Publications
33 papers · * denotes equal contribution
2026
NeurIPS 2026
Beyond Correctness: Robustness-Driven Evolutionary Self-Training for Large Language Models
TREST combines evolutionary search with supervised fine-tuning to prioritize robust reasoning trajectories, favoring insight-driven mathematical strategies over brittle solutions.
NeurIPS 2026
The Mirage of Optimizing Training Policies: Monotonic Inference Policies as the Real Objective for LLM Reinforcement Learning
MIPU addresses training-inference mismatch in LLM reinforcement learning through sampler-referenced updates and inference-gap-aware acceptance, improving reasoning performance and training stability.
NeurIPS 2026
Does Your Large Language Model Have An Intuitive Sense of The Difficulty of A Question?
RLDP estimates problem difficulty from hidden activations in a single forward pass, while RLDP-AdaSwitch uses that signal to allocate reasoning effort efficiently.
DAI 2026 Oral
Scaling Reinforcement Learning via Agent Ensemble
An agent-ensemble approach to scaling reinforcement learning by combining multiple agents with complementary behaviors.
ICML 2026 Spotlight
ScaleMoE: Mixture-of-Experts for Scalable Continuous Control in Actor-Critic Reinforcement Learning
ScaleMoE brings a mixture-of-experts architecture to actor-critic reinforcement learning, scaling network capacity to improve sample efficiency and performance on continuous control tasks.
SPOT@ICLR 2026
Beyond Scalar Critics: A Distributional Perspective on Reinforcement Learning with Verifiable Rewards for LLMs
DistRLVR is a distributional RL framework for LLM post-training with verifiable rewards that models token-level return distributions and uses tail-aware advantages to improve sample efficiency and reasoning performance.
ICLR 2026
CellAgent: LLM-Driven Multi-Agent Framework for Natural Language-Based Single-Cell Analysis
An LLM-driven multi-agent system for end-to-end single-cell and spatial transcriptomics analysis through natural language, combining hierarchical planning, expert tools, and self-reflective optimization.
ICLR 2026
Squeeze the Soaked Sponge: Efficient Off-policy RFT for Large Language Models
ReMix brings off-policy reinforcement finetuning to LLM post-training by reusing rollout data from past policies, dramatically reducing training cost while staying competitive on math reasoning benchmarks.
LLA@ICLR 2026
Benchmarking Continual Agent Memory for Online Learning, Transfer, and Forgetting
AgentMemoryBench is a unified benchmark for continual agent memory that measures improvement, retention, forgetting, transfer, and conflict resolution over time, together with a multi-memory baseline called MEMs.
WWW 2026
AFE-Master: Enhancing LLM-Driven Autonomous Feature Engineering with Domain-Specific Language Parsing and Guided Local Search
AFE-Master introduces domain-specific language parsing with guided local search to strengthen LLM-driven autonomous feature engineering.
KDD 2026
PACE: Unleashing the Power of Code Embeddings to Boost AutoML Agents
PACE leverages code embeddings to enhance LLM-driven AutoML agents, improving feature engineering and model selection through structured code representations.
2025
NeurIPS 2025
Improving Reward Models with Proximal Policy Exploration for Preference-Based Reinforcement Learning
A reward-modeling approach for preference-based reinforcement learning built on proximal policy exploration.
ICCV 2025
RoboAnnotatorX: A Comprehensive and Universal Annotation Framework for Accurate Understanding of Long-horizon Robot Demonstration
RoboAnnotatorX presents a comprehensive and universal framework for annotating long-horizon robot demonstrations to enable accurate understanding.
ACL 2025
DualRAG: A Dual-Process Approach to Integrate Reasoning and Retrieval for Multi-hop Question Answering
A dual-process framework that combines retrieval and reasoning to improve multi-hop question answering.
ACL 2025
War of Thoughts: Competition Stimulates Stronger Reasoning in Large Language Models
Investigates how competitive mechanisms enhance the reasoning capabilities of LLMs, leading to improved performance on complex tasks.
AAAI 2025
Key Decision-Makers in Multi-Agent Debates: Who Holds the Power?
Analyzes mediator roles and decisive voices within multi-agent debate frameworks, revealing how influence shifts throughout deliberation.
WWW 2025
SheetAgent: Towards a Generalist Agent for Spreadsheet Reasoning and Manipulation
SheetAgent is a generalist LLM agent for spreadsheet reasoning and manipulation across realistic multi-step tasks.
2024
IJCAI 2024
vMFER: Von Mises-Fisher Experience Resampling Based on Uncertainty of Gradient Directions for Policy Optimization
An experience resampling method that uses gradient-direction uncertainty for more stable policy improvement.
IJCAI 2024
ENOTO: Improving Offline-to-Online Reinforcement Learning with Q-Ensembles
An offline-to-online reinforcement learning method that improves transition efficiency with Q-ensembles.
AAMAS 2024
A Trajectory Perspective on the Role of Data Sampling Techniques in Offline Reinforcement Learning
Organizing samples in a trajectory manner can substantially improve the learning efficiency for offline RL algorithms.
AAAI 2024
OVD-Explorer: Optimism Should Not Be the Sole Pursuit of Exploration in Noisy Environments
OVD-Explorer achieves noise-aware optimistic exploration by separating useful uncertainty from stochastic noise for continuous control.
NeurIPS 2024
PERIA: Perceive, Reason, Imagine, Act via Holistic Language and Vision Planning for Manipulation
A holistic language-and-vision planning framework that unifies perception, reasoning, imagination, and action for manipulation.
ICML 2024
KISA: A Unified Keyframe Identifier and Skill Annotator for Long-horizon Robotics Demonstrations
A unified keyframe identification and skill annotation method for long-horizon robot demonstrations.
NeurIPS 2024 Workshop
Optimizing Reward Models with Proximal Policy Exploration in Preference-Based Reinforcement Learning
A preference-based reinforcement learning study on improving reward models with proximal policy exploration.
ICLR 2024
Uni-RLHF: Universal Platform and Benchmark Suite for Reinforcement Learning with Diverse Human Feedback
A unified platform and benchmark suite for reinforcement learning with diverse human feedback.
2023
TNNLS 2023
Exploration in Deep Reinforcement Learning: From Single-Agent to Multi-Agent Domain
A survey of exploration methods in deep reinforcement learning, spanning single-agent and multi-agent settings.
CAAI AIR 2023
OSCAR: OOD State-Conservative Offline Reinforcement Learning for Sequential Decision Making
An offline reinforcement learning method that stays conservative on out-of-distribution states for sequential decision-making.
ICLR 2023
EUCLID: Towards Efficient Unsupervised Reinforcement Learning with Multi-choice Dynamics Model
An unsupervised reinforcement learning method that improves efficiency with a multi-choice dynamics model.
2021
ASE 2021
FIGCPS: Effective Failure-Inducing Input Generation for Cyber-Physical Systems with Deep Reinforcement Learning
A deep reinforcement learning approach for generating failure-inducing inputs in cyber-physical systems testing.