Research map / Architectures and efficiency / RNNs and normalization

Many Circuits, One Mechanism: Input Variation and Evaluation Granularity in Circuit Discovery

NeurIPS 2026 · Architectures and efficiency › RNNs and normalization · approach: Mechanistic circuit interpretability · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Shows that ACDC circuits extracted for Literal Sequence Copying from different token-frequency bands in Pythia models (70M–1.4B) differ structurally but are functionally interchangeable, a pattern termed "phantom specialization", and that edge-level evaluation exposes this while source-level evaluation inflates faithfulness.

Problem

Circuit discovery papers often interpret structural differences between discovered circuits as evidence of distinct mechanisms, but it is unclear whether such differences arise when only input token frequency varies with the task fixed, and whether they are functionally meaningful.

Approach

Partition vocabulary into token-frequency bands from The Pile, extract ACDC circuits per band (3 draws each) on a fixed copying task, and test structural, functional (cross-band transfer, edge-level vs source-level evaluation) and representational (interchange interventions) similarity; find structurally distinct circuits that transfer across bands and share a core recovering nearly all accuracy.

Key results

• Within-band Jaccard exceeds between-band Jaccard in all five Pythia models, but gaps are small (0.013–0.032).
• Transfer efficiency of band-specific edges ranges from 81.4% (Pythia-70m) to 97.0% (Pythia-1.4b), indicating generic rather than band-specific benefit.
• A shared core (edges in at least 3 of 5 conditions) recovers at least 99% of the accuracy gap between the universal core and full circuit for models of 160M and above.

Surprising findings

• Source-level evaluation inflates apparent circuit accuracy by up to 85 percentage points relative to edge-level evaluation.
• Band-specific edges achieve 0% accuracy in isolation and are about 1–2% stable across draws, suggesting they are extraction byproducts.

Limitations

• Transfer test has limited power for very small effects (80% power only for advantages of 0.013–0.026 accuracy points or larger).
• Circuits defined over edges between components, so token-position or feature-level specialization is not tested.
• Only one main task (LSC) with a smaller SVA replication covering only core structural and transfer analyses.

Experimental setup

Datasets: The Pile (token frequency measurement) Literal Sequence Copying synthetic sequences subject-verb agreement synthetic sentences

Benchmarks: Literal Sequence Copying (cross-band transfer) subject-verb agreement replication

Metrics: KL divergence (faithfulness) accuracy (sufficiency) ablation accuracy (necessity) edge fraction (minimality) Jaccard similarity transfer efficiency Cohen's d

Baselines: EAP EAP-IG random edge subgraphs

Builds on: ACDC AutoCircuit EAP-IG Literal Sequence Copying causal abstraction

Closest papers by meaning

• Towards Understanding Fine-Tuning Mechanisms of LLMs via Circuit Analysis (ICML 2025 · similarity 0.87)
• The Computational Complexity of Circuit Discovery for Inner Interpretability (ICLR 2025 · similarity 0.87)
• LLM Circuit Analyses Are Consistent Across Training and Scale (NeurIPS 2024 · similarity 0.87)
• Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition (ICLR 2025 · similarity 0.86)
• Hypothesis Testing the Circuit Hypothesis in LLMs (NeurIPS 2024 · similarity 0.86)
• How connectivity structure shapes rich and lazy learning in neural circuits (ICLR 2024 · similarity 0.85)