Search papers, labs, and topics across Lattice.
UC San Diego
1
0
Self-looping reasoning in LLMs can be effectively mitigated through targeted hidden-state interventions, leading to improved reasoning quality and efficiency.