LEMON: zero-shot LLM misbehavior monitor flags hallucinations from attention maps
yoavartzi · x · 2026-10-26
- Researchers from Technion, Tel Aviv University and NVIDIA introduce LEMON, an online misbehavior monitor that flags hallucinations and jailbreaks token by token from attention maps, with <10% overhead.
- Key challenge: attention heads/layers aren't aligned across LLMs (lookback ratios between Llama-2-7B and Mistral-7B correlate at only r=0.04). LEMON treats these mismatches as symmetries and builds an equivariant monitor, enabling zero-shot transfer to unseen LLMs.
- Paper on arXiv; code coming soon.
More from Safety
- AI shrinks the patch window to hours; air-gapped backups are your last line — aakashgupta · 2026-10-26
- NYT roundtable asks five experts what the rules for AI companies should actually be — ShakeelHashim · 2026-10-26
- Google DeepMind pilots world's first double-blind evaluations for frontier models — scychan_brains · 2026-10-26
- LLM back-of-envelope estimate puts PHP vulnerability impact at $1bn–$4bn — S_OhEigeartaigh · 2026-10-26
- US states halt new data centers as 72% of Americans oppose local builds — YvesMulkers · 2026-10-26
- a16z podcast: Diode pairs AI PCB design with Brooklyn factory to revive US hardware manufacturing — MattPerault · 2026-10-26