Scholarlink

<p>As foundation models evolve from disembodied inference to embodied intelligence, their capacity for self-evolution introduces new risks to cognitive stability. This study examines how autonomous adaptation in embodied foundation models can undermine internal coherence and mission reliability. Using the OODA loop to structure the cognition-action cycle, we identify functional evolution boundaries that distinguish adaptive from unsafe learning. We formalize key cognitive degradation mechanisms, including representation drift, reward misalignment, and coupling amplification. We further analyze how such degradation propagates across perception, reasoning, and control. Building on this analysis, we propose the Trusted Evolution Framework (TEF), a governance architecture for bounded self-evolution. TEF introduces formal stability criteria and quantitative indicators to regulate adaptive dynamics and maintain cognitive coherence. Theoretical validation and application pathways demonstrate that TEF enables adaptive agents to evolve safely while preserving flexibility and trust. This work provides a theoretical foundation for cognitive robustness in self-evolving systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

The trusted evolution framework: governing cognitive stability in self-evolving foundation models

  • 期刊论文

  • 发表日期：

    2026年2月5日

查看全文
  • Yuwei Chen,
  • Shiyong Chu

摘要

As foundation models evolve from disembodied inference to embodied intelligence, their capacity for self-evolution introduces new risks to cognitive stability. This study examines how autonomous adaptation in embodied foundation models can undermine internal coherence and mission reliability. Using the OODA loop to structure the cognition-action cycle, we identify functional evolution boundaries that distinguish adaptive from unsafe learning. We formalize key cognitive degradation mechanisms, including representation drift, reward misalignment, and coupling amplification. We further analyze how such degradation propagates across perception, reasoning, and control. Building on this analysis, we propose the Trusted Evolution Framework (TEF), a governance architecture for bounded self-evolution. TEF introduces formal stability criteria and quantitative indicators to regulate adaptive dynamics and maintain cognitive coherence. Theoretical validation and application pathways demonstrate that TEF enables adaptive agents to evolve safely while preserving flexibility and trust. This work provides a theoretical foundation for cognitive robustness in self-evolving systems.