2026

CESAR concept: audio LLM reasoning with process rewards
ICLR 2026

CESAR: Audio LLM Reasoning via Process Rewards

J. Fan*, R. Ren, J. Li, R. Pandey, P.G. Shivakumar, I. Bulyko, A. Gandhe, G. Liu, Y. Gu
SOTA MMAU · Beats Gemini 2.5 Pro & GPT-4o Audio

Resolves test-time inverse scaling in Audio LLMs by rewarding the reasoning process.

Project Page OpenReview arXiv

2025

ADRPO concept: adaptive per-sample regularization
NeurIPS 2025

ADRPO: Adaptive Divergence for Generative Models

J. Fan*, T. Wei, C. Cheng, Y. Chen, G. Liu
2B SD3 surpasses 4.8B & 12B models

Sample-level adaptive divergence regularization — high-value samples explore freely, poor samples stay constrained.

Project Page OpenReview arXiv
ORW-CFM-W2 concept: flow matching with diverse clusters
ICLR 2025

ORW-CFM-W2: Flow Matching Self-Evolution

J. Fan*, S. Shen, C. Cheng, Y. Chen, C. Liang, G. Liu
First online RLHF for flow matching models

No human data, no mode collapse. W2 regularization preserves generation diversity.

Project Page OpenReview
AC-Flow concept: actor-critic intermediate feedback on a flow trajectory
Preprint 2025

AC-Flow: Actor-Critic for Flow Matching

J. Fan*, C. Cheng, S. Shen, X. Zhou, G. Liu
Under Review

Intermediate feedback + dual-stability for robust flow matching fine-tuning on SD3.

Project Page arXiv

2023

LBC concept: behavior selection breaking the human-record line
ICLR 2023 ⭐ Oral · 5/4176

LBC: Breaking 24 Atari World Records

J. Fan*, Y. Zhuang, Y. Liu, J. Hao, B. Wang, J. Zhu, H. Wang, S.-T. Xia
10,077% HNS · 24 records · 78× data efficient

Learnable behavior control via hybrid policy mapping + bandit meta-controller.

Project Page OpenReview

2022

GDI concept: data distribution driving policy iteration
ICML 2022

GDI: Generalized Data Distribution Iteration

J. Fan*, C. Xiao
22 world records · 9,620% HNS · 500× vs Agent57

Unified RL framework showing data distribution is the key to superhuman efficiency.

Project Page PMLR