Clone-Robust Weights in Metric Spaces: Handling Redundancy Bias in Benchmark Aggregation
The discussion of a paper from ETH Zurich examines 'redundancy bias' in AI benchmarks, where multiple similar tasks skew results. The authors propose a method using 'local voting' and metric space geometry to assign weights based on task similarity. This approach ensures fair evaluation by preventing near-identical tasks from dominating the final score.