AI Science Tracker
EN

Verifier feedback raised both correct and incorrect model verdicts

The University of Manchester; University of Rajshahi; Varendra University · 27 Oct 2026 · computer science

Discovery

Verifier feedback raised both correct and incorrect model verdicts

Source check: Source checked. Not independent scientific confirmation.

Finding: On 32 software verification tasks, Qwen2.5-Coder-7B averaged 22.0 correct verdicts with verifier feedback versus 12.7 without it. Incorrect verdicts also rose from one to seven. A deterministic validation baseline returned 26 correct verdicts and left six unresolved.

Technology: Qwen2.5-Coder-7B-Instruct; verifier output interpretation

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

AI contribution
The trained coding model interpreted formal-verification evidence, allowing a direct comparison of verdicts under controlled feedback conditions.
Test and result
The fixed assertion set was evaluated with and without feedback for three seeds. Correct, incorrect and unresolved decisions were counted separately.
Limitation / still unknown
The subset was small and selected from one verification benchmark. The deterministic baseline included an additional validation step and is not an equivalent raw-feedback condition.
Evidence sources and annotation context

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

Evidence sources: https://arxiv.org/abs/2610.36947v1 · https://arxiv.org/html/2610.36947v1

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: §§III–IV, E3 frozen-subset design, Figure 3 and interpretation limits.

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.

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