AI Science Tracker
EN

Universitat Autònoma de Barcelona

1 discovery from 2026

Discoveries

A hidden Gemma arithmetic state steered the next answer digit

Source check: Source checked. Not independent scientific confirmation.

Finding: In Gemma 4 12B-IT, a learned activation subspace carried arithmetic results across text and image tasks apart from the directions used to read the current digit. With the same first digit supplied, changing the remaining numerical state made the counterfactual units digit the top prediction in 77.6% of 1,536 interventions.

Technology: Gemma 4 12B-IT; Distributed Alignment Search; activation interchange

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

AI contribution
Distributed Alignment Search learns activation subspaces in Gemma 4 from paired arithmetic instances. Component-specific activation swaps isolate a part orthogonal to the current digit readout and test its downstream effect.
Test and result
Across four text/image and addition/subtraction conditions, the 13-dimensional component orthogonal to the current digit readout showed 75.5% numerical retrieval across conditions among 20 values excluded from alignment fitting. In 1,536 cache tests with the same first digit supplied, the counterfactual units digit was the top prediction in 77.6% of cases.
Limitation / still unknown
The experiments used controlled arithmetic prompts and fitted subspaces. Numerical values held out for cross-condition alignment could still occur during earlier subspace training. A second model was tested in the same study.
Evidence sources and annotation context

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

Evidence sources: https://arxiv.org/abs/2610.37422v1 · https://arxiv.org/html/2610.37422v1

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: arXiv v1 primary HTML §§3–6, Appendix D.1–D.2, E.1–E.2 and F.

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 the interactive profile for the timeline, or browse all organizations and model families.