Search papers, labs, and topics across Lattice.
Peking University
3
0
5
4
LLM agents still fail to reliably automate real-world workflows, with even the best models succeeding on only two-thirds of tasks in a new live benchmark.
Current autonomous agent benchmarks miss nearly half of safety violations and over 10% of robustness failures because they only check final outputs, a problem Claw-Eval directly addresses.
LLMs trained with a novel "second-order rollout" that generates critiques in addition to responses learn more effectively from the same data, unlocking better reasoning.
