Browsecomp/HLE Reproducibility | 结果复现
Hi Zhipu team, thank you so much for open-sourcing such impressive models and sharing your research!
Just a question regarding reproducibility of the GLM-4.7 search-agent benchmarks: How can the BrowseComp and HLE(w/ tools) evaluation results be replicated? Is the search-agent and "context management" framework you used for BrowseComp/HLE evaluation open-source, or do you plan to open-source it?
Also, if I can also ask the same question concerning the Code agent, that was used for SWE-Bench verified/multilingual and Terminal bench 2.0? Are there any plans to open-source that code-agent framework?
Thanks again for your great work! 🙏
你好，智谱团队，非常感谢你们开源了如此出色的模型并分享相关研究成果！
我有一个关于 GLM-4.7 搜索代理（search-agent）基准测试可复现性的问题：BrowseComp 和 HLE（with tools）的评测结果是如何复现的？你们在 BrowseComp / HLE 评测中使用的搜索代理和“上下文管理（context management）”框架是否已经开源，或者是否有计划将其开源？
另外，如果可以的话，我也想就代码代理（Code agent）提出同样的问题：该代码代理被用于 SWE-Bench（verified / multilingual）以及 Terminal Bench 2.0。是否有计划开源这一代码代理框架？
再次感谢你们出色的工作！🙏