代理式AI在罕見症狀上出現悖論性失效

🔑 增強重點摘要

• •Optimization instability in autonomous agentic workflows causes performance oscillation that worsens with class imbalance, such as 3% prevalence for Long COVID brain fog, leading to wild sensitivity swings between 1.0 and 0.0 in the Pythia framework[1].
• •Guiding agents intended to monitor and redirect optimization paradoxically amplify overfitting and instability, failing to improve generalization on low-prevalence symptoms[1].
• •Selector agents that retrospectively select the best iteration outperform guiding agents and expert lexicons, achieving a 331% F1 score gain on brain fog detection[1].
• •This instability represents a key failure mode in agentic AI systems, exacerbated by sparse positive signals in imbalanced datasets, with broader implications for clinical NLP and autonomous systems[1].
• •Mitigating such issues requires strategies like retrospective selection over active intervention, alongside general agentic AI challenges including explainability, bias, and unintended behaviors[1][2][4].

🛠️ 技術深入

• •Pythia framework uses the target LLM for all optimization operations, ensuring intrinsic compatibility and full interpretability via interpretable error analysis[1].
• •Guiding agent intervention: Monitors performance post-iteration; pauses and redirects if no improvement, but leads to aggressive exploitation of development sets[1].
• •Selector agent: Passively identifies optimal iteration post-hoc, stabilizing performance without active guidance[1].
• •Tested on clinical symptoms with varying prevalence: shortness of breath (23%), chest pain (12%), Long COVID brain fog (3%), revealing prevalence-dependent instability[1].
• •Central failure mode: Oscillation between overcorrection and collapse due to sparse positives amplifying noise in self-optimization loops[1].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線