A weighted training method that distinguishes likely shortcut examples improved AI robustness across four image and language benchmarks without adding an inference-time cost.
Some results have been hidden because they may be inaccessible to you
Show inaccessible results