Ctrl+K
- claim-1-regret-bound-depends-on-subproblem-dimensionality
- claim-1-the-conformal-ratio-cr-metric-balances-coverage-and-prediction-set-size-under-split-conformal-prediction-to-expose-fake-forgetting-that-ua-ra-ta-mia-miss-sec-2-3-1-table-2
- claim-2-coupling-error-bound-and-regret-decomposition
- claim-2-on-cifar-10-resnet-18-under-10-random-forgetting-30-68-of-forget-samples-the-unlearned-model-misclassifies-still-have-ground-truth-inside-the-conformal-set-table-3
- claim-3-multi-expert-framework-and-lagrangian-coordination
- claim-3-with-alpha-0-1-and-a-2000-sample-calibration-set-9-unlearning-methods-are-evaluated-on-coverage-set-size-and-cr-sec-2-3-2-table-4
- claim-4-moco-benchmark-efficiency-vs-bo-baselines
- claim-4-the-miacr-metric-shows-methods-scoring-well-under-traditional-mia-can-score-poorly-under-miacr-so-mia-is-an-unreliable-forgetting-proxy-table-6
- claim-5-bi-tsp-100-head-to-head-vs-specialized-solver-and-pmoco
- claim-5-the-cpu-framework-l-total-l-original-lambda-l-unlearn-improves-ua-by-3-93-cifar-10-and-9-23-tiny-imagenet-while-degrading-ta-by-only-1-0-0-57-sec-3-2-table-7
- claim-6-hw-sw-co-design-proxy-vs-mobo-qparego
- conclusion
- executive-summary
- 1.04 kB