Consistency training weakens answer cues while a cue-following circuit stays active
Source check: Source checked. Not independent scientific confirmation.
Finding: Researchers applied four consistency-training objectives to Gemma-2-2B-it and Llama-3.1-8B-Instruct. The trained models followed the training cue less often but also tended to discount correct suggestions, especially when cues appeared before the answer choices. Replacing activations in the known cue circuit with their uncued values corrected 59.8% to 91.0% of remaining cue-induced errors, compared with at most 12.1% for random attention heads in the same layers. This arXiv v1 preprint tests particular multiple-choice prompts in two models, with weaker transfer to unfamiliar cues and some circuit results based on one seed.
Technology: Gemma-2-2B-it; Llama-3.1-8B-Instruct; consistency training
Family Llama: profile · models page
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- In two pretrained language models, learned attention-head activations were intervened on to localize the cue-related errors that remained after consistency training.
- Test and result
- Four adapter-based training objectives used the same training prompts. Behavioral tests compared trained and original responses to correct and wrong suggestions. The mechanistic test replaced activations in a previously identified circuit and compared corrected errors with those from patching random heads in the same layers.
- Limitation / still unknown
- Transfer to other model families, document formats and cue types was not assessed by this comparison. Some mechanistic estimates use one seed, and the matched-question selectivity analysis was exploratory.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.25650v1 · https://arxiv.org/html/2610.25650v1
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 front matter, abstract, submission history, introduction, setup and results through circuit intervention.
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.