Repeated Sampling Beats Self‑Refine and Reflexion Across 1.5B‑7B LMs
A systematic study of 1.5 B, 3 B and 7 B open models shows that every self‑refine, reflexion, or debate method underperforms a baseline that repeatedly samples the same question and picks the most common answer when token budgets are equal. The result calls into question the value of model‑self‑critique for accuracy gains.