#Paper/Research

AIThinking Machines published a post on on-policy distillation, a training approach combining the error-correcting relevance of RL with the reward density of SFT. The quoted post reports that in math reasoning and an internal chat assistant, on-policy distillation can outperform other approaches at a fraction of the cost.

AIThinking Machines Lab describes on-policy distillation, which samples rollouts from a student model and has a teacher grade each token with reverse KL. The authors report that this matches Qwen3-style reasoning results at a fraction of RL's cost, with AIME'24 reaching 70% in about 150 steps from a 400k SFT checkpoint. The method also helps recover instruction-following behavior lost during fine-tuning on internal documents.

Why it matters: The post explains why on-policy distillation gives dense per-token feedback, letting a small model match RL results at much lower compute cost.

AIStanford and Cognition AI researchers introduced Kevin-32B, a 32B-parameter model trained with multi-turn reinforcement learning to write CUDA kernels. On KernelBench, it solves 89% of tasks at best@16 and achieves 65% average correctness over eight refinement steps, versus 53% for o4-mini and 51% for o3. Its best@16 speedup is 1.41x, and multi-turn training outperforms single-turn training as refinement steps increase.

AILiquid AI reports STAR, an evolutionary algorithm that synthesizes tailored neural network architectures from a numerical genome representation. The authors say it produced hundreds of designs that outperform Transformer and hybrid architectures in quality, with smaller caches and parameter counts, and can optimize for latency on target hardware. The full method is described in the arXiv technical report 2411.17800.

Why it matters: The post explains how evolutionary search over a new architecture design space produced designs beating Transformers and hybrids, giving a concrete method for quality versus latency and memory trade-offs.

AICognition reports that its agent Devin resolved 79 of 570 sampled SWE-bench issues, a 13.86% success rate, without being given the files to edit. The report says this exceeds the best previous unassisted baseline of 1.96% and the best assisted result of 4.80%. It also describes the adapted evaluation setup, a 45-minute runtime limit, and cases where Devin failed on multi-file edits.

Why it matters: The report explains how SWE-bench was adapted for end-to-end agent evaluation, with failure cases that clarify where the 13.86% result comes from and its limits.