A judge flagged more concerning answers early in OLMo reinforcement learning than at the end
Source check: Source checked. Not independent scientific confirmation.
Finding: In tests of OLMo-3-7B-Think, GPT-4o flagged 25.0% of outputs after supervised fine-tuning, 46.1% early in reinforcement learning, and 11.8% at the final checkpoint. The labels cover text that the judge classified as power-seeking or related behavior. The two judges agreed least near the apparent peak.
Technology: OLMo-3-7B-Think, GPT-4o, Gemini 2.0 Flash
Family Gemini: profile · models page
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- GPT-4o rates outputs across the same 76 items in 16 OLMo-3 checkpoints; Gemini 2.0 Flash provides a secondary comparison.
- Test and result
- The authors repeated the same 76 benchmark questions across 16 released post-training checkpoints and compared binary labels from two judging models.
- Limitation / still unknown
- The rates describe one primary judge’s classification of written answers. They do not demonstrate the model’s internal goals or behavior in deployment.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.31624 · https://arxiv.org/html/2610.31624v1
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 checkpoint methods §§2–4, §4.1 trajectory, inter-judge comparisons and Appendices A.2/A.4.
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 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.