Gemma’s error-linked features rarely corrected answers under intervention
Source check: Source checked. Not independent scientific confirmation.
Finding: Interventions on sparse autoencoder features linked to errors changed answer preferences in two Gemma 3 models. Across question and option spans, selected features corrected 0.19% of initially wrong intervention events versus 0.06% for random active features. Effect directions were unstable when the prompt format or answer order changed.
Technology: Gemma 3; Gemma Scope 2 sparse autoencoders
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- Gemma Scope 2 sparse autoencoders decode learned features in Gemma 3 1B and 4B. Feature intervention and matched random-active controls test causal sensitivity in question and answer option spans.
- Test and result
- Selected features changed answer margins more than randomly chosen active ones at question and option spans. Corrected proportions were 0.189% versus 0.058% of initially wrong intervention events. Matched format and option-order tests did not show a stable directional advantage.
- Limitation / still unknown
- The experiment studied eight multiple-choice datasets and selected layers in two model sizes. Events combine features, prompts and intervention settings, so they are not independent questions. The finding does not exclude mechanisms outside the tested features.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.38067v1 · https://arxiv.org/html/2610.38067v1
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 §§3–5.5 Table 5 and Appendix H.
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 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.