Grounded but Not Immune: Confirmation Bias in Agent-as-a-judge
Source check: Source checked. Not independent scientific confirmation.
Finding: In controlled evaluations across four agent judges and five datasets, a false prescreening note overturned up to 40% of correct judgments even when judges gathered decisive evidence. The study characterizes this as confirmation bias under its tested conditions, not as a general attack success rate.
Technology: Tool-using language-model judges evaluated with controlled false prescreening cues
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- Language-model judges gathered evidence from tool-use trajectories; false contextual cues were manipulated to test verdict changes.
- Test and result
- Across four judges and five datasets, a false prescreening note overturned up to 40% of correct tool-using judgments despite decisive evidence collection. Repeated paired judgments characterized the observed failure modes.
- Limitation / still unknown
- The cue is a controlled diagnostic, not an adversarial attack; results are bounded to studied agents, datasets and task settings.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/html/2611.05662v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review date: 22 Nov 2026. Review scope: Primary result-bearing material for the stated method, result and evaluation, with bibliographic version information..
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 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.