MétaCan
Menu
Back to cohort
Record W4415967295 · doi:10.48550/arxiv.2510.18148

Extracting Rule-based Descriptions of Attention Features in Transformers

2025· preprint· W4415967295 on OpenAlexaboutno aff
Dan Friedman, Adithya Bhaskar, Alexander Wettig, Danqi Chen

Bibliographic record

VenueArXiv.org · 2025
Typepreprint
Language
FieldComputer Science
TopicExplainable Artificial Intelligence (XAI)
Canadian institutionsnot available
Fundersnot available
KeywordsInterpretabilitySecurity tokenTransformerInterpretation (philosophy)Word (group theory)Value (mathematics)Pattern recognition (psychology)

Abstract

fetched live from OpenAlex

Mechanistic interpretability strives to explain model behavior in terms of bottom-up primitives. The leading paradigm is to express hidden states as a sparse linear combination of basis vectors, called features. However, this only identifies which text sequences (exemplars) activate which features; the actual interpretation of features requires subjective inspection of these exemplars. This paper advocates for a different solution: rule-based descriptions that match token patterns in the input and correspondingly increase or decrease the likelihood of specific output tokens. Specifically, we extract rule-based descriptions of SAE features trained on the outputs of attention layers. While prior work treats the attention layers as an opaque box, we describe how it may naturally be expressed in terms of interactions between input and output features, of which we study three types: (1) skip-gram rules of the form "[Canadian city]... speaks --> English", (2) absence rules of the form "[Montreal]... speaks -/-> English," and (3) counting rules that toggle only when the count of a word exceeds a certain value or the count of another word. Absence and counting rules are not readily discovered by inspection of exemplars, where manual and automatic descriptions often identify misleading or incomplete explanations. We then describe a simple approach to extract these types of rules automatically from a transformer, and apply it to GPT-2 small. We find that a majority of features may be described well with around 100 skip-gram rules, though absence rules are abundant even as early as the first layer (in over a fourth of features). We also isolate a few examples of counting rules. This paper lays the groundwork for future research into rule-based descriptions of features by defining them, showing how they may be extracted, and providing a preliminary taxonomy of some of the behaviors they represent.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.

How this classification was reachedexpand

Full frame distilled prediction

Teacher imitation

Not calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.

metaresearch head score (Codex)0.002
metaresearch head score (Gemma)0.000
Version: codex-gemma-dda1882f352aValidation status: machine_predicted_unvalidated
Candidate categoriesMeta-epidemiology (narrow)
Consensus categoriesnone
DomainCandidate signal: none · Consensus signal: none
Study designCandidate signal: Observational · Consensus signal: Observational
GenreCandidate signal: Empirical · Consensus signal: Empirical
Teacher disagreement score0.297
Threshold uncertainty score0.999

Codex and Gemma teacher scores by category

CategoryCodexGemma
Metaresearch0.0020.000
Meta-epidemiology (narrow)0.0010.001
Meta-epidemiology (broad)0.0010.001
Bibliometrics0.0010.002
Science and technology studies0.0000.000
Scholarly communication0.0000.001
Open science0.0020.001
Research integrity0.0010.002
Insufficient payload (model declined to judge)0.0000.000

Machine scores (provisional)

The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.

Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.

Opus teacher head0.073
GPT teacher head0.317
Teacher spread0.244 · how far apart the two teachers sit on this one work
Validation statusscore_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it

Classification

machine, unvalidated

Machine predicted; a candidate call from one teacher head, not a consensus.

Study designObservational
Domainnot available
GenreEmpirical

How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".

Quick stats

Citations0
Published2025
Admission routes1
Has abstractyes

Explore more

Same venueArXiv.orgSame topicExplainable Artificial Intelligence (XAI)French-language works237,207