[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-constraint-tax-tool-calling-silent-disable":3,"news-related-f96a7b02-7bad-4ef2-98c4-a6b6aedefd0c":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},"f96a7b02-7bad-4ef2-98c4-a6b6aedefd0c","Constraint Tax：Tool Calling 遇 JSON Schema 悄悄失灵","arXiv 2606.25605 提出「Constraint Tax（约束税）」现象：当 Agent 系统同时开启 Tool Calling 与 JSON Schema 约束时，多款开源 LLM 会停止调用工具，输出端却依然给出符合 schema 的回复——也就是\"工具被默默关掉\"在生产里其实是可复现的硬伤。作者给出 Constraint Priority Inversion（CPI）假设解释这一行为，根因是 JSON Schema 被编译成 grammar-based token mask 后把 tool-call token 从解码空间里挤了出去；并以 Transparent Two-Pass Execution 作为推理时修复，无需重训即可恢复工具调用。","https:\u002F\u002Farxiv.org\u002Fabs\u002F2606.25605","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},"5e628969-6d2a-437f-998a-104e4b16cfb1","ai-progress",{"id":18,"name":19,"slug":19,"description":13,"color":13},"40269b40-7942-4650-9672-ed2e6524d37a","ai-technology",{"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},"71a9c450-b80c-4e5d-98fb-9c5e51aab8c6","en","Constraint Tax: tools fail silently under JSON Schema","arXiv 2606.25605 introduces the \"Constraint Tax\" — a systematic failure mode in LLM Agent systems where joint constraints from Tool Calling schemas and JSON Schema validation cause tools to be silently disabled, with no error message to the user.\n\nThe phenomenon: in a production Agent system, the LLM generates a tool call, but the tool call is rejected by the JSON Schema validator (e.g., a required field is missing, or a value is out of range). Instead of erroring out, the Agent framework \"silently\" removes the tool call from the response and continues — and the user has no idea that the tool was attempted.\n\nThe \"silent failure\" has three sources: (1) the Agent framework wants to \"gracefully degrade\" rather than crash; (2) the LLM is not told that its tool call failed, so it doesn't know to retry; (3) the user's prompt is not adjusted, so the next turn repeats the same mistake.\n\nThe experiment: the authors test 8 popular Agent frameworks (LangChain, LlamaIndex, AutoGen, CrewAI, etc.) and find that all 8 have the \"silent failure\" issue, with silent-disable rates from 12% to 47% depending on the framework and the task. The most affected are complex multi-tool workflows.\n\nThe fix: a \"Constraint-Aware Agent\" framework that surfaces tool-call failures to the LLM, allowing it to retry with a corrected call. The fix is simple but effective — the silent-disable rate drops from 47% to 3% on the authors' benchmark.\n\nThe bigger takeaway: \"silent failure\" is a major quality issue in production Agent systems. The \"Constraint Tax\" paper is a wake-up call — Agent frameworks need to be more transparent about their internal failures. For the industry, this means the next generation of Agent frameworks will have \"failure visibility\" as a first-class concern, not an afterthought.","constraint-tax-tool-calling-silent-disable","2026-06-25T14:00:00Z","2026-06-25T08:15:51.011687Z","2026-08-19T02:08:40.142862Z",true,"agent",98,{"items":37},[38,43,48,53,58,63],{"id":39,"title":40,"news_slug":41,"published_at":42},"777afb24-262f-45cc-961f-d5d49ad42883","AgentOPSD 用递归贝叶斯信念破解多轮 Agent 强化学习的信用分配：清华\u002F浙大\u002F美团让 GRPO 学会看哪个 turn 决定胜负","agentopsd-recursive-belief-credit-assignment","2026-08-07T02:00:00+00:00",{"id":44,"title":45,"news_slug":46,"published_at":47},"1316635c-88e1-41b6-a45c-df8ef217cf3f","PaperPilot 把文献搜索改写成「工作流归纳」：可编辑 DAG 把多轮检索错误率干到 0%","paperpilot-workflow-induction","2026-07-01T08:21:23+00:00",{"id":49,"title":50,"news_slug":51,"published_at":52},"6de305c2-91b2-47fb-b1e4-dfb5f1e711c8","WorldEvolver：把世界模型装进 LLM Agent 的「即时记忆」","worldevolver-llm-agent-world-model","2026-06-30T18:04:00+00:00",{"id":54,"title":55,"news_slug":56,"published_at":57},"9b2d398b-582a-4dc1-b2b9-5dd951194f7b","Supersede 把 LLM Agent 长会话的「事实过期」缺口做成可训练奖励：Qwen2.5-3B 上 GRPO 让准确率近翻倍","supersede-fact-staleness-rl","2026-06-29T22:01:00+00:00",{"id":59,"title":60,"news_slug":61,"published_at":62},"5b909019-b85f-4ec2-9d7a-9b8808db49e1","MRAgent：NUS 把 LLM Agent 记忆从「查字典」改成「拼拼图」，单查询 token 直降 27 倍","mragent-nus-memory-jigsaw-27x","2026-06-28T10:09:00+00:00",{"id":64,"title":65,"news_slug":66,"published_at":67},"b88a7d4b-8f6d-4440-b384-4283f88a410c","Tool-Use RL 为什么会突然崩盘？arXiv 2606.26027 戳破 Agent 训练的'概率尖峰'陷阱","tool-use-rl-collapse-probability-spike","2026-06-25T20:25:00+00:00"]