Pith Number

pith:BOLR37ZK

pith:2026:BOLR37ZKRVR2XXZQZD44HGFVZZ

not attested not anchored not stored refs resolved

Finding Interpretable Prompt-Specific Circuits in Language Models

Azalea Rohr, Gabriel Franco, Lucas M. Tassis, Mark Crovella

ACC++ extracts causal attention signals from language models in a single forward pass, revealing many are interpretable via natural language.

arxiv:2602.13483 v2 · 2026-02-13 · cs.LG · cs.AI

Open paper page JSON Open Graph Bundle Merged state Verified badge What is a Pith Number?

Add to your LaTeX paper

\usepackage{pith}
\pithnumber{BOLR37ZKRVR2XXZQZD44HGFVZZ}

Prints a linked badge after your title and injects PDF metadata. Compiles on arXiv. Learn more · Embed verified badge

Record completeness

1 Bitcoin timestamp

2 Internet Archive

3 Author claim open · sign in to claim

4 Citations open

5 Replications open

✓ Portable graph bundle live · download bundle · merged state

The bundle contains the canonical record plus signed events. A mirror can host it anywhere and recompute the same current state with the deterministic merge algorithm.

Claims

C1strongest claim

ACC++ extracts circuits from a single forward pass, without replacement models or patching. Circuits identified by ACC++ consist of components that are causal for the model's attention decisions, together with the low-dimensional signals used to communicate between them. Across multiple models, a substantial portion of ACC++ signals are interpretable: many signals admit a short natural-language description.

C2weakest assumption

That the low-dimensional subspaces identified by the attention-causal communication principle are in fact the causal signals driving attention decisions, and that the natural-language descriptions assigned to them reflect genuine model mechanisms rather than post-hoc interpretations.

C3one line summary

ACC++ traces prompt-specific circuits in language models from one forward pass by extracting interpretable low-dimensional causal signals, revealing clustered mechanisms for indirect object identification and language-specific signals in multilingual settings.

References

18 extracted · 18 resolved · 0 Pith anchors

[1] doi:10.48550/ARXIV.2307 2025 · doi:10.48550/arxiv.2307

[2] Set up the counterfactual.Choose an attention head and a destination–source pair, then decide whether we are searching fordestinationsignals (in the destination token) orsourcesignals (distributed acr

[3] Enumerate candidate signals.Decompose the residual stream into outputs of upstream components, and project each component’s contribution onto the head’s singular-vector directions to form a set of can

[4] Build a contribution table.Convert candidates into a contribution matrix whose rows correspond to candidate signals and whose columns correspond to source positions in the destination row of attention

[5] Score candidates with attribution.Use Integrated Gradients to assign each candidate a fixed importance score for the attention weight on the chosen source token, while accounting for Softmax competiti

Cited by

1 paper in Pith

Data-driven Circuit Discovery for Interpretability of Language Models

Receipt and verification

First computed	2026-05-17T23:39:16.171443Z
Builder	pith-number-builder-2026-05-17-v1
Signature	Pith Ed25519 (`pith-v1-2026-05`) · public key
Schema	pith-number/v1.0

Canonical hash

0b971dff2a8d63abdf30c8f9c398b5ce5a61795ae7e3d5587651a8476c0f9a29

Aliases

arxiv: 2602.13483 · arxiv_version: 2602.13483v2 · doi: 10.48550/arxiv.2602.13483 · pith_short_12: BOLR37ZKRVR2 · pith_short_16: BOLR37ZKRVR2XXZQ · pith_short_8: BOLR37ZK

Agent API

Resolver JSON Graph JSON Events JSON Schema Signing key

Verify this Pith Number yourself

curl -sH 'Accept: application/ld+json' https://pith.science/pith/BOLR37ZKRVR2XXZQZD44HGFVZZ \
  | jq -c '.canonical_record' \
  | python3 -c "import sys,json,hashlib; b=json.dumps(json.loads(sys.stdin.read()), sort_keys=True, separators=(',',':'), ensure_ascii=False).encode(); print(hashlib.sha256(b).hexdigest())"
# expect: 0b971dff2a8d63abdf30c8f9c398b5ce5a61795ae7e3d5587651a8476c0f9a29

Canonical record JSON

{
  "metadata": {
    "abstract_canon_sha256": "29dc341a63e705f5542be6f9b1beb90f833899d66d98bae03ca3acaa88ddf930",
    "cross_cats_sorted": [
      "cs.AI"
    ],
    "license": "http://creativecommons.org/licenses/by/4.0/",
    "primary_cat": "cs.LG",
    "submitted_at": "2026-02-13T21:41:17Z",
    "title_canon_sha256": "86d076f178dc5ed41fb42c4f95499884d4ed051998d7844be5f96d493f8bee6b"
  },
  "schema_version": "1.0",
  "source": {
    "id": "2602.13483",
    "kind": "arxiv",
    "version": 2
  }
}