Topic
Evaluating Model Capabilities After Benchmark Saturation
Latest news
- 16 OctPaper proposes using model transcripts to measure capabilities after benchmark saturation
The new arXiv paper argues that when benchmark scores saturate, full solution transcripts from models contain richer information than scalar scores. The authors simulate saturation and investigate whether transcripts can recover capability differences between models.
Research · 1 source
Articles
No articles about Evaluating Model Capabilities After Benchmark Saturation yet.