Fin-SD tests how much hint signal self-distillation needs
Original title: How Much Does the Hint Matter? Scope Conditions for Hindsight Self-Distillation in Financial Reasoning
On GPT-OSS-20B, a generic hint matched the full 3.03-point gain and used 9.5% fewer tokens than full-solution distillation.
The paper introduces Fin-SD, a localized hindsight self-distillation method for multi-step financial reasoning without an external teacher. When GPT-OSS-20B fails a problem, it marks the first wrong step, writes a short critique, and a frozen copy conditioned on the correct prefix plus hint is distilled back under reverse KL. On a custom accounting benchmark, replacing the critique with generic feedback matched the entire 3.03 percentage-point gain on FinanceReasoning Hard. Same-family GPT-OSS-120B hints were not materially better, Fin-SD matched full-reference on-policy self-distillation with 9.5% fewer tokens, and MMLU-Pro plus a FinTrust audit showed no material regression.
Unconfirmed: Results are the authors' findings on their custom benchmark.
- No external model contributes a gradient.
- The reported gain is 3.03 percentage points on FinanceReasoning Hard.
- GPT-OSS-120B hints, about 6x larger, added no material benefit.
- Token use was 9.5% lower than full-reference self-distillation.
From the original article
1 min read at huggingface.co
Financial institutions operating under data-residency constraints are restricted to open-weight models small enough to locally host, which trail frontier systems on multi-step quantitative reasoning.
By Johnny Yu, Cyril Gorlla · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.