Search papers, labs, and topics across Lattice.
4
0
7
7
Decentralized agent swarms can learn to self-orchestrate their collective behavior and individual capabilities simultaneously, outperforming fixed coordination strategies.
PRISM enables rapid and interpretable training data attribution, achieving 500x faster insights into model behavior while maintaining high accuracy.
Economic interactions among agents can spontaneously generate sophisticated reasoning strategies, outperforming traditional monolithic approaches.
Escaping the confines of autoregressive generation, Bidirectional Evolutionary Search unlocks significant performance gains by recombining partial trajectories and recursively decomposing tasks into checkable subgoals.