Sakana AI's RLT: a 7B teacher distills a 32B student that beats large-teacher baselines
青稞AI · wechat · 2026-10-22
Sakana AI published a paper flipping the distillation paradigm: instead of training an ultra-strong teacher for students to mimic, train the teacher to explain solutions — explaining with a known answer is far easier than solving from scratch, just as a physics teacher needn't out-Newton Newton.
Method: teachers are optimized with dual rewards —
- rSS: given the question plus the teacher's explanation, can the student solve new problems? (measures usefulness)
- rKL: does the student's self-generated reasoning match the teacher's? (measures internalization)
Rewards are dense and immediate, and the teacher's task is simpler than the student's, enabling small teachers.
Results (AIME2024/MATH500/GPQA Diamond composite):
- 7B student: RLT with a 7B teacher hits 49.50%, beating Bespoke (DeepSeek-R1 traces) at 46.60%; even 1K samples yield 47.43%
- 32B student: a 7B RLT teacher reaches 73.23%, surpassing large-teacher Bespoke-32B at 71.47% — "small teacher, big student"
- References: vanilla Qwen2.5-7B 39.17%, QwQ-32B 65.03%, DeepSeek-R1 82.87%
The authors frame it as common-sense-driven research: train models to do what they're trained for — teaching — enabling cheaper deployment of capable models.
More from Research
- RFD4-Proteina's 'underrated' docs detail partial diffusion for protein redesign — AllThingsApx · 2026-10-23
- MPI-Sec PhD Thesis on Vulnerability Detection Benchmarks Includes Pre-LLM Classic "Top Score on the Wrong Exam" — mboehme_ · 2026-10-23
- LiveNerf's 29-day data finds no nerfing of Claude Opus 5.5 — TheOnlyVibemaster · 2026-10-23
- New paper makes 9 falsifiable predictions toward a major AI-enabled terror attack by 2027-2028 — StephenLCasper · 2026-10-23
- HumanoidPose recovers humanoid robots' full-body joint states from RGB video alone — cholz · 2026-10-23
- Sophont's Holon-S1 predicts genes and biomarkers from routine pathology slides — iScienceLuvr · 2026-10-23