Execution feedback enables coding agents to rewrite programs and learn from their own corrections. The learning weight of a correction should reflect both the number of transitions supported by its executions and the amount of evidence behind that support. We introduce Effective‑Evidence Self‑Distillation (EESD), which models these two quantities separately. Normalized execution relevance determines relative transition support and yields an effective pseudo‑count mass; a Dirichlet posterior then provides an uncertainty‑penalized, KL‑anchored correction learning weight. Under a symmetric prior, changing the mass preserves category ordering, and the effective mass produces a supervised coefficient bounded by its matched fixed‑mass counterpart. Experiments span four model‑domain history sweeps; increasing visible observations from one to eight reduces future‑outcome NLL by 55.0‑59.3%. At eight observations, effective mass achieves lower NLL than fixed mass in all comparisons. In the primary DeepSeek/RunBugRun matched study, argmax predictions agree on all 3,000 examples, with the largest NLL gain under concentrated relevance. After one correction‑learning round, DeepSeek/CodeARC Pass@1 rises from 15.0% to 20.4%, with a paired 95% bootstrap interval of [+2.8, +8.0] percentage points. A twelve‑setting downstream evaluation confirms the model‑domain scope of this update. These findings show that separating evidence support from evidence mass changes probability estimation and improves correction learning in coding agents.
Review