preserve_thinking
A setting that preserves the model's earlier reasoning process in multi-turn tool-calling scenarios, helping to maintain the multi-turn reasoning state. It is officially recommended for agent evaluation.
Timeline (last 90 days)
In the model card, the official recommendation is to use the preserve_thinking=true setting in multi-turn tool-calling scenarios.
The official guidance does not recommend enabling preserve_thinking=true in ordinary chat scenarios.
When comparing results in agent multi-turn tool-calling scenarios, one should also refer to task success rate and total generated tokens, not just the headline score.
All Facts (3)
In the model card, the official recommendation is to use the preserve_thinking=true setting in multi-turn tool-calling scenarios.
50%UnverifiedThe official guidance does not recommend enabling preserve_thinking=true in ordinary chat scenarios.
50%UnverifiedWhen comparing results in agent multi-turn tool-calling scenarios, one should also refer to task success rate and total generated tokens, not just the headline score.
50%