Language-model credence updates moved outside the starting opinion range
Source check: Source checked. Not independent scientific confirmation.
Finding: In a controlled study of eight instruction-tuned language models, 22% to 68% of updates fell outside the interval between the two stated starting credences. Depending on model, 4% to 38% did so opposite the drift measured without peer input. Small model populations also sometimes spread farther apart under schedules that never widened opinions in a classical averaging comparison.
Technology: Qwen2.5; Gemma; Llama 3.1; Mistral 7B; token-probability credence readout
Family Llama: profile · models page
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- Eight instruction-tuned Qwen, Gemma, Llama and Mistral models responded to controlled scalar credence exchanges. Output token probabilities gave observable estimates of the updated credence, with matched no-peer controls.
- Test and result
- Across 73,080 operator measurements, 22.3%–68.3% of model updates left the initial opinion interval, and 3.7%–38.0% did so opposite the no-peer drift. Small model populations widened their opinion range in 5.9%–25.6% of rounds, versus zero for a DeGroot averaging comparison using the same schedules.
- Limitation / still unknown
- The readout concerns single-claim scores in a fixed prompt design, not general belief or human behavior. The study used models no larger than 9 billion parameters, some with quantization. The 10,000-agent analysis was a simulation of fitted functions.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.37699v1 · https://arxiv.org/html/2610.37699v1
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 §§3.2–3.5, 5.1–5.4 Table 2 and §7.
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.