FLEET: Entropy-Trajectory Memory Beats Repeated Sampling with 3x Speedup
Oleksii Streltsov · hf · 2026-09-24
FLEET addresses the diminishing returns of temperature-based repeated sampling: because it's memoryless, more samples yield a growing share of semantically duplicated answers. FLEET adds memory by representing each generation as a sparse trajectory through high-entropy states, inferring per-token utility scores that adjust logits.
- Same accuracy as repeated sampling with a 3x speedup under the same budget.
- LiveCodeBench Pass@32 jumps from 59.9% to 66.2% on complex coding tasks.
- Deterministic in the evaluated greedy configuration, with hyperparameters derived from a single calibration pass and minimal pipeline changes.
More from Research
- Gilbert Strang on the most important equation in statistics and linear regression — tetsuoai · 2026-09-24
- Altman claims OpenAI model solved Navier-Stokes, a Millennium Prize problem — victor_explore · 2026-09-24
- Paper: LLM reasoning operations are internally separable — wrong math still carries the right signature — rohanpaul_ai · 2026-09-24
- Together launches Jev-style Tev1-4B classifier on Qwen3.5, train your own for $17 — usamawahabkhan · 2026-09-24
- Multiple postdoc roles in Melbourne to build formally verified seL4 sandboxes for AI agents — mboehme_ · 2026-09-24
- Researcher Who Spent a Year on Feature Attribution Vows Never Again — aryaman2020 · 2026-09-24