Should we benchmark conceptual capabilities using judgment prediction tasks?

Short summary

The post proposes using judgment prediction tasks—asking AIs to predict a specified expert's judgment on subjective conceptual questions—as a benchmarking methodology for AI capabilities on disagreement-laden tasks. Key challenges include noisy human judgments, knowledge cutoff confounds, and difficulty generating objective ground truth. The approach aims to isolate conceptual reasoning ability from taste/prior disagreements between models and judges.

• •Proposes judgment prediction as a benchmarking method for subjective conceptual AI tasks
• •Identifies noise, knowledge cutoffs, and lack of ground-truth feedback as key downsides
• •Goal is to measure conceptual reasoning capability separately from model-judge taste disagreements

Generated with AI, which can make mistakes.