用因果图揭示大模型内部概念如何一步步相互作用。
Causal Concept Graphs in LLM Latent Space for Stepwise Reasoning
- 构建概念因果图，捕捉语言模型中概念间的因果依赖关系。
- 在多个推理任务上，因果图引导的干预效果显著优于随机方法。
- 图结构稀疏稳定，适合用于解释大模型的逐步推理过程。
稀疏自编码器能定位语言模型中的概念位置，但无法揭示多步推理中概念间的交互方式。本文提出因果概念图（CCG）：一种基于稀疏可解释潜变量的有向无环图，边表示概念间的学习到的因果依赖。结合任务条件稀疏自编码器进行概念发现与基于DAGMA的可微结构学习实现图恢复，并引入因果保真度评分（CFS）评估图引导干预是否比随机干预产生更大下游影响。在ARC-Challenge、StrategyQA和LogiQA三个数据集上，使用GPT-2 Medium模型，五次种子实验（共15组配对运行），CCG获得CFS=5.654±0.625，显著优于ROME式追踪（3.382±0.233）、仅用SAE排序（2.479±0.196）和随机基线（1.032±0.034），校正后p<0.0001。学习到的图结构稀疏（5-6%边密度），具有领域特异性且跨种子稳定。
原文摘要 · Abstract (English)
Sparse autoencoders can localize where concepts live in language models, but not how they interact during multi-step reasoning. We propose Causal Concept Graphs (CCG): a directed acyclic graph over sparse, interpretable latent features, where edges capture learned causal dependencies between concepts. We combine task-conditioned sparse autoencoders for concept discovery with DAGMA-style differentiable structure learning for graph recovery and introduce the Causal Fidelity Score (CFS) to evaluate whether graph-guided interventions induce larger downstream effects than random ones. On ARC-Challenge, StrategyQA, and LogiQA with GPT-2 Medium, across five seeds ($n{=}15$ paired runs), CCG achieves $\CFS=5.654\pm0.625$, outperforming ROME-style tracing ($3.382\pm0.233$), SAE-only ranking ($2.479\pm0.196$), and a random baseline ($1.032\pm0.034$), with $p<0.0001$ after Bonferroni correction. Learned graphs are sparse (5-6\% edge density), domain-specific, and stable across seeds.
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。