Generic one-line feedback matches self-critiques in hindsight self-distillation, study finds
jhzyu · reddit · 2026-10-29
Authors of an ACM ICAIF '26 paper study hindsight self-distillation for multi-step financial reasoning on GPT-OSS-20B: the student locates its first wrong step and writes a critique, then a frozen copy conditioned on prefix plus critique is distilled back under reverse KL. Surprisingly, replacing the critique with a fixed generic feedback line matches the full 3-point gain, and critiques from a 6x larger model don't help — suggesting positional information, not critique content, drives the benefit. Fin-SD matches OPSD from full reference solutions with 9.5% fewer tokens. Paper and code are open.
More from Research
- New paper challenges JEPA feature-selection theory: unpredictable features can be kept — andreea7b · 2026-10-29
- SREGym to debut at Bench Summit, targeting AI SRE evaluation frontier — tianyin_xu · 2026-10-29
- AA-Robotics benchmark: GPT-6 Astra tops robot-arm control, still ~100x slower than humans — ArtificialAnlys · 2026-10-29
- Delta-Matching closes the FP8 attention gap: all 7 core matmuls in FP8, loss matches BF16 — songhan_mit · 2026-10-29
- RFD4-Proteina lands on Hugging Face as HF publishes its October AI for Science roundup — AllThingsApx · 2026-10-29
- Google's PLE: 85 real engineering tasks, best agent config scores only 50.6% — rohanpaul_ai · 2026-10-29