Self evolving agents learn to reason efficiently and adaptively
More: Mode-Optimized Reasoning with Self-Adaptive Effort for Self-Evolving Agents
Artificial Intelligence
Summary
Some AI agents can think through problems in many ways but often spend a lot of time doing so for every task, which can be slow and costly. The authors propose a new method called MORE that helps these agents choose the best way to think and how hard to work on each problem. It learns to decide whether to give a quick answer or use a longer reasoning process, saving time while improving accuracy. This approach works well on tasks involving coding, math, and general problem solving with various model sizes.
What this means in practice
- •For software development teams: Optimize AI coding assistants to tailor reasoning effort and techniques to each coding task for faster and more accurate suggestions.
- •For automated math problem solvers: Improve math-solving AI by dynamically adjusting reasoning steps to balance speed and correctness across diverse problem types.
Authors
Dianxing Shi, Dingjie Fu, Tianyu Lei, Wangjie Qiu
Abstract
Self-evolving training improves the performance of agents on difficult tasks through verifiable feedback, but it often applies costly deliberation uniformly across tasks. To address this problem, we propose Mode-Optimized Reasoning with Self-Adaptive Effort (MORE), a self-evolving training framework that adaptively selects both reasoning form and computational effort for each task. Specifically, MORE follows a two-stage training paradigm: (i) \emph{Stage 1} establishes a compact prior over direct answering, short and long chain-of-thought, and code-based reasoning, and (ii) \emph{Stage 2} self-evolves this prior with Mode-Adaptive and Test-time Effort PPO (MATE-PPO) on generated, verifiable task streams without curated data. By learning mode selection and token expenditure jointly, MORE allocates concise reasoning to tasks that permit it while retaining extended deliberation where it is useful. Across coding, mathematics, and general-reasoning benchmarks, MORE achieves consistent performance gains on Base and Coder backbones at model scales from 3B to 14B. During inference, MORE demonstrates higher accuracy and better efficiency, with markedly shorter response lengths and reduced inference time.