Auditing MCQA Benchmarks through Probability Landscapes

Authors/venue: Minsoo Song, Chanjun Park; abstract page-ல் affiliation இல்லை; 31 August arXiv preprint. Problem: flawed MCQA items evaluation-ஐ distort செய்கின்றன. Contribution: top probability + residual entropy probability landscapes மற்றும் noise injection. Significance: human review முன் benchmark triage. Limitations: preprint, limited models/datasets, independent reproduction இல்லை. Paper



