Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Aug 10·5 min read
MEDLEY-BENCH: Benchmarking Behavioural Metacognition and Belief Revision Under Social Pressure in Large Language Models
TL;DR
MEDLEY-BENCH is a new benchmark designed to evaluate how large language models revise their beliefs under social pressure and conflicting evidence. It reveals significant variability in model responses, highlighting the need for deeper understanding of metacognitive processes in AI.
✦ Why It Matters
Evaluate your AI models using MEDLEY-BENCH to better understand their belief revision capabilities under social pressure.
Key Takeaways
- Implement MEDLEY-BENCH to assess belief revision in your models.
- Analyze how models respond to conflicting evidence and social pressure.
- Use the Medley Metacognition Score for a structured evaluation.
- Prepare for the upcoming v1.5 release for improved scoring reproducibility.
Full Summary
Current benchmarks for large language models primarily assess the quality of final answers, neglecting how these models adjust their beliefs when faced with disagreement or conflicting information. MEDLEY-BENCH introduces a structured evaluation framework that compares self-review and social revision processes across 35 models from 12 families using the Medley Metacognition Score (MMS).
The study found that the Evaluation-mapped composite score was the lowest in 30 out of 35 models, indicating variability in how models handle belief revision. An exploratory analysis showed that sensitivity to manipulated consensus labels varied widely among models.
Additionally, a human rubric study indicated moderate agreement between human reviewers and model responses. This initial release, MEDLEY-BENCH v1.0, aims to complement traditional accuracy assessments by focusing on belief revision under ambiguity and social disagreement, with plans for a more robust v1.5 version.