Understand this story
Princeton University and Google published a paper studying the performance of a 4B model trained on agent states in agent tasks. The paper reports that the most accurate decision model on isolated problems solved only 69.5% of agent tasks, while the 4B model trained on agent states solved 91.5%. Its harness delegates tool selection at each step to a fast decision model, adopting its answer only when confidence exceeds a threshold fitted on the agent's own runs, and trained AgentDecision-4B on the states the agent actually reached, including states after mistakes. On tasks similar to the training data, this type of 4B model reduced agent costs by 81.1% to 89.0% and increased speed by up to 2.5 times, with a success rate gap of within 3 percentage points; delegating all tool selections without a threshold lowered the success rate by 19.0 percentage points. The paper is available at arxiv.org/abs/2610.21412.
AI-generated from public reports
Latest development · Report
Princeton and Google Paper: 4B Decision Model Shares Agent Tool Selection, Cutting Costs by 81.1% to 89.0%Developments
1 confirmed developments
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Oct 19
Rohan Paul
Princeton and Google Paper: 4B Decision Model Shares Agent Tool Selection, Cutting Costs by 81.1% to 89.0%
A new paper from Princeton University and Google finds that the most accurate decision model on isolated problems solved only 69.5% of agent tasks, while a 4B model trained on agent states solved 91.5%. Its harness delegates each step's tool selection to a fast decision model, adopting its answer only when confidence exceeds a threshold fitted on the agent's own runs, and it trained AgentDecision-4B on the states the agent actually reached (including states after mistakes). On tasks similar to the training data, such 4B models cut agent costs by 81.1% to 89.0% and boost speed by up to 2.5x, with a success-rate gap within 3 percentage points; without a threshold, delegating all tool selection drops the success rate by 19.0 percentage points. The paper is available at arxiv.org/abs/2610.21412.