A study accepted to ACM ICAIF '26 found that on GPT-OSS-20B, a generic feedback line matched self-written critiques when used for hindsight self-distillation of multi-step financial reasoning. In the method, the student model identifies its own first wrong step and writes a critique, and a frozen copy conditioned on the prefix plus the critique is distilled back into the student under reverse KL divergence. The finding indicates that the specific content of the critique may matter less than the presence of feedback in this setup.
We study hindsight self-distillation for multi-step financial reasoning on GPT-OSS-20B. The student finds its own first wrong step and writes a critique. A frozen copy, conditioned on the prefix plus the critique, is distilled back into the student under reverse KL. Notably, replacing the critique w
EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.