Conditional Co-Ablation: Recovering Self-Repair Backups in Transformer Circuits

2026-07-02

AI Summary

This work addresses the failure of conventional single-component ablation-based attribution methods in Transformers, which overlook backup pathways due to the model’s self-repair mechanisms, often misclassifying critical components as irrelevant. To overcome this limitation, the authors propose Conditional Collaborative Ablation (CoAx), the first approach that formalizes self-repair as a conditional circuit completion task. CoAx quantifies the conditional increase in ablation effects across remaining units after removing a primary component, thereby uncovering latent second-order interactions and redundant pathways. The method operates in an unsupervised manner—requiring no labels and relying solely on model outputs—and integrates counterfactual patching, structured pruning, and cross-model transfer for precise attribution. Evaluated on the IOI task with GPT-2-small, CoAx improves AUC for backup head identification from 0.33 to 0.91, substantially outperforming baselines, and successfully generalizes across eight distinct models, enabling capability knockout and scalable pruning.

Abstract

Mechanistic interpretability often relies on component-level interventions to discover how a model produces a behavior. This guides attribution, capability knockout, and model pruning downstream to operate by scoring each unit by the effect of ablation in isolation. Such first-order scoring is natural when component importance is additive, but becomes misleading when a transformer self-repairs: after a primary component is removed, a dormant backup can take over, muting the primary's measured effect while the backup itself appears irrelevant on the intact model. We recast this failure as a recovery task, conditional circuit completion, and introduce Conditional Co-Ablation (CoAx), a label-free, output-grounded score that asks how much each remaining unit's ablation effect grows once a primary set has been removed. This conditional growth exposes the second-order interaction that single-unit scores discard. On the GPT-2-small IOI circuit, CoAx raises backup-head recovery from 0.33 to 0.91 ROC-AUC, outperforming all baselines, including self-repair-aware gradient scores (best 0.82); counterfactual patching verifies that the recovered heads causally carry the repair. The same label-free procedure transfers to induction across eight models. Beyond discovery, the recovered backups correct self-repair-masked attribution, identify the components required for capability knockout, and yield repair-aware structured pruning scaling from 124M to 7B. Component importance is therefore not merely an isolated-unit property: in robust circuits, the components that matter can become visible only under the interventions that make them necessary.

Problem

• self-repair
• transformer circuits
• mechanistic interpretability
• component ablation
• backup recovery

Innovation

• Conditional Co-Ablation
• self-repair
• mechanistic interpretability
• transformer circuits
• structured pruning