Open Models Almanac
Methodspost-trainingsupervised fine-tuning

single mention · post-training

Prompt-based cold start for reasoning with tool use

A cold-start strategy that combines reasoning and tool-use capabilities through carefully designed prompting rather than new training data.1 It uses system prompts to instruct the model to incorporate tool calls within its reasoning process, occasionally yielding desired trajectories that provide a basis for subsequent reinforcement learning.2,3

Trade-off The reasoning in tool-use patterns may lack robustness, but the model can still occasionally generate the desired trajectories.4

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. a straightforward strategy for integrating these two capabilities is through carefully designed prompting — DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
  2. the system prompt we designed to instruct the model to incorporate multiple tool calls within its reasoning process — DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
  3. providing a basis for subsequent reinforcement learning stages — DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models
  4. although the reasoning in tool-use patterns may lack robustness, the model is occasionally able to generate the desired trajectories — DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models

Also called Cold-Start.

Filed under post-training › supervised fine-tuning