大模型能察觉自身状态变化，但能力随层数递减。
Detecting the Disturbance: A Nuanced View of Introspective Abilities in LLMs
- 用激活控制技术测试模型对内部干扰的感知能力
- 早期层可准确识别10句中哪句被篡改（88%准确率）
- 发现模型具备部分自省能力，适合研究认知机制
大型语言模型能否自省，即准确检测其内部状态的变化？我们基于Meta-Llama-3.1-8B-Instruct系统性地研究了这一问题。首先，我们发现以往研究采用的二元检测范式将自省与方法学误差混淆：看似高准确率完全由全局logit偏移导致，使模型无论问题内容如何都倾向于肯定回答。但在需要差异敏感的任务中，我们发现了部分自省的稳健证据：模型在10个句子中定位被注入的句子准确率达88%（随机为10%），并区分注入强度的准确率达83%（随机为50%）。这些能力仅存在于早期层，后续层数迅速退化至随机水平——我们通过注意力信号路由和残差流恢复机制进行了解释。结果表明，大模型能够对内部扰动计算有意义函数，自省是一种真实但依赖层级的现象，值得进一步研究。代码已开源：https://github.com/elyhahami18/llama-introspection-new
原文摘要 · Abstract (English)
Can large language models introspect, that is, accurately detect perturbations to their own internal states? We systematically investigate this question using activation steering in Meta-Llama-3.1-8B-Instruct. First, we show that the binary detection paradigm used in prior work conflates introspection with a methodological artifact: apparent detection accuracy is entirely explained by global logit shifts that bias models toward affirmative responses regardless of question content. However, on tasks requiring differential sensitivity, we find robust evidence for partial introspection: models localize which of 10 sentences received an injection at up to 88\% accuracy (vs.\ 10\% chance) and discriminate relative injection strengths at 83\% accuracy (vs.\ 50\% chance). These capabilities are confined to early-layer injections and collapse to chance thereafter -- a pattern we explain mechanistically through attention-based signal routing and residual stream recovery dynamics. Our findings demonstrate that LLMs can compute meaningful functions over perturbations to their internal states, establishing introspection as a real but layer-dependent phenomenon that merits further investigation. Our code is open-sourced here: https://github.com/elyhahami18/llama-introspection-new
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。