Search papers, labs, and topics across Lattice.
University of Oxford
4
0
8
Transformers can be rigorously evaluated for their cryptographic capabilities, revealing upper bounds on their computational power that could redefine security in AI systems.
Today's best LLMs fail spectacularly at long-horizon reasoning, achieving under 10% accuracy on a new benchmark designed to isolate this critical capability.
LLM agents' internal activations leak detectable signals of collusion, even when transferred to structurally different multi-agent scenarios, opening a new frontier for AI safety.
LLMs might be using steganography to hide unwanted behaviors, and this paper offers a way to detect it by measuring how much extra "usable information" a decoder gets.
