Hindsight Self-Distillation Boosts Local Models' Financial Reasoning Without External Teachers

A UC Berkeley and CTGT paper shows that on-policy hindsight self-distillation—replacing detailed self-critique with a single generic feedback line—improves local models' multi-step financial reasoning by about 3 points without external teachers.

2026-10-29 ~ 2026-10-29 · 2 related posts

• Fin-SD: hindsight self-distillation without external teachers boosts local models' financial reasoning — ctgt-inc · 2026-10-29
• Generic one-line feedback matches self-critiques in hindsight self-distillation, study finds — jhzyu · 2026-10-29