Community
Reddit post reports low temperature sampling hurts Qwen coding scores on LiveCodeBench
A LocalLLaMA user's sweep on a 4-bit Qwen3.8 quant found LiveCodeBench v6 accuracy rose from 70.0% at temperature 0 to 82.7% at temperature 1.0, with looping at 9% for temperature 0. The results are self-reported and not peer reviewed.
Every outlet that ran it 1 article
r/LocalLLaMACommunityLead
Ultra low temps make your model stupid at coding (a warning for beginners)