Author(s): Chew Loong Nian – AI ENGINEER

Originally published on Towards AI.

Temperature 0 vs 1.0: Greedy Decoding Collapsed 14% of Llama’s 128K Calls

Temperature 0 is the most copy-pasted line in production LLM code. It is also, on long context, the single worst value you can pick.

After introducing the core finding, the article explains that a 172B-token study shows temperature 0 dramatically increases “coherence loss” (requests that fall into repetition loops and produce no usable output), especially at long contexts like 128K–200K. The author then walks through the underlying mechanism: with greedy decoding, temperature=0 makes the argmax choice unable to flip, so escape from a loop becomes effectively impossible, and expected escape time grows (roughly doubly exponentially) as temperature decreases toward zero. The piece further argues that temperature=0 is not truly deterministic in practice—batch-dependent kernel behavior can still cause different outputs—so teams gain little reproducibility while inheriting severe reliability risk. It compares accuracy tradeoffs, noting temperature=0 can win on some metrics for short prompts, but often loses on grounding/fabrication and becomes increasingly problematic as context grows. Finally, it offers practical guidance: for long-context tasks, use a moderate temperature (about 0.4–0.7), cap max_tokens to limit runaway loops, explicitly pin decoding parameters instead of relying on defaults, and monitor truncation/finish_reason distributions to detect coherence-loss issues.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI