信念還是電路？脈絡內圖形學習的因果證據

🔑 增強重點摘要

• •The study challenges the 'Bayesian inference' vs 'pattern matching' dichotomy by demonstrating that LLMs utilize a hybrid mechanism where local transition probabilities and global graph structure are processed in parallel.
• •The research identifies that the model's internal representation of graph topology is not monolithic but is decomposed into distinct, linearly separable subspaces within the residual stream.
• •The causal patching experiments suggest that the model's reliance on specific graph features can be dynamically modulated, indicating that in-context learning is a steerable process rather than a static retrieval of pre-trained knowledge.

🛠️ 技術深入

• •Task Design: Utilized synthetic random-walk sequences on graphs with varying edge probabilities to isolate the model's ability to infer global connectivity versus local sequence prediction.
• •PCA Methodology: Applied Principal Component Analysis to the residual stream activations at specific layers to identify the dimensionality of the graph-topology encoding.
• •Causal Patching: Implemented activation patching by replacing activations from a 'clean' graph sequence with those from a 'corrupted' sequence to measure the causal effect on the next-token prediction.
• •Steering Mechanism: Employed activation steering by adding learned vectors to the residual stream to bias the model toward specific graph-path interpretations without modifying model weights.

🔮 前景展望AI analysis grounded in cited sources