Self‑evolving training improves agent performance on difficult tasks via verifiable feedback, yet it usually applies costly deliberation uniformly across tasks. To address this, we propose Mode‑Optimized Reasoning with Self‑Adaptive Effort (MORE), a self‑evolving training framework that adaptively selects both reasoning form and computational effort per task. MORE follows a two‑stage paradigm.
Stage 1 builds a compact prior over direct answering, short chain‑of‑thought, long chain‑of‑thought, and code‑based reasoning.
Stage 2 self‑evolves this prior on generated, verifiable task streams using Mode‑Adaptive and Test‑time Effort PPO (MATE‑PPO) without curated data. By jointly learning mode selection and token expenditure, MORE allocates concise reasoning to tasks that allow it while retaining extended deliberation where beneficial.
Across coding, mathematics, and general‑reasoning benchmarks, MORE consistently improves Base and Coder backbones from 3B to 14B parameters. During inference it yields higher accuracy and better efficiency, with markedly shorter responses and reduced inference time.
Review