Discovers minimal causal subgraphs in transformer models via ablation and gradient attribution.
npx skills add https://github.com/zjunlp/Mechanist --skill circuit-zjunlp
Please help me install this Agent Skill. Skill: Circuit Source: https://github.com/zjunlp/Mechanist/tree/main/skills/mechanism-skills/circuit-discovery Command: npx skills add https://github.com/zjunlp/Mechanist --skill circuit-zjunlp
High-intent search queries and answers about installing and using this skill.
Use either exact iterative search (ACDC), which prunes edges by activation patching with corrupted inputs, or fast gradient-based attribution (EAP-IG), which scores all edges in a few passes and thresholds the scores. Both produce a minimal subgraph evaluated for faithfulness on a task distribution.
ACDC performs exact iterative edge pruning via per-edge activation patching, giving causal, intervention-grounded circuits at high compute cost. EAP-IG approximates edge effects with gradients and integrated gradients, trading some exactness for dramatically lower cost.
Yes. The EAP-IG demo builds a computational graph directly from a TransformerLens HookedTransformer such as gpt2-small using Graph.from_model, then attributes, selects a top-n circuit, and evaluates it. ACDC similarly targets TransformerLens-style hooked models.
Ablation-based search scales poorly with model and graph size, while gradient-based approximations can miss edges with non-linear effects. Results also depend on the task distribution, corrupted inputs, and faithfulness metric, so different configurations can yield different circuits.
Faithfulness is measured by ablating components outside the recovered circuit and checking whether the task metric is preserved on a held-out distribution. The EAP-IG evaluate_graph function reports metric values and the drop from baseline under interventions such as zero or mean ablation.
handsomelong922Guide mechanistic interpretability research on transformer models using TransformerLens.
arsityInspect transformer activations, attention, and circuits with TransformerLens.
choice5346Trains and analyzes sparse-coded neural network features using SAELens and TransformerLens workflows.
BordaAutomate end-to-end ablation studies on completed ML research runs.
