2 papers
cs.LG2026
ReNFT: Repairing Mode Collapse in Reward Post-Training via Internal Probability-Mass Recalibration
Yuchen Bao, Chao Wen, Haowei Wang +10
Reward post-training of diffusion generators inevitably concentrates probability mass on a few reward-favored modes, a mode collapse that erases within-prompt diversity. Existing m…
cs.RO2026
STEAM: Self-Supervised Temporal Ensemble Advantage Modeling for Real-World Robot Learning
Zhihao Liu, Qiuyi Gu, Yitao Wang +16
Real-world robot learning increasingly relies on heterogeneous data, but demonstrations and rollouts often mix useful progress with stalls, corrections, and suboptimal behavior. Ef…