FLEET algorithm adds memory to search instead of sampling

The authors introduce FLEET, which attributes external rewards to specific tokens and uses MCTS to adjust logits, making generation aware of past rewards. It matched the sampling baseline with half the iterations on GSM8K and raised LiveCodeBench v6 from 0.59 to 0.69 in 9 iterations versus 32.

• Five Qwen3.8-27B models match Fable 5 in benchmark tests2026-10-25