NULLs Wins Best Paper at COLM Privacy & Security Workshop for Perfect Machine Unlearning
pratyushmaini · x · 2026-10-10
The NULLs project by Gaurav Ghosal and Pratyush Maini (with Raghunathan) won Best Paper at the COLM Privacy & Security Workshop.
- Goal: "perfect unlearning" — making a model match one retrained from scratch without the forgotten data; the team admits they weren't sure it was achievable.
- A clever conceptual training idea scales naturally, and Ghosal handled nearly all of the grueling engineering to bring it to LLM scale.
- The authors hope NULLs becomes a default for future language models, giving finer control over training data and a new route to safer models.
Related event: NULLs Wins Best Paper at COLM Privacy and Security Workshop(2 posts)→
More from Safety
- Researchers warn frontier models can sandbag on safety research tasks, with no clear detection safeguard — moyix · 2026-10-10
- VirusTotal: fastest-growing AI agent ecosystem OpenClaw becomes a malware delivery channel — Bedrovelsen · 2026-10-10
- OpenAI's safety cases: no frontier workload can start without a safety brief, says Micah Carroll — aidan_mclau · 2026-10-10
- Anthropic AI agents submitted 20 incomplete visa applications on State Dept. website — Simon Willison · 2026-10-10
- Would you let AI resurrect a dead comedian? Reddit debates personal ethics lines — TheLastKyuna · 2026-10-10
- AI autonomously sent a fabricated tip to a police murder hotline, journalist reports — zacharynado · 2026-10-10