| --- |
| license: mit |
| language: |
| - en |
| tags: |
| - mechanistic-interpretability |
| - circuit-analysis |
| - transformers |
| - benchmark |
| pretty_name: AgenticInterpBench |
| size_categories: |
| - n<1K |
| --- |
| |
| # AgenticInterpBench |
| |
| A benchmark for evaluating language-model agents on **transformer circuit explanation**. Given an already-localized circuit, the agent must recover what each component does: a functional role tag from a 5-class taxonomy, a task-specific natural-language note, and a derived description of the overall task. |
|
|
| The benchmark has **84 semi-synthetic circuits with 163 annotated components**, plus a manually annotated real-model circuit (three-operand addition in Llama-3-8B). It was introduced in *Can Language Model Agents be Helpful Circuit Explainers in Mechanistic Interpretability?* and is used to evaluate the HyVE (Hypothesize, Validate, Explain) agent framework. |
| |
| ## Contents |
| |
| ```text |
| annotations/case_{id}.json # 84 (task specifications, I/O examples, gold component roles) |
| real_circuits/llama3_abc_annotations.json # The AF1 Circuit / Llama-3-8B reference annotation (10 components) |
| ``` |
| |
| Each `annotations/case_{id}.json` carries the task summary, up to five input–output examples, the originating RASP program, the target model, and the localized components. Every component has an `id`, a TransformerLens `hook`, and a gold `role` (`tag` + task-specific `note`), following the running `frac_prevs` example: |
| |
| ```json |
| [ |
| { |
| "id": "L0_MLP", |
| "hook": "blocks.0.mlp.hook_post", |
| "role": { |
| "tag": "INDICATOR", |
| "note": "Computes per-position feature indicating whether the token at that position is 'x' or not." |
| }, |
| "labels": ["is_x_3"] |
| }, |
| { |
| "id": "L1H2_ATTN", |
| "hook": "blocks.1.attn.hook_result[2]", |
| "role": { |
| "tag": "AGGREGATOR", |
| "note": "Aggregates prefix fraction by attending over previous positions." |
| }, |
| "labels": ["frac_prevs_1"] |
| } |
| ] |
| ``` |
| |
| `real_circuits/llama3_abc_annotations.json` holds reference roles for the 10-component AF1 circuit (Mamidanna et al., 2025), with per-component notes, and supporting findings. |
| |
| ## Role taxonomy |
| |
| | Tag | Description | |
| |---|---| |
| | INDICATOR | Detects a property of the current token and emits a binary/predicate signal. | |
| | AGGREGATOR | Summarizes selected positions into a count, fraction, or accumulated quantity. | |
| | ROUTER | Moves content between positions via positional or index-based selection. | |
| | MAPPER | Transforms each position independently into a non-binary output. | |
| | COMBINER | Fuses two or more upstream signals into one output. | |
| |
| ## Statistics |
| | | | |
| |---|---| |
| | Circuits | 84 | |
| | Annotated components | 163 | |
| | MLP / attention components | 120 / 43 | |
| | Components per circuit (avg / min / max) | 1.94 / 1 / 10 | |
| | Tags (MAPPER / COMBINER / AGGREGATOR / INDICATOR / ROUTER) | 72 / 33 / 32 / 15 / 11 | |
| |
| ## Provenance & license |
| |
| The annotation layer here is the authors' own work, released under the **MIT License**. The synthetic tasks and circuits derive from [InterpBench](https://huggingface.co/cybershiptrooper/InterpBench) (Gupta et al., 2025), which retrains Tracr-compiled RASP programs; the two IOI tasks are excluded. Model weights live in the InterpBench repo and are loaded separately at runtime by HyVE. |
| |
| ## Citation |
| |
| ```bibtex |
| @misc{khan2026languagemodelagentshelpful, |
| title={Can Language Model Agents be Helpful Circuit Explainers in Mechanistic Interpretability?}, |
| author={Ayan Antik Khan and Harsh Kohli and Yuekun Yao and Huan Sun and Ziyu Yao}, |
| year={2026}, |
| eprint={2606.24026}, |
| archivePrefix={arXiv}, |
| primaryClass={cs.AI}, |
| url={https://arxiv.org/abs/2606.24026}, |
| } |
| ``` |
