[HUGGINGFACE]score: 0.62
5WBENCH Reveals Machine Unlearning Methods Fail on Causal Why-Type Questions
May 27, 2026
5WBENCH is a 1,000-sample-per-category benchmark across Who/What/When/Where/Why question types, exposing that Why-type causal questions are under 0.06% of CounterFact and under 1.3% of TOFU, masking systematic unlearning failures. No existing baseline simultaneously achieves high forgetting and high retention on Why-type questions.
paper
HOW THIS AFFECTS YOU
●
researcher5WBENCH exposes a structural evaluation gap in machine unlearning benchmarks and provides a balanced test set that makes causal knowledge forgetting failures measurable for the first time.
●
policyWorth watching because current unlearning compliance claims based on CounterFact or TOFU scores may be systematically misleading for causal knowledge removal.