Bad Genius: Counterfactual-Guided Harness Evolution Beyond Task-Specific Shortcuts
This work introduces Counterfactual Harness Search and Evolution (CHASE), which casts harness evolution as constraint generation over valid counterfactual benchmarks and formalizes an ideal shortcut-neutralized benchmark $B_0$ and establishes theoretical guarantees linking finite counterfactual archives to $B_0$ and ch...
Guo-Jun Zhu, Xu Huang, Peng Yin et al.
· 0 citations