safety security
The Paraphrase Kept What the Seed Forgot
GRAPHSU expands deletion pressure from named forget examples into neighboring support routes.
Summary
GRAPHSU expands deletion pressure from named forget examples into neighboring support routes.
The method builds a weighted graph around aliases, paraphrases and connected training samples, then applies graded forgetting to high-risk neighbors. On TOFU and PISTOL with GPT-2 Medium and Llama-3.2-3B-Instruct, the authors report up to a 49.5-percentage-point leakage reduction over a matched seed-only baseline while retaining feasible utility. Both benchmarks are controlled evaluation settings, so enterprise deletion claims still need domain-specific testing.
Why it matters
GRAPHSU expands deletion pressure from named forget examples into neighboring support routes.
Limits and context
- On TOFU and PISTOL with GPT-2 Medium and Llama-3.2-3B-Instruct, the authors report up to a 49.5-percentage-point leakage reduction over a matched seed-only baseline while retaining feasible utility.
Key claims
GRAPHSU expands deletion pressure from named forget examples into neighboring support routes.
Qualification: On TOFU and PISTOL with GPT-2 Medium and Llama-3.2-3B-Instruct, the authors report up to a 49.5-percentage-point leakage reduction over a matched seed-only baseline while retaining feasible utility.
Evidence: source-2026-08-30-004
Sources
- arXiv preprint 2608.26743arXiv · primary research
Corrections
No corrections have been recorded for this story.