DeepSeek-V3.2
DeepSeekHistorical reasoning content is discarded only when a new user message is introduced to the conversation. If only tool-related messages ... are appended, the reasoning content is retained throughout the interaction.
specific method · filed under inference & serving
Thinking Context Management is a context-management rule for tool-calling conversations that decides when the model's reasoning content is kept or discarded.1,2 It retains reasoning across tool-related messages, unlike the DeepSeek-R1 strategy of discarding it at the second message round, which forces the model to redundantly re-reason for each tool call.3,4
Trade-off Preserving reasoning avoids redundant re-reasoning and its token inefficiency, but agent frameworks that simulate tools via user messages may not benefit.5,6
Summarised by glm-5.3-flash from 2 sources; every sentence is checked against the quote it cites.
we developed a context management strictly tailored for tool-calling scenarios— paper
Historical reasoning content is discarded only when a newuser messageis introduced— paper
If only tool-related messages (e.g., tool outputs) are appended, the reasoning content isretainedthroughout the interaction— paper
discarding reasoning content upon the arrival of the second round of messages—results in significant token inefficiency— paper
forces the model to redundantly re-reason through the entire problem for each subsequent tool call— paper
These frameworks may not fully benefit from our enhanced reasoning persistence— DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
Also called Thinking Context Management.
2 quotes from 2 sources, one card per model, each opening with its strongest.
Historical reasoning content is discarded only when a new user message is introduced to the conversation. If only tool-related messages ... are appended, the reasoning content is retained throughout the interaction.