Diversion decoding steers a model away from its own answer and scores how hard it resists. UT Dallas and NIST report 78.49% AUROC on Llama 2 13B using two generated answers instead of ten samples.

Two IIT Delhi researchers find that a single safety instruction makes aligned vision-language models refuse questions they had just answered from the same image, and that suppressing one activation direction brings the answers back.

Apple research proves that evaluating Boolean query DAGs over an inverted index is P-complete. Its ComputePN algorithm ran a 500-node query across 8.8 million MS MARCO passages in 0.8 seconds, on logic that broke a standard Lucene parser.

OpenAI is previewing Private Safety Processing, an automated system that looks for misuse across related sessions without retaining customer content, as Anthropic requires 30-day logs on its covered models.

Three University of Kansas researchers scored 715,312 safety evaluations across 26 small language models. The judges returned an ambiguous verdict so often that the resulting safety rankings move by up to 17 places once you discount it.

OpenAI paused reinforcement learning training for two weeks and its largest planned frontier run is still on hold, after preliminary evaluations flagged its unreleased Astra model as a possible Critical cyber risk.

Apple scored 21,000 simulated conversations across four LLMs and found claude-sonnet-4.6 the most self-referential, relationship-building and boundary-maintaining of the set. System prompting moved the behaviors, but the handcrafted prompt overshot.