RapidUn reweights LoRA unlearning by cross-sample influence: lower seen- and OOD-trigger ASR than Fisher, GA, and LoReUn on Llama-3-8B, with a 77x speedup over clean-corpus retraining

RapidUn converts cross-sample influence estimates into fixed sample-specific weights for weighted LoRA unlearning, removing targeted behavioral contamination under a LoRA-only PEFT setting. Relative to existing approximate unlearning methods, the key change is fixing the influence signal as sample weights rather than relying on continued post-deployment supervision, matching a practical setting with a small forget set, a limited retain buffer, and LoRA-only updates. The abstract states the framework's composition and target setting and reports results on Llama-3-8B (Dolly-15k, Alpaca-57k) and Mistral-7B + Dolly-15k; the weight computation details and hyperparameters are not expanded in the abstract.

On trigger-based behavior removal, RapidUn shows lower seen-trigger and OOD-trigger-family ASR than Fisher, GA, and LoReUn while keeping clean utility competitive. This provides comparative evidence for influence-guided sample reweighting against three baselines on a controlled trigger benchmark, rather than a single-method self-report. Evidence comes from comparisons on Llama-3-8B with Dolly-15k and Alpaca-57k plus cross-model validation on Mistral-7B + Dolly-15k; the abstract does not give specific ASR or utility numbers.

On Llama-3-8B + Alpaca-57k, RapidUn achieves a 77x wall-clock speedup over the clean-corpus LoRA retraining reference. This directly contrasts the efficiency gain against an explicit retraining reference, indicating approximate unlearning can be substantially cheaper in time than the retraining path. The figure comes directly from the setting described in the abstract; the measurement protocol and hardware environment are not stated there.

Complementary TOFU, semantic LLM-judge, and IFEval evaluations support influence-guided sample reweighting beyond the controlled trigger benchmark. The evaluation surface extends from a single trigger benchmark to other unlearning and general-capability tests, so the conclusion is not confined to one test construction. The abstract phrases these as evaluations that 'further support' the approach and does not report specific scores or scales for them.

The result targets a PEFT setting with a small forget set, a limited retain buffer, and LoRA-only updates, suited to engineering scenarios that must remove targeted behavioral contamination after deployment without affording full retraining; beneficiaries include teams handling model behavior governance and compliance cleanup. Cross-model validation covers Mistral-7B + Dolly-15k, indicating the method is not tied to a single model, while the applicable boundary remains the setting described in the abstract.

The abstract does not give specific ASR and utility numbers for each baseline, the measurement protocol and hardware for the 77x speedup, or the computational cost and weight stability of the influence estimation; specific scores for TOFU, the semantic LLM judge, and IFEval are likewise undisclosed. In addition, what was loaded here is the arXiv abstract page and metadata, not the body, figures, or appendix, so method details and full experimental tables still require consulting the original.