Review of 1,322 Papers Finds 92% of LLM Privacy Research Fixates on Memorization
tianshi_li · x · 2026-09-25
A systematic literature review, now accepted at NeurIPS after a year in the pipeline, argues privacy in LLMs is not just memorization.
The team reviewed 1,322 papers (2016–25) across ML, NLP, and security, finding 92% fixate on memorization and chat leaks. They map five urgent overlooked problems — surveillance, inference, aggregation, and other negative downstream outcomes — and propose a research roadmap to broaden the field.
More from Research
- Visual Jenga scored three 5s at NeurIPS 2025 but still landed only a poster — anand_bhattad · 2026-09-25
- Asking Claude to introspect after experimenting with interpretability tooling — KhoslasBiggestOpp · 2026-09-25
- After Posting a NeurIPS Meme, She Got 4 Papers Accepted, 1 Spotlight — AnkaReuel · 2026-09-25
- AAAI reviewer rant: template-violating, AI-sounding reviews advance broken papers — OutsideSimple4854 · 2026-09-25
- NeurIPS Datasets and Benchmarks Track notifications are live on OpenReview — tfburns · 2026-09-25
- First Recorded LLM NetHack Ascension: GPT 6 Astra Wins in 37,140 Turns — egrefen · 2026-09-25