超越行为权衡：在LLM中疼痛-愉悦决策的机制追溯
Beyond Behavioural Trade-Offs: Mechanistic Tracing of Pain-Pleasure Decisions in an LLM
AI总结：
研究揭示了LLM在疼痛-愉悦决策中的内部机制，通过机制追溯揭示了价值信号的表示和因果作用，为AI意识和福利的讨论提供了证据基础。
摘要：
先前的行为研究表明，某些大语言模型在选项被描述为导致痛苦或愉悦时会改变选择，并且这种偏差可以随着声明的强度而扩大。为了将行为证据（模型的行为）与机制可解释性（支持其的计算）联系起来，我们研究了与价值相关的信息是如何被表示并在变压器内部被因果使用。使用Gemma-2-9B-it和一个受先前工作启发的最小化决策任务模型，我们（i）通过跨流的逐层线性探测映射表示可用性，（ii）通过激活干预（引导；修补/删除）测试因果贡献，（iii）量化在epsilon网格上的剂量-反应效应，读出2-3个logit边距和数字对归一化的选择概率。我们发现（a）价值符号（痛苦 vs. 娱乐）在非常早期的层（L0-L1）中可以完美线性分离，而一个词典基线保留了显著的信号；（b）渐进强度强烈可解码，峰值出现在中后期层，尤其是注意力/MLP输出中，决策对齐在最终标记之前稍早达到最高；（c）沿着数据衍生的价值方向的加法引导在晚期站点因果调节2-3个logit边距，最大的效果出现在晚期层注意力输出（attn_out L14）中；（d）头部级别的修补/删除表明这些效果分布在多个头部而不是集中在单一单元中。这些结果将行为敏感性与可识别的内部表示和干预敏感的站点联系起来，为更严格的情境测试和更广泛复制提供了具体的机制目标。这项工作支持更基于证据的（a）关于人工智能意识和福利的辩论，以及（b）在制定政策、审计标准和安全保障时的治理。
英文摘要：
Prior behavioural work suggests that some LLMs alter choices when options are framed as causing pain or pleasure, and that such deviations can scale with stated intensity. To bridge behavioural evidence (what the model does) with mechanistic interpretability (what computations support it), we investigate how valence-related information is represented and where it is causally used inside a transformer. Using Gemma-2-9B-it and a minimalist decision task modelled on prior work, we (i) map representational availability with layer-wise linear probing across streams, (ii) test causal contribution with activation interventions (steering; patching/ablation), and (iii) quantify dose-response effects over an epsilon grid, reading out both the 2-3 logit margin and digit-pair-normalised choice probabilities. We find that (a) valence sign (pain vs. pleasure) is perfectly linearly separable across stream families from very early layers (L0-L1), while a lexical baseline retains substantial signal; (b) graded intensity is strongly decodable, with peaks in mid-to-late layers and especially in attention/MLP outputs, and decision alignment is highest slightly before the final token; (c) additive steering along a data-derived valence direction causally modulates the 2-3 margin at late sites, with the largest effects observed in late-layer attention outputs (attn_out L14); and (d) head-level patching/ablation suggests that these effects are distributed across multiple heads rather than concentrated in a single unit. Together, these results link behavioural sensitivity to identifiable internal representations and intervention-sensitive sites, providing concrete mechanistic targets for more stringent counterfactual tests and broader replication. This work supports a more evidence-driven (a) debate on AI sentience and welfare, and (b) governance when setting policy, auditing standards, and safety safeguards.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
