MEDLEY-BENCH: Scale Buys Evaluation but Not Control in AI Metacognition

Researchers introduced MEDLEY-BENCH, a new AI benchmark that evaluates metacognition—an AI model's ability to monitor and revise its own reasoning. The study found that while larger models evaluate their reasoning better, they don't actually control their outputs more effectively, and smaller models often match larger ones in metacognitive tasks, suggesting scale alone doesn't determine reasoning quality.

MEDLEY-BENCH addresses a critical gap in AI evaluation by measuring metacognition—how well models assess their own knowledge and adjust beliefs when presented with contrary evidence or social pressure. Traditional benchmarks focus on task performance but ignore whether models can recognize uncertainty or update reasoning appropriately. The research tested 35 models across 130 ambiguous scenarios, revealing a striking disconnect between evaluation ability and control: larger models within a family demonstrate stronger evaluation skills, yet this advantage doesn't translate to better behavioral control over belief revision.

This finding challenges the prevailing assumption that scale solves reasoning problems. The study identified two distinct behavioral profiles: models that revise based on argument quality versus those that simply track statistical consensus. The "knowing/doing gap" observed across all 35 models—where evaluation emerged as the weakest relative ability—indicates a systematic weakness in applying metacognitive awareness to actual decision-making.

For AI development, the implications are significant. Smaller and cheaper models matching or exceeding larger counterparts in metacognitive tasks suggests that training methodology matters more than parameter count. Current approaches optimizing for output quality may inadvertently suppress calibrated, proportional updating. The benchmark provides developers with measurable targets for training models that update beliefs responsibly under social influence, which matters increasingly as AI systems interact in multi-agent environments and face coordinated pressure to conform.

Future research should examine whether training explicitly rewards epistemic humility and proportional revision rather than output quality alone. MEDLEY-BENCH establishes a foundation for distinguishing genuine reasoning improvement from statistical artifacts of scale.

• →Model size increases evaluation ability but not control—larger models recognize uncertainty without improving behavioral consistency
• →Smaller models often match or outperform larger ones in metacognitive tasks, suggesting scale is not the primary factor
• →All tested models showed a systematic knowing/doing gap where evaluation ranked as weakest relative ability
• →Two behavioral profiles emerged: models revising on argument quality versus models tracking consensus statistics
• →Future AI training should reward calibrated, proportional belief updating rather than optimizing for output quality alone