Context Management Reproducibility | 可复现性 ?

[panademo]

Hi StepFun team, thank you so much for open-sourcing such impressive models and sharing your research!

Just a quick question on the discard-all strategy used for BrowseComp:

When the context length exceeds the threshold and the agent “discards its entire context,”(ref from the Step 3.5 Flash paper), does that mean everything accumulated (tool calls, reasoning, observations, etc.) is removed except the system prompt and initial user message/question?

Also, is the agent framework used for BrowseComp/HLE evaluation the same as (or similar to) the one in Step-DeepResearch?

Thanks again for your amazing work! 🙏

---

Hi 阶跃星辰团队，感谢你们开源如此出色的模型并分享你们的研究成果！

有一个关于 BrowseComp 中使用的 discard-all 策略的小问题：

当 context length 超过阈值，agent “discards its entire context”（引用自 Step 3.5 Flash 论文）时，是否意味着此前累计的所有内容（tool calls、reasoning、observations 等）都会被移除，仅保留 system prompt 和 initial user message/question？

另外，用于 BrowseComp/HLE 评测的 agent framework，是否与 Step-DeepResearch 中使用的框架相同或类似？

再次感谢你们出色的工作！🙏

[hengm3467]

Pasted the reply from our researcher to the HuggingFace discussion on the same question:

To answer your questions:

Discard-all Strategy: You are correct. When the context length hits the threshold, the agent clears all accumulated tool calls, reasoning steps, and observations. It effectively resets the memory, retaining only the system prompt and the initial user message to keep the core objective in focus while freeing up space for new exploration.
Agent Framework: The framework used for BrowseComp/HLE is indeed very similar to the Step-DeepResearch architecture. Both rely on a core ReAct loop managed by a dedicated Context Manager. The primary difference lies in the toolsets: BrowseComp utilizes a specialized suite of internal optimization tools tailored specifically for complex web browsing and high-level reasoning tasks.