Research map / Architectures and efficiency / RNNs and normalization
IBCircuit: Towards Holistic Circuit Discovery with Information Bottleneck
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 IBCircuit, an information-bottleneck framework that learns sigmoid-gated Gaussian noise weights on attention heads and residual edges to find holistic circuits in GPT-2 without hand-designed corrupted activations.
Problem
Circuit discovery in language models usually needs task-specific corrupted activations and treats components independently, which is inaccurate and inefficient; the paper seeks a holistic, task-agnostic way to find minimal faithful circuits.
Approach
Cast circuit discovery as an information bottleneck: learnable IB weights gate Gaussian noise injected into node (attention head) and edge activations, trained end-to-end with a KL-to-original-output term plus a KL-based mutual information regularizer; components with high retained weight form the circuit, using the model's own outputs as supervision.
Key results
- On IOI, IBCircuit recovers the canonical circuit with a better TPR/FPR trade-off than SP, ACDC and AP, but underperforms ACDC on Greater-Than.
- For edge circuits on IOI with GPT-2 small, IBCircuit achieves higher Logit Difference and lower KL Divergence than ACDC and EAP as edge count falls.
- On GPT-2 XL (IOI), IBCircuit is comparable to EAP and better in Logit Difference at 30k-70k edges (about 97%-98.5% of 2,235,025 edges).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove MI loss (ablation) | IBCircuit · IOI | Logit Difference | Hurt |
| remove KL loss (ablation) | IBCircuit · IOI | Logit Difference | Hurt |
| increase trade-off coefficient α (ablation) | IBCircuit · IOI | average λ | Hurt |
| increase trade-off coefficient α (ablation) | IBCircuit · IOI | KL Divergence | Hurt |
| replace Gaussian perturbations with hard-concrete masking (ablation) | IBCircuit · Greater-Than | Greater Probability | Mixed |
| replace MI regularization with the SP objective (ablation) | IBCircuit · Greater-Than | Greater Probability | Hurt |
| replace MI regularization with the SP objective (ablation) | IBCircuit · Greater-Than | KL Divergence | Hurt |
| use IBCircuit instead of baseline node-selection methods | GPT-2 circuit discovery · IOI | Logit Difference | Helped |
| use IBCircuit instead of ACDC | GPT-2 circuit discovery · Greater-Than | Greater Probability | Mixed |
| use IBCircuit instead of ACDC and EAP | GPT-2 edge circuit discovery · IOI | Logit Difference | Helped |
| use IBCircuit instead of ACDC and EAP | GPT-2 edge circuit discovery · IOI | KL Divergence | Helped |
| use IBCircuit instead of EAP | GPT-2 XL edge circuit discovery · IOI | Logit Difference | Helped |
Surprising findings
- IBCircuit underperforms ACDC on the Greater-Than task, which the authors attribute to the lack of a single well-defined expected output.
- At very small node counts, all methods including IBCircuit show high KL divergence, so extreme sparsity is less meaningful.
Limitations
- Underperforms ACDC on Greater-Than, especially at low node counts.
- Sparsity-faithfulness trade-off controlled by hyperparameter alpha; sparser circuits sacrifice KL divergence.
- Evaluated mainly on GPT-2 small with two tasks.
Experimental setup
Datasets: IOI task prompts Greater-Than task prompts
Benchmarks: IOI Greater-Than
Metrics: ROC/TPR-FPR vs canonical circuit Logit Difference Greater Probability KL Divergence
Baselines: Subnetwork Probing ACDC Attribution Patching Edge Attribution Patching EAP-IG
Builds on: variational information bottleneck ACDC subnetwork probing attribution patching edge attribution patching
Closest papers by meaning
- The Computational Complexity of Circuit Discovery for Inner Interpretability (ICLR 2025 · similarity 0.86)
- Rethinking Circuit Completeness in Language Models: AND, OR, and ADDER Gates (NeurIPS 2025 · similarity 0.85)
- Many Circuits, One Mechanism: Input Variation and Evaluation Granularity in Circuit Discovery (NeurIPS 2026 · similarity 0.85)
- Circuit Component Reuse Across Tasks in Transformer Language Models (ICLR 2024 · similarity 0.84)
- Efficient Automated Circuit Discovery in Transformers using Contextual Decomposition (ICLR 2025 · similarity 0.84)
- Finding Transformer Circuits With Edge Pruning (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.