FLEET: From Logits Entropy to Enhanced Trajectories in Text Generation
FLEET adds memory to sampling so repeated generations stop collapsing into near-duplicates.
The paper represents each output as a sparse trajectory through high-entropy token states, then uses those trajectories to adjust logits on later generations. In benchmarks, it matches repeated temperature sampling accuracy with a reported 3x speedup. On LiveCodeBench, Pass@32 rises from 59.9% to 66.2% under the same budget. The authors say the greedy-decoding setup is deterministic and needs only a single calibration pass. HF Daily Papers' note
The paper represents each output as a sparse trajectory through high-entropy token states, then uses those trajectories to adjust logits on later generations. In benchmarks, it matches repeated temperature sampling accuracy with a reported 3x speedup. On LiveCodeBench, Pass@32 rises from 59.9% to 66.2% under the same budget. The authors say the greedy-decoding setup is deterministic and needs only a single calibration pass. HF Daily Papers' note
score 5