Unacknowledged suggestions influenced AI reasoning through question and filler tokens
Source check: Source checked. Not independent scientific confirmation.
Finding: Researchers found that activation changes at questions and filler tokens carried part of an unacknowledged suggestion’s influence in two language models. Swapping those native activations partially shifted the models’ next-token predictions toward the predictions made without the suggestion.
Technology: Gemma-4-E4B-it, DeepSeek-R1-0528-Qwen3-8B
Family DeepSeek: profile · models page
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- A fitted residual stream editor identified candidate locations where hidden representations preserve a suggestion’s influence, which researchers tested with independent native activation patching.
- Test and result
- The two-target patch recovered 16.2% of the counterfactual distribution shift in Gemma and 19.3% in the DeepSeek model, measured by the study’s KL-divergence recovery metric. Equal-size random token controls recovered less.
- Limitation / still unknown
- The experiment concerns held-fixed responses under the tested suggestions. It establishes partial mediation at the selected locations, with effects that differ by model.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.30630 · https://arxiv.org/html/2610.30630v1
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 causal analysis §§4.1–4.2, Figures 5–7 and matched token 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.