Bridged Self-Distillation: On-Policy Reasoning Transfer from Any Teacher, Even Black-Box
TUM · hf · 2026-10-23
Researchers at TU Munich and LMU propose Bridged Self-Distillation (BSD), a unified framework addressing two mismatches when distilling from a stronger teacher: its solutions may be hard for a smaller student to imitate, and its token-level probabilities may be incompatible or unavailable (closed APIs).
Method: BSD uses privileged reasoning traces to score student rollouts and guides updates via outcome verification, enabling on-policy reasoning transfer from any teacher across three scenarios:
- Same family, shared tokenizer: standard OPD via token-level distribution matching
- Cross family, different tokenizer: cross-tokenizer OPD via vocabulary mapping
- Closed API, text only: black-box OPD based on response-based feedback
More from Research
- LLMs make code so cheap that recursive rewrites could make verified software tractable — jd_pressman · 2026-10-23
- Hugging Face hands OpenEnv RL environments project to Linux Foundation for neutral governance — lmoroney · 2026-10-23
- Formal verification debate: narrowing code search space to the verifiable 1% — jd_pressman · 2026-10-23
- New paper: AI reviews beat human peer review at its best, with policy implications — RexDouglass · 2026-10-23
- Vibe-proving ASAN equivalence with Claude unearths a wave of PHP memory bugs — jd_pressman · 2026-10-23
- File-Retrieved Tool Output Makes GPT Agents Substitute Approximations, Costing Accuracy — ucriverside · 2026-10-23