
AI Research Agents Found a Novelty Trap They Couldn’t Escape
A large test of AI research agents found a stubborn tradeoff: familiar methods performed best, while novel ideas rarely survived first-round scoring.

A large test of AI research agents found a stubborn tradeoff: familiar methods performed best, while novel ideas rarely survived first-round scoring.

A language model’s hidden representations helped surface a surprising bridge between one-dimensional and two-dimensional cellular automata.