[Daily Picks]A LocalLLaMA post warns beginners that ultra-low sampling temperature (t=0 / greedy), often used to make speculative-decoding speed claims look great, can hurt coding quality. On Qwen3.8 (IQ4_XS-M) with fixed top_p/top_k, LiveCodeBench hard problems rose from about 70% at temp 0 to about 83% at temp 1.0; looping was also far more common at the low end. Knowledge/recall tasks were less hurt, but answers got longer at low temp.
Practical takeaway: don’t copy benchmark “greedy” settings into your daily coding agent. Start nearer the model card’s recommended sampler (often temp around 0.6–1.0 with sensible top_p/top_k). If you need reproducible evals, keep a separate greedy profile—don’t make it your interactive default.
What temperature do you use for local coding, and have you seen greedy loops in the wild?