ASAP: Evaluating Agentic Capabilities During Pretraining
October 13, 2026
The ASAP framework allows for the evaluation of agentic behavior in pretrained base models without converting them to chat-style agents. It uses structured trajectory continuation and minimal recovery mechanisms to assess software engineering and research capabilities directly from the completion interface.
HOW THIS AFFECTS YOU
●
builderThis framework helps determine if a base model is suitable for agentic tasks.
●
researcherYou can evaluate how base models handle interactivity before the expensive post-training phase.