AI Science Tracker
EN

Attention heads selected on symbols also affected evidence judgments in text

University of California, Irvine · 27 Oct 2026 · computer science

Discovery

Attention heads selected on symbols also affected evidence judgments in text

Source check: Source checked. Not independent scientific confirmation.

Finding: Ablating eight Pythia-410M attention heads chosen using symbolic prompts reduced accuracy on a natural-language evidence task by 8.2 percentage points. Ablating random sets matched by layer and size reduced it by 0.8 points. A separate full-precision analysis showed an effect in the same direction.

Technology: Pythia-410M; attention-head ablation; related native full-precision models

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
The learned attention heads were selected using controlled changes in symbolic evidence and then intervened on during a separate natural-language evidence comparison.
Test and result
The study selected the eight heads without natural-language prompts, then compared ablation on held-out domains and paraphrases. The original comparison averaged five seeds.
Limitation / still unknown
The task asks which state has more supporting reports. The result establishes influence under intervention in these models, not a general account of reasoning or identical head functions across models.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2610.36524v1 · https://arxiv.org/html/2610.36524v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 21 Nov 2026. Review scope: §§3–5.6, Tables 5–6 and native precision controls.

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open in the full catalog for filters, charts and related discoveries. Browse computer science discoveries.

Related catalog entries

Links use documented relationships where available, plus shared catalog context. A shared topic or model family does not establish a scientific dependency.