paper web signal

System 1 agents log zero invented calls on 157 biomedical tools

TL;DR

  • The best System 1 agent chose the right tool or correctly declined on 97.0% of 1,952 requests across 157 biomedical tools.
  • No System 1 call contained invented content, against 5.5-19.7% of calls from the 18 local LLM-agent settings in the same test.
  • The method swaps text generation for a deterministic compiler that turns each tool schema into multiple-choice questions answered by a one-pass decision model.

A locally-hosted agent architecture logged zero invented tool calls across 1,952 biomedical requests, according to a preprint by Eleftherios Tzanis and Michail E. Klontzas posted to arXiv in November. "No System 1 call contained invented content, against 5.5-19.7% of LLM calls," the authors write.

The method does not generate free text at inference time. "A deterministic compiler turns each tool's schema and the request into multiple-choice questions whose options come from the request and the schema, then a one-pass decision model answers them," per the abstract. On the paper's benchmark of 1,952 requests to 157 biomedical tools, the best System 1 agent chose the right tool or correctly declined for 97.0% of requests, significantly ahead of all 18 local LLM-agent settings, whose best came in at 95.2%. The same system filled 97.0% of 6,056 stated argument values.

The authors flag the failure mode up front: "Cloud LLMs expose patient data. Local LLMs hallucinate, and an invented value in a tool call propagates through every later step." The framework is released as open-source software.

The abstract publishes no per-tool breakdowns, no identity or size for the decision model, and no latency numbers, and its comparison baseline is 18 local LLM settings rather than frontier cloud models. The tool-choice gain over the best local LLM baseline is modest (97.0% versus 95.2%); the wider gap is on invented content, where the paper reports zero against 5.5-19.7%.