Talks and presentations

Covering the Pareto Frontier with LLM-Coordinated Interpretable Policy Library Permalink

August 25, 2026

Conference Talk, IFAC 2026 (23rd IFAC World Congress), Busan, South Korea

Conference talk at IFAC 2026, which proposes a vision of self-designing industrial autonomy, and a paradigm shift from tuning opaque neural weights to searching over explicit code logic. We provide a proof-of-concept: a dual-agent LLM loop to autonomously build an interpretable policy library. On industrial multi-objective tasks, it matches deep RL in 30 iterations (<90 min) with fully transparent, deployment-ready code.

Enhancing Human Feedback Reliability in Preference-Based Reinforcement Learning Permalink

November 27, 2025

Defense, PhD proposal defense, Tsinghua University, Beijing, China

PhD proposal defense on enhancing the reliability of human feedback in Preference-Based RL. The work identifies three fundamental causes of unreliable feedback: behavioral indistinguishability, judgment ambiguity, and multi-objective conflicts; and proposes a three-part technical framework: skill-driven PbRL (S-EPOA), contrastive learning-based PbRL (CLARIFY), and preference-based multi-objective RL (Pb-MORL).

S-EPOA: Overcoming the Indistinguishability of Segments with Skill-Driven Preference-Based RL Permalink

August 27, 2025

Conference Talk, IJCAI 2025, Guangzhou, China

Conference talk at IJCAI 2025 on S-EPOA, which leverages unsupervised skill discovery to select trajectory segments that differ widely in skill, so humans can compare them efficiently in preference-based RL.

Safe-GPI: Safety-Guaranteed Policy Composition Via Generalized Policy Improvement for Autonomous Vehicles Permalink

August 15, 2025

Conference Talk, IEEE CASE 2025 (Outstanding WiRA Student Paper Award), Los Angeles, California, USA

Conference talk at IEEE CASE 2025 on safety-guaranteed policy composition via Generalized Policy Improvement for autonomous vehicles, which resolves the ‘impossible triangle’ of performance vs. safety vs. sample-efficiency. This paper received the Outstanding WiRA Student Paper Award.

CLARIFY: Contrastive Preference Reinforcement Learning for Untangling Ambiguous Queries Permalink

March 24, 2025

Seminar, Group seminar, IIIS, Tsinghua University, Beijing, China

Seminar on CLARIFY (ICML 2025). We introduced a contrastive learning framework that learns a structured trajectory embedding space to quantify human judgment clarity, coupled with a rejection-sampling query mechanism that actively filters ambiguous comparisons. The method significantly improves both human annotation reliability and offline preference-based RL performance.

Ni Mu