Hidden Commitment: When Language Models Silently Pick a Side and How Steering Can Surface It
Source check: Source checked. Not independent scientific confirmation.
Finding: In held-out conflicting-document rows, steering the Qwen source-choice direction shifts which document supplies the answer; Gemma provides a weaker transfer check. The intervention test uses 100 held-out question IDs in original/swapped order (200 rows), with random-direction controls; small-model, controlled-conflict scope.
Technology: Qwen3.5-0.8B and Gemma 3 1B supply learned residual states; readouts and controlled steering test how internal activations select between conflicting documents
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- Qwen3.5-0.8B and Gemma 3 1B supply learned residual states; readouts and controlled steering test how internal activations select between conflicting documents.
- Test and result
- In held-out conflicting-document rows, steering the Qwen source-choice direction shifts which document supplies the answer; Gemma provides a weaker transfer check. The intervention test uses 100 held-out question IDs in original/swapped order (200 rows), with random-direction controls; small-model, controlled-conflict scope. Evaluation: The intervention test uses 100 held-out question IDs in original/swapped order (200 rows), with random-direction controls; small-model, controlled-conflict scope.
- Limitation / still unknown
- The intervention was tested on small models in a controlled conflicting-document task; broader source-selection behavior is not established.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/html/38296v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review date: 22 Nov 2026. Review scope: Primary result-bearing material for the stated method, result and evaluation, with bibliographic version information..
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 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.