Skill Path: Unveiling Language Skills from Circuit Graphs

2024-10-02

AI Summary

Existing circuit discovery methods suffer from two key limitations: atomic ablation disrupts causal dependencies among connected components and fails to adequately control for confounding effects, leading to inaccurate skill attribution. To address this, we propose “skill paths”—a linear, disentangled representation of language skills that enables precise isolation of individual skills from circuit graphs. We introduce a three-stage extraction framework—decomposition, pruning, and causal mediation—that achieves the first full-chain linear decomposition of Transformers and enables hierarchical skill validation. Leveraging counterfactual reasoning and causal mediation estimation, we successfully extract three fundamental skill paths—preceding-token prediction, inductive reasoning, and in-context learning—across multiple models. Empirical results demonstrate their hierarchical structure and compositional inclusivity, significantly improving both the accuracy and interpretability of skill attribution.

Technology Category

• Machine Learning: Causal Learning
• Knowledge Representation and Reasoning: Diagnosis and Abductive Reasoning
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data

Abstract

Circuit graph discovery has emerged as a fundamental approach to elucidating the skill mechanistic of language models. Despite the output faithfulness of circuit graphs, they suffer from atomic ablation, which causes the loss of causal dependencies between connected components. In addition, their discovery process, designed to preserve output faithfulness, inadvertently captures extraneous effects other than an isolated target skill. To alleviate these challenges, we introduce skill paths, which offers a more refined and compact representation by isolating individual skills within a linear chain of components. To enable skill path extracting from circuit graphs, we propose a three-step framework, consisting of decomposition, pruning, and post-pruning causal mediation. In particular, we offer a complete linear decomposition of the transformer model which leads to a disentangled computation graph. After pruning, we further adopt causal analysis techniques, including counterfactuals and interventions, to extract the final skill paths from the circuit graph. To underscore the significance of skill paths, we investigate three generic language skills-Previous Token Skill, Induction Skill, and In-Context Learning Skill-using our framework. Experiments support two crucial properties of these skills, namely stratification and inclusiveness.

Problem

• Circuit graphs lose causal dependencies during atomic ablation
• Circuit graphs capture extraneous effects beyond target skills
• Skill paths isolate individual skills through linear component chains

Innovation

• Skill paths isolate skills in linear chains
• Three-step framework decomposes and prunes circuits
• Causal mediation extracts paths using counterfactuals interventions