AI Science Tracker
EN

Gemini flags local execution errors in successful multi-agent runs

Hong Kong University of Science and Technology; Adobe Research · 19 Oct 2026 · computer science

Discovery

Gemini flags local execution errors in successful multi-agent runs

Source check: Source checked. Not independent scientific confirmation.

Finding: In 1,828 benchmark traces from nine multi-agent configurations, a Gemini 3.1 Pro judge flagged local execution errors in 54.3% of successful runs (140 of 258) and 75.2% of failed runs (1,181 of 1,570). In a 45-trace sample, two human auditors both retained 64 of 73 existing flags and found additional errors. The result describes the errors the judge flagged under the study's definition and does not estimate how many errors it missed across the full corpus.

Technology: Gemini 3.1 Pro

Family Gemini: profile · models page

Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.

AI contribution
A pretrained Gemini 3.1 Pro model assigned candidate local error labels within traces produced by other agent systems.
Test and result
The authors annotated 1,828 existing benchmark traces from nine configurations. Two people audited a stratified random sample of 45 traces. Three further language models reannotated a separate sample of 90 traces.
Limitation / still unknown
Annotation completeness was uncertain. The human audit found 21 additional errors across 11 traces, including five found by both auditors. The observed percentages do not measure the full prevalence of execution errors or operational reliability.
Evidence sources and annotation context

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

Evidence sources: https://arxiv.org/abs/2610.25852v1 · https://arxiv.org/html/2610.25852v1

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 front matter, abstract, submission history, trace collection and annotation audit, matched replay methods and results.

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.

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

Related catalog entries

Links use documented relationships where available, plus shared catalog context. A shared topic or model family does not establish a scientific dependency.