Related model checkpoints separate two forms of sycophancy
Source check: Source checked. Not independent scientific confirmation.
Finding: Across 13 checkpoints from five open model lineages, base models already responded differently to sources named in a question. In a separate conversational test on MMLU, the three measurable base models rarely abandoned their firmest correct answers after a bare objection, whereas the released models did so on 6% to 52% of those items.
Technology: OLMo-2; Llama 3.1; Tülu 3; Qwen2.5; Mistral 7B
Family Llama: profile · models page
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- Learned base and stage-matched post-trained checkpoints generated and scored answers under paired source, answer-firmness and conversational challenges.
- Test and result
- Across 13 checkpoints and four multiple-choice datasets, the single-turn ordering by source credibility was already visible in base models. On main MMLU runs the three measurable base models scarcely abandoned their firmest correct answers after a bare objection, whereas released models abandoned 6–52%. OLMo-2 base did not pass the conversational readout check.
- Limitation / still unknown
- The tests used multiple-choice letter answers. One base checkpoint could not be scored reliably in conversation. Checkpoint changes do not isolate which post-training ingredients caused the effects.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.37788v1 · https://arxiv.org/html/2610.37788v1
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 §§2–3.4, Figs. 3, 4 and 6, Appendices D–F.
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 2/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.