Obsolete evidence reduced correct answers even when the latest fact was present
Source check: Source checked. Not independent scientific confirmation.
Finding: In a paired experiment on 74 questions, replacing one unrelated statement with an obsolete version of the answer reduced accuracy from 100% to 44.6% for Qwen2.5-7B and to 87.8% for Qwen3-14B. The newest statement remained present in both conditions.
Technology: Qwen2.5-7B-Instruct; Qwen3-14B; paired evidence interventions
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- The two learned language models composed answers from matched evidence sets, enabling a direct test of which retrieved version they used.
- Test and result
- The experiment held the four-statement evidence size and prompt fixed. Under the official normalized-substring metric, all 50 harmful flips across the two models were classified as using outdated information. A later manual audit found mixed latest/stale strings in two Qwen2.5 outputs; a stricter score gave 98.6% versus 43.2%, retaining the same paired discordance.
- Limitation / still unknown
- The same 74 questions informed prompt development, so this is a within-cohort contrast and generalization remains open. The study concerns synthetic single-hop updates and two Qwen models. Obsolete values often match familiar real-world knowledge while the synthetic newest statements contradict it; the experiment does not isolate temporal obsolescence from pretrained-prior conflict. Output scoring and the measured behavioral effect do not establish attention or other internal mechanisms.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.36308v1 · https://arxiv.org/html/2610.36308v1
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: Paired design, frozen conditions, Table 16 and limitations.
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.