FLEET: From Logits Entropy to Enhanced Trajectories in Text Generation
Oleksii Streltsov, Oleksandra Vitko
Kharkiv National University of Radio Electronics
Published September 23, 2026 · on the feed September 24, 2026 · ▲ 4 · arXiv 2609.27657
The abstract
Solutions based on large language models (LLMs) often rely on temperature sampling to improve accuracy and stability by aggregating multiple samples from the completion distribution. However, this memoryless approach is inherently suboptimal: because it lacks awareness of prior generations and their evaluations, it produces an increasing proportion of semantically duplicate answers as more samples are drawn, leading to diminishing returns. To address this limitation, we introduce FLEET, a novel method that integrates a memory mechanism into the generation process. FLEET represents each generation as a sparse trajectory through states whose entropy exceeds a predefined threshold and uses these trajectories to infer per-token utility scores that adjust the logits. Benchmark evaluations demonstrate that FLEET achieves the same accuracy as the repeated sampling baseline, with a 3x speedup, and substantially improves accuracy on complex coding tasks (LiveCodeBench Pass@32 increases from 59.9% to 66.2%) under the same budget. Furthermore, in the greedy-decoding configuration evaluated here, the approach is deterministic and uses a single calibration pass to derive its principal hyperparameters, requiring only minimal modifications to existing LLM pipelines.
The Brief
2:12A short podcast summary of the paper. Written and voiced by AI.
Transcript
FLEET. From Logits Entropy to Enhanced Trajectories in Text Generation. By Oleksii Streltsov and one other, from the Kharkiv National University of Radio Electronics. Published on the twenty-third of September. Its code repository is public and real: thirteen source files.
Solutions built on large language models often rely on temperature sampling, aggregating several samples from the completion distribution to improve accuracy and stability. That approach is memoryless. It has no awareness of earlier generations or of how they were evaluated, so as more samples are drawn, a growing share of them are semantically duplicate answers, and returns diminish. Each new sample is drawn as though the earlier ones never happened, so the same answer can be paid for again and again.
FLEET integrates a memory mechanism into the generation process itself. Each generation is represented as a sparse trajectory, through only those states whose entropy exceeds a predefined threshold. From these trajectories FLEET infers per-token utility scores, and those scores adjust the logits. The sparse trajectory keeps only the moments where the model was uncertain, which are the points where one generation could have gone differently from another. Those are the points the memory records.
The method is designed to sit lightly on existing pipelines. In the greedy- decoding configuration evaluated, it is deterministic, and it uses a single calibration pass to derive its principal hyperparameters. It needs only minimal modifications to the pipelines already in use. Determinism is a notable property for a method built around diversity. Run it again in that configuration and it gives the same result.
On benchmarks, FLEET matches the accuracy of the repeated sampling baseline with a three times speedup. Under the same budget, it substantially improves accuracy on complex coding tasks. On LiveCodeBench, pass at thirty-two rises from fifty- nine point nine percent to sixty-six point two percent. The two results are two sides of one trade. Hold accuracy fixed and FLEET gets there faster. Hold the budget fixed and it gets further.
Drawing more samples is usually a bet on variety. FLEET keeps a record of where earlier generations went at their high-entropy moments, and lets that record shape the logits of the next one, so that additional samples are less often spent on answers already seen.
Paper Quest · 7 chapters · about 39 min
Play this paper
Learn it by doing: make the calls the authors made, guess the result before you see it, and defend it in a viva graded on your own words. Chapter 1 is right here.
Chapter 1 🧩 · The problem
Sampling again and again, and paying for the same answer each time. That is the wall.
You ask a model the same coding problem many times and keep the best answer. Each sample is drawn with temperature, as if the earlier ones never happened.
The road ahead
- 1 · The problem ~6 min Drawing many samples of a model's answer wastes effort, because each sample ignores the earlier ones and repeats answers you already checked.
- 2 · What came before ~3 min
- 3 · The key idea ~8 min Remember each generation only at its uncertain branching points, score those branches by how the completion did, and steer later generations away from failed ones.
- 4 · Choose your own research ~3 min
- 5 · Build the method ~8 min Calibrate once, then for each generation record its uncertain states, turn past evaluations into per-token utilities, and lower the logits of failed tokens.
- 6 · The experiment ~6 min You’ll guess: To reach the same accuracy as the repeated sampling baseline, how much faster was FLEET?
- 7 · The viva ~5 min Three questions from the examiners, answered in your own words. Win it for the paper’s trophy.
Free, with Google or an email address. Your stars and progress are kept.