SkillAdam:让 AI 智能体技能稳定自进化

DAIR.AI · @dair_ai · X·2026-09-11 05:00·46分钟前
AI 导读

SkillAdam 将 Adam 的两个动量估计迁移到离散、不可微的技能文档上,解决技能自进化循环中的方向不稳定与更新幅度固定两大失效模式。它用优化记忆记录已识别问题及历史尝试结果以稳定更新方向,用波动驱动的编辑预算控制每次修订幅度。在覆盖短程与长程任务的七个基准上达到 SOTA,优化动态更稳定,迭代次数和成本均显著低于此前方法。

DAIR.AI@dair_ai
43AI 编辑部评分,满分 100

SkillAdam:让 AI 智能体技能稳定自进化

2026-09-11 05:00· 46分钟前
AI 导读

SkillAdam 将 Adam 的两个动量估计迁移到离散、不可微的技能文档上,解决技能自进化循环中的方向不稳定与更新幅度固定两大失效模式。它用优化记忆记录已识别问题及历史尝试结果以稳定更新方向,用波动驱动的编辑预算控制每次修订幅度。在覆盖短程与长程任务的七个基准上达到 SOTA,优化动态更稳定,迭代次数和成本均显著低于此前方法。

Another interesting approach to self-evolve agent skills.

But it's important to know that skill self-evolution loops fail in two specific ways:

  1. Direction instability. Effective corrections get overwritten by iteration-local feedback instead of accumulating, so the loop keeps undoing its own fixes.
  1. Fixed update scope. Every revision changes about the same amount regardless of whether recent case-level improvements were consistent or noisy.

SkillAdam addresses both by porting Adam's two moment estimates to discrete, non-differentiable skill documents.

As a functional analogue of the first moment, an optimization memory records identified problems and the outcomes of prior solution attempts, which stabilizes the update direction. As an analogue of the second moment, a volatility-driven edit budget tracks the history-weighted variation of recent case-level improvements and controls how large each revision is allowed to be.

Across seven benchmarks spanning short and long-horizon tasks it reaches state of the art with more stable optimization dynamics, and it gets there in substantially fewer iterations and at lower cost than prior methods.

Paper: https://academy.dair.ai/papers/skilladam-stable-and-efficient-skill-evolution-for-agents-2609.08944

来源:DAIR.AI· x.com