Unverified50% confidenceFactExact time
Under temperature=0 (greedy decoding) conditions, multiple runs of a coding agent on the same SWE-bench task still lead to different final states, meaning full reproducibility cannot be achieved.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn multi-turn autonomous tasks, numerical non-determinism is systematically amplified, so one should not blindly trust the reproducibility of temperature=0.75% similarUnverifiedTemperature=0（贪心采样）也不能保证模型输出的确定性，至少有三个原因：浮点运算非结合性、MoE路由不确定性、服务端配置漂移70% similarVerifiedEven when the temperature is set to 0, some inference frameworks still have non-determinism in underlying floating-point computations, causing output differences across hardware and across batches.70% similarUnverifiedHigh temperature combined with a missing repetition penalty can easily cause the model to fall into repetitive loops or generate meaningless content.68% similarUnverifiedUnder a Temperature 0 setting, QuantaMind always produces the same rating for the same input; its judgment rules include that required calls must be triggered, the appearance of prohibited calls results in immediate failure, and the end state must match exactly.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/901335API
curl https://kongchang.com/api/v1/knowledge/claims/901335MCP
get_claim(id=901335)