FLEET lifts coding Pass@32 to 66.2%

FLEET, a memory-augmented decoding method, lifts LiveCodeBench Pass@32 from 59.9% to 66.2% under the same budget, a 6.3-point gain. The paper says it also matches repeated-sampling accuracy at 3× speed, but its evaluation uses Llama 3.2-3B on 222 post-cutoff coding tasks and remains a research result. Coding-agent teams should treat search memory as a candidate alternative to simply buying more samples, then benchmark latency and verifier quality against the output budget exposed by Opus 5.5’s price cut.