AI Science Tracker
EN

A hidden-state shift can cause a language model to repeat one token

The University of Tokyo; TIER IV; Huawei · 21 Oct 2026 · computer science

Discovery

A hidden-state shift can cause a language model to repeat one token

Source check: Source checked. Not independent scientific confirmation.

Finding: Researchers isolated a shift shared across responses during language-model preference training. Inserting the learned shift into an untreated model reproduced a failure in which one token dominated generation. The failure appeared in 14 of 17 tested model lineages under the study’s per-token ranking objective.

Technology: Preference-optimized language models, including Llama-2 and Llama-3.1

Family Llama: profile · models page

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

AI contribution
The researchers isolated a common direction learned during preference training and tested its contribution by transplanting it into an untreated language model.
Test and result
The authors isolated the shared offset, compared transplantation with magnitude-matched random directions, and tested masking of the component in the training loss.
Limitation / still unknown
The failure depends on model and objective. Related hidden shifts were also observed without generation collapse in tested standard preference-training runs.
Evidence sources and annotation context

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

Evidence sources: https://arxiv.org/abs/2610.29410 · https://arxiv.org/html/2610.29410v1

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, §§4–5 and 7, Appendix B.17, and primary intervention 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.

Open in the full catalog for filters, charts and related discoveries. Browse computer science discoveries.

Related catalog entries

Links use documented relationships where available, plus shared catalog context. A shared topic or model family does not establish a scientific dependency.