arxivcs.CLcs.AIcs.LG2026-07-21
Verifiable Self-Evolution for Open-Ended Dialogue Skills via Future-Feedback Prediction
Textual skills provide a lightweight way to improve frozen language-model agents, but their self-evolution normally requires a stable validation signal. Such signals are natural in mathematics or code, where an answer can be checked after it changes, yet are problematic in open-e…