Search papers, labs, and topics across Lattice.
State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China
2
0
4
Major LLM failures can be traced back to specific disruptions in cognitive organization, providing a roadmap for targeted interventions and deeper understanding of model behavior.
Overthinking in RL models can amplify during training, but Dynamic Rollout Editing effectively curtails this behavior, enhancing reasoning efficiency.
