Search papers, labs, and topics across Lattice.
University of Maryland, College Park
4
0
8
Early hidden states in language models can predict steering success with surprising accuracy, enabling more efficient control over model behavior.
Scoring fixed-length reasoning chunks with LLMs can outperform traditional majority voting by up to 28 percentage points, all without the need for reward model training.
LLMs can slash multi-hop retrieval latency by 40% using SpecHop, a framework that speculatively executes multiple reasoning paths and rolls back incorrect ones.
Chain-of-thought reasoning can be sped up by 25-50% without sacrificing accuracy, simply by watching how the model's confidence changes as it thinks.
