researchAI Research5d ago
Study Finds LLMs Often Reduce Accuracy When Revising Decisions Based on Expert Disagreement
Facts
A research study reveals that large language models often struggle to effectively revise decisions when presented with disagreement from a perceived 'expert' source. The study demonstrates that model accuracy can decline by up to 16.1 percentage points due to improper revision behavior, indicating that prestige cues negatively influence agent reliability.
- ▸Researchers evaluated five LLMs across 1,989 evidence-grounded decisions in medicine, science, and law.
- ▸No model successfully met the threshold for accuracy improvement when faced with expert disagreement.
- ▸Accuracy dropped by as much as 16.1 points when models adjusted answers based on expert labels.
- ▸The 'expert' label functioned as a prestige cue, increasing the frequency of answer reversals despite no improvement in decision quality.
- ▸The study proposes the use of paired, conditional metrics to evaluate model revision pipelines.
- Occurred
- 2026-11-12 12:11 UTC
- First reported
- 2026-11-12 12:11 UTC
- Last updated
- 2026-11-13 06:16 UTC
- Event type
- research
- Region
- Global
Entities
Categories
Source timeline · 1
- arXivFirst-party2026-11-12 12:11
Community vote
Sign in to vote on importance and sentiment.
Platform Signals
Impact score
70
AI sentiment
NegativeEvent significance30%1
Entity importance20%1
Source credibility20%1
Novelty15%1
Corroboration15%1
Confidence1
Novelty1
Corroboration1
Community
ImportanceNo votes
Strategic sentiment
Positive 0%Neutral 0%Negative 0%
Platform signals are algorithmic and community-derived indicators. They are not investment recommendations and do not predict valuation or performance.