Research map / Large language models / Language, safety and interpretability
Query Circuits: Explaining How Language Models Answer User Prompts
Tung-Yu Wu, Fazl Barez
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Best-of-N paraphrase sampling for circuit discovery instead of single-query EAP-IG | EAP-IG query circuits · MMLU (Llama-3.2-1B-Instruct) | edges needed to reach NDF 0.6: 200k -> 5k edges | Improved |
| average edge scores over the query and its paraphrases instead of single-query scoring (ablation) | EAP-IG query circuits · IOI, arithmetic, MMLU, ARC | NDF | No clear effect |
| add more paraphrases to BoN (ablation) | BoN · IOI, arithmetic multiplication, MMLU Astronomy | NDF | Improved |
| use iBoN or BoN-CSM instead of full BoN (ablation) | BoN · IOI, arithmetic, MMLU, ARC | NDF | Hurt |
| increase the integrated gradients step count (ablation) | EAP-IG · MMLU Astronomy | NDF | No clear effect |
| use query-specific circuit instead of dataset-level capability circuit | EAP-IG on GPT-2 Small · IOI (1000 edges) | fraction of performance recovered: about 65% -> less than 50% | Hurt |
| choose circuit with highest NDF instead of lowest NDF for SAE gender-feature ablation (ablation) | query circuits, GPT-2 Small · Gender Bias (32 samples), logit scale | absolute bias reduction: 0.234 -> 0.810 | Improved |
Closest papers by meaning
- Rethinking Circuit Completeness in Language Models: AND, OR, and ADDER Gates (NeurIPS 2025 · similarity 0.88)
- Towards Understanding Fine-Tuning Mechanisms of LLMs via Circuit Analysis (ICML 2025 · similarity 0.88)
- Finding Interpretable Prompt-Specific Circuits in Language Models (NeurIPS 2026 · similarity 0.87)
- Extracting Interpretable Task-Specific Circuits from Large Language Models for Faster Inference (AAAI 2025 · similarity 0.87)
- Hypothesis Testing the Circuit Hypothesis in LLMs (NeurIPS 2024 · similarity 0.87)
- Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models (ICLR 2025 · similarity 0.87)
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.
