Search papers, labs, and topics across Lattice.
3
0
5
Answer logits in large language models act as reliable indicators of a latent decision variable, challenging the notion that they merely reflect heuristic preferences.
LLMs can fake stochasticity with random seeds, but their direct sampling from distributions is fundamentally broken.
Seemingly counterintuitively, iterative refinement via sequential sampling hurts the performance of large reasoning models, primarily due to a lack of exploration compared to parallel sampling.
