Temperature Sweep Links Low Decoding Settings to Worse Coding
AI summary
A Reddit author tested Qwen3.8 in IQ4_XS-M quantization on 100 hard LiveCodeBench v6 problems while varying only the decoding temperature. Pooled coding accuracy rose from 70.0% at temperature 0 to 82.7% at 1.0, while the loop rate fell from 9% to 0.3%. The author also reported that lower temperatures did not reduce performance on GPQA Diamond knowledge questions.
In-depth insight
The disclosed test suggests that very low-temperature or greedy decoding can trade coding quality and looping behavior for determinism, offering local-model users a practical decoding consideration.