Research map / Architectures and efficiency / RNNs and normalization

EAP-GP: Mitigating Saturation Effect in Gradient-based Automated Circuit Identification

NeurIPS 2025 · 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.

Proposes EAP-GP, a gradient-based circuit discovery method that builds an adaptive integration path from clean to corrupted activations to avoid gradient saturation, improving circuit faithfulness over EAP and EAP-IG on GPT-2 models.

Problem

Gradient-based edge attribution for automated circuit discovery in transformer LMs (EAP, EAP-IG) suffers from the zero-gradient problem and saturation effects along fixed straight-line integration paths, yielding unreliable edge attributions and unfaithful circuits.

Approach

Replace EAP-IG's fixed straight-line integration path with GradPath, a model-dependent path built by iterative gradient-descent steps on the squared output distance to the corrupted activation, normalized by gradient norm, so that the path avoids saturated regions where gradients vanish.

Key results

• At 97.5% edge sparsity on GPT-2 Small, EAP-GP reaches 80.1% NFS on IOI versus 62.4% for EAP-IG and 56.9% for EAP.
• On Greater-Than at 97.5% sparsity, EAP-GP reaches 99.8% NFS versus 97.6% for EAP-IG and 96.3% for EAP (Table 1).
• EAP-GP is roughly five times slower than EAP-IG (232.5 s vs 49.7 s on IOI at 97.5% sparsity).

Measured effects reported in the paper

Surprising findings

• EAP generates fully unfaithful (score 0) circuits at 99.0-99.9% sparsity on IOI due to parentless heads being pruned.
• Increasing the number of GradPath steps beyond about 5, or reducing sparsity, can decrease faithfulness, attributed to amplified high-frequency oscillations in the path.

Limitations

• Runtime is roughly five times that of EAP-IG due to the extra forward and backward passes for constructing the path.
• Gains on Greater-Than and Gender-Bias are small, and manual-circuit precision/recall results are mixed, with EAP-GP initially worse on IOI.
• Faithfulness can decline with more steps or lower sparsity, possibly from selecting noise-encoding edges.

Experimental setup

Datasets: Indirect Object Identification (IOI) Subject-Verb Agreement (SVA) Gender-Bias Capital-Country Hypernymy Greater-Than

Benchmarks: IOI SVA Gender-Bias Capital-Country Hypernymy Greater-Than

Metrics: Normalized Faithfulness Score (NFS) runtime (s) precision-recall vs manual circuits

Baselines: EAP EAP-IG ACDC

Builds on: Edge Attribution Patching (EAP) EAP-IG Integrated Gradients Automated Circuit Discovery (ACDC)

Closest papers by meaning

• Finding Transformer Circuits With Edge Pruning (NeurIPS 2024 · similarity 0.88)
• Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition (ICLR 2025 · similarity 0.88)
• Transcoders find interpretable LLM feature circuits (NeurIPS 2024 · similarity 0.86)
• The Computational Complexity of Circuit Discovery for Inner Interpretability (ICLR 2025 · similarity 0.85)
• DeepGate4: Efficient and Effective Representation Learning for Circuit Design at Scale (ICLR 2025 · similarity 0.85)
• Circuit Representation Learning with Masked Gate Modeling and Verilog-AIG Alignment (ICLR 2025 · similarity 0.85)