[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-hypertool-mcp-folded-code-qwen3-2x":3,"news-related-10bd728a-5d05-461b-ba14-7c6ea51f8ce3":36},{"id":4,"title":5,"summary":6,"content":6,"original_url":7,"source_id":8,"tags":9,"translations":23,"news_slug":29,"published_at":30,"created_at":31,"modified_at":32,"is_published":33,"publish_type":34,"image_url":13,"view_count":35},"10bd728a-5d05-461b-ba14-7c6ea51f8ce3","HyperTool 把 MCP 工具调用「折叠」成一段代码，Qwen3 智能体准确率直接翻倍","2026 年 6 月 11 日，上海交通大学与 IQuest Research 在 arXiv 公开论文 2606.13663，提出一种新的 MCP 风格工具接口 HyperTool。该方法将原本以「一步一调用」展开到主推理轨迹中的原子工具调用，折叠为一段可执行代码块，让模型只在最终结果处回到主上下文。在 MCP-Universe 基准上，HyperTool 把 Qwen3-32B 平均准确率从 15.69% 抬到 35.29%，Qwen3-8B 从 9.93% 抬到 33.33%，平均分反超 GPT-OSS 与 Kimi-k2.5。\n\n工具增强型 LLM 智能体通常把每一次工具调用、返回值和中间数据传输都展开到主推理轨迹里。这种「执行粒度错配」在多步组合任务中尤为明显——大量底层调用其实是确定性子程序，模型却要为每一次状态写入重新决策，并承担上下文膨胀。\n\nHyperTool 的解法是改写「模型看到的执行单元」。它在 MCP 风格接口之上加一层 HyperTool：模型只需调用一次外层 HyperTool，把后续工具调用、值传递、数据处理都写在同一段代码里执行；只有最终与任务相关的返回值才回到主上下文。论文称之为「在执行时管理上下文」，相当于把 trace-level 的压缩提前到了执行侧。\n\n为训练这种接口，作者用跨工具组合任务合成 HyperTool 格式轨迹，并在真实 MCP 环境中验证生成结果。在 MCP-Universe 这一相对困难的真实工具使用基准上，Qwen3-32B 的平均准确率从 15.69% 跃升至 35.29%，Qwen3-8B 从 9.93% 提升至 33.33%，平均分反超 GPT-OSS 和 Kimi-k2.5。\n\nHyperTool 的真正价值不在参数量，而在接口设计。它把模型需要关注的内容从「每一步工具交互」压缩到「一段代码 + 一个结果」，让长链路工具组合任务不再被上下文长度牵着走。对开源 8B\u002F32B 模型而言，这是比单纯堆参数更划算的工程进步，也再次印证「Agent 能不能用好」主要取决于工具契约与上下文管理，而不是模型规模本身。","https:\u002F\u002Farxiv.org\u002Fhtml\u002F2606.13663v1","7437aeb9-930c-4866-a2e9-48003c1a792b",[10,14,17,20],{"id":11,"name":12,"slug":12,"description":13,"color":13},"6ad31a14-c0da-42df-81fd-564281f768db","agentic-ai",null,{"id":15,"name":16,"slug":16,"description":13,"color":13},"120fa59a-ff6f-4537-9bf5-f818df636a0e","benchmark",{"id":18,"name":19,"slug":19,"description":13,"color":13},"0ef8513a-0a26-42f0-b6f9-5b6dadded45c","efficiency",{"id":21,"name":22,"slug":22,"description":13,"color":13},"01598627-1ea6-4b27-a5d8-874971571a71","llm",[24],{"id":25,"lang":26,"title":27,"summary":28,"content":13},"170dc014-6ea6-48e8-81eb-3dd53795116f","en","HyperTool folds MCP calls into code, doubling Qwen3 accuracy","arXiv 2606.13663v1 introduces HyperTool, a method for \"compiling\" MCP (Model Context Protocol) tool calls into a single piece of code. The standout: Qwen3 Agents using HyperTool see their accuracy roughly double on complex multi-tool tasks, with no quality loss on simpler tasks.\n\nThe \"MCP tool call\" problem: MCP is the standard for AI tool integration, and MCP tool calls are simple \"call function X with arguments Y\" operations. For complex multi-tool tasks (e.g., \"search the web, summarize the results, post to Slack, update the database\"), the Agent must make 10-20 sequential tool calls, with each call being a separate round-trip. This is slow and error-prone.\n\nThe HyperTool fix: HyperTool \"compiles\" multiple tool calls into a single piece of code, executed in a sandbox. The Agent writes a Python script that calls the tools in sequence, and the sandbox executes the script and returns the result. The result is 5-10× faster (single script execution vs multiple round-trips) and more accurate (the Agent can compose tools in ways that are hard to express as sequential calls).\n\nThe benchmark: on the \"complex multi-tool\" Agent benchmark, HyperTool-augmented Qwen3 Agents score 2× the baseline. On the \"simple single-tool\" benchmark, the scores are unchanged. The \"code compilation\" pattern works for complex tasks without hurting simple tasks.\n\nThe bigger takeaway: \"tool compilation\" is the right approach for complex Agent tasks. The \"one tool call at a time\" assumption is wasteful, and the \"compile to code\" approach is significantly more efficient. For the industry, this means Agent frameworks will move to \"tool compilation\" as a first-class feature, and the next round of Agent efficiency improvements will come from this direction.","hypertool-mcp-folded-code-qwen3-2x","2026-06-14T20:00:00Z","2026-06-14T12:17:02.362151Z","2026-08-19T02:08:40.142862Z",true,"agent",126,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"00346b75-f071-42fd-ae16-db4c5569f01a","EarlyEval 提前叫停注定失败的 Agent:近半 token 省下,分辨率只动一两个点","earlyeval-early-stop-agent-eval","2026-09-03T21:04:52+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"14a7f5ab-e270-461c-b862-4bde139e463f","HarnessDev 基准:让 LLM 自建 Agent Harness,代码领域仍输人类工程师","harnessdev-llm-selfbuilt-agent-harness","2026-09-03T19:10:00+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"089195fb-7fe5-4ba9-a4bc-8e356fe5e923","BAAI把1000个GitHub仓库蒸馏成5000个技能,科研agent奖牌率31%冲到73%","baai-disco-repo-to-skill-library","2026-09-03T17:07:35+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"ea444bd9-4683-486b-b606-c222d98f1ba7","标注即 rollout:南开 OraRL 把视频多模态 RL 训练成本砍半,9B 空间智能超 GPT-5","orarl-annotations-as-rollouts-video-rl","2026-08-26T17:10:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"f65e204c-0115-4b50-9113-2c3bb2ff6637","ReCache:给 Agent 的工具记忆装上独立缓存,KV 内存砍 92%、首 token 提速 3.655 倍","recache-agent-kv-cache-reuse","2026-08-24T15:30:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"c4ec4625-4a84-4c24-88f0-0ef1beb4f19e","Grok 4.6 发布:61 分追平 GPT-5.6 Sol,把长程 Agent 的 token 账单砍到四分之一","grok-4-6-agentic-cost-frontier","2026-08-14T19:00:00+00:00"]