🫨简直无敌！DeepSeek Harness被清华微软团队直接“瘦身”：Token消耗砍掉87.93%，成功率反而暴涨14%！
This is a dev post classified by AGK Intelligence as AI dev tools (a tool drop), kept by the Dev Radar because it carries real work, not commentary.
🫨简直无敌！DeepSeek Harness被清华微软团队直接“瘦身”：Token消耗砍掉87.93%，成功率反而暴涨14%！ Agent每次工具反馈后都要重新问一遍LLM，哪怕下一步早定死了（写完代码→编译，编译成功→测试）。清华+微软团队用HarnessMM直接把这些“已决定”的步骤缓存下来，跳过重复推理。 核心成果（98个HumanEval任务）： 🔹DeepSeek Harness：Token从96.28k降到11.62k，砍掉87.93%！成功率从71.43%直接飙到85.71% 🔹Codex：成功率稳在90.82%，Token再降29.54%（66.72k→47.02k） 🔹原理：模型一次规划未来命令+精确参数+守卫条件（如“编译成功”），匹配就直接执行，跳过LLM回调。模型权重完全不动 🔹额外：MBPP上也大幅降本，递归自我改进（RSI）让Harness配置持续变强 Agent循环里大量Token浪费在“重新决定已经决定的事”上。这波直接把成本打下来，性能还往上走。 论文：https://arxiv.org/abs/2610.32813 #DeepSeek #AIAgent #Token效率 #清华 #微软 #HarnessMM #大模型降本 #Agent优化 #梁文峰 #梁圣
Posted by SuSu_酥酥👅 (25.6k followers) 1 h ago · 49 likes · 4.3k views · view the original post on X. Kept by the Dev Radar as AI dev tools. Tools mentioned: arXiv.org.
More dev work like this
- 美团发布 LongCat-2.5 正式版，距离今年 9 月预览版上线已有一个多月。 — @0xLogicrw
- Meet Ministral 3 14B Instruct 2512, abliterated and GGUF-ready. This uncensored 14B… — @HuggingModels
- 💥疯了！Qwen自己长出了一颗真GPU，在权重里跑真实CUDA！ — @NFT_Chen
- Made another tiny Claude Code mod: every GitHub PR link in the chat gets its status… — @adrianthedev
- 周末趁我们休息的时候，DeepSeek 又突然放出一个王炸：Super-agent（超级 Agent 系统）。 — @GitHub_Daily
- Meet Ministral 3 8B Instruct 2512, now in abliterated GGUF form. This uncensored,… — @HuggingModels
- A new study from Qwen finds that AI agents are less likely to complete tasks… — @Chinazhidx
- Let AI help administer Linux — without handing it an unrestricted shell. — @DanKornas
Every post is read and classified by AGK Intelligence: what it is, which market it belongs to, and whether the link is a real tool. 31.3k posts from 5k X accounts over the last 21 days, 9k tools, 12 markets. New posts are collected several times an hour and the full ranking is recomputed every hour — last update 2026-11-02 10:21 UTC. Full method.