Training and inference alignment in distillation and few-step models

Recent work covers rollout-mismatch tolerance in on-policy distillation, visual-agent post-training, few-step diffusion RL, and sparse model repair. Reframe reports gains across 14 visual benchmarks and Tree-Ising reports gains including 54.6% versus 51.7% on MATH-500, but effect sizes, edit-selection protocols, and transfer need verification; FewStepRL’s recommendations depend on reward-gradient cost.

Sources (3)

Updated Nov 12, 2026

AI Research Pulse

Training and inference alignment in distillation and few-step models

Scaling Mathematical Proof Supervision from Structured Textbooks

Incentivizing Self-Learning in Large Language Models via Reinforcement Learning

Memorization vs. In-Context Learning in State Space Models