← News

Agents & MCPSources: arXiv · 19 Nov 2026

AnyS1 paper describes a small model that handles routine agent decisions

The arXiv abstract reports a framework that lets a lightweight model take over common choice points in LLM agent workflows, cutting cost and latency while a large model handles the rest.

In plain words

AI agents built on large language models often call a big, expensive model for every single step, even for simple picks between a few known options. The paper's authors propose AnyS1, which learns from written skill instructions where those simple choices occur, then uses a small 4B model to make them instantly. When the small model is unsure, the workflow hands the decision back to the large model.

What was announced

According to the abstract, current dual-system designs already route easy decisions to a lighter model, but each deployment needs custom hooks written for that specific workflow. AnyS1 instead treats human-written procedural skills as a general source of decision points: the instructions themselves indicate where a choice arises and which candidates apply. The authors say they mined such decisions from more than 1,000 public skills and collected trajectory states inside isolated containers to build a training set called Any-DP.

The resulting System-1 is a single 4B pointer model trained with label smoothing and calibrated so its confidence can be trusted. At run time, routine choices resolve in one forward pass, and control returns to the large model when confidence drops below a set threshold. On three benchmarks the system had not seen, and across different large-model families, the authors report cost per episode falling by 75-91% and model-call time by 63-92% versus an uncompiled agent, with task reliability preserved and thresholds transferring zero-shot. These figures come from the paper's own evaluation and have not been independently verified.

What it does not show

  • Evidence is limited to the paper's abstract; the full methodology, experimental setup and failure cases are not available for review.
  • The reported cost and latency reductions are the authors' own benchmark results and have not been independently confirmed.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.