Paper   Code

Influence-Based Fair Selection for Sample-Discriminative Backdoor Attack

Qi Wei, Shuo He, Enneng Yang, Tingcong Liu, Haobo Wang, Lei Feng, Bo An

AAAI 2025 (Oral)   Paper

We find that when triggers are very small, sample-discriminative backdoor attacks pick poisoned samples unfairly across classes and get uneven success, so we propose IFS that uses influence scores with efficient pruning and class-aware thresholds to select high-impact but class-balanced poisoned samples, improving attack success on four datasets.

Candidate Label Set Pruning: A Data-centric Perspective for Deep Partial-label Learning

Shuo He, Chaojie Wang, Guowu Yang, Lei Feng

ICLR 2024 (Oral)   Paper   Code

Partial-label Learning with Mixed Closed-set and Open-set Out-of-candidate Examples

Shuo He, Lei Feng, Guowu Yang

SIGKDD 2023   Paper   Code

Candidate-aware Selective Disambiguation Based on Normalized Entropy for Instance-dependent Partial-label Learning

Shuo He, Guowu Yang, Lei Feng

ICCV 2023   Paper   Code

A Generalized Unbiased Risk Estimator for Learning with Augmented Classes

Senlin Shu, Shuo He, Haobo Wang, Hongxin Wei, Tao Xiang, Lei Feng

AAAI 2023   Paper   Code

Partial Label Learning with Semantic Label Representations

Shuo He, Lei Feng, Fengmao Lv, Wen Li, Guowu Yang

SIGKDD 2022   Paper   Code

Discriminatively Relabel for Partial Multi-label Learning

Shuo He, Ke Deng, Li Li, Senlin Shu, Li Liu

ICDM 2019 (Oral)   Paper   Code

Collaboration Based Multi-label Learning

Lei Feng, Bo An, Shuo He

AAAI 2019   Paper

Estimating Latent Relative Labeling Importances for Multi-label Learning

Shuo He, Lei Feng, Li Li

ICDM 2018   Paper   Code