DecisionGym

Headlines · 1

1. Oct 14
   JevAgentHarness delegates agent tool choices to small decision models

   An arXiv paper submitted 15 Oct 2026 by Kunlun Zhu and co-authors introduces JevAgentHarness, which delegates the multiple-choice decisions LLM agents make at each step — such as which tool to call next — to decision models like Jev, falling back to the large LLM when a per-choice confidence threshold fitted on the agent's own trajectories is not met. The authors built AgentDecisionBench and report that the decision model most accurate on isolated questions solved only 69.5% of agent tasks, and that delegating tool choices without a fitted confidence threshold lowered success by 19.0 points. They also trained AgentDecision-4B, an open 4B decision model, on post-mistake agent states labeled with Jev's answers; per the authors, on tasks unlike its training data the agent solved 91.5% with AgentDecision-4B and 82.0% with Jev, while on tasks like the training data such models made agents 81.1–89.0% cheaper and up to 2.5 times faster without losing 3 or more points of task success. The models, data and harness are said to be released.

   Hugging Face · Papers · 🔥 0

Experience and discussion from the community

Share my DecisionGym experienceAsk about DecisionGym