Skip to content
Thursday, 15 October 2026
Tenesys AI News
Subscribe

9B-parameter fine-tuned policy model

Change history

  1. 15 Oct 2026
    Small Fine-Tuned Model Learns to Follow New Regulated Call Scripts Without Retraining
    ParameterBeforeNow
    Documented check pass rate67-71%93-100%
    Grounded check pass rate97-100% (pre-training)97-100% (post-training, unchanged)
    Scenario pass rate15/4340/43
    Training dataNot specifiedFine-tuned on recorded calls spanning nearly 300 randomly generated regulated procedures
    Model sizeNot specified9 billion parameters

News

Voice AI· 🧪 Worth Testing

Small Fine-Tuned Model Learns to Follow New Regulated Call Scripts Without Retraining

Researchers describe a system called FlowSpec that converts a creditor's legal call procedure into machine-readable instructions, tool restrictions, training signals and test cases for a voice agent. A 9-billion-parameter model was trained on calls covering nearly 300 randomly generated procedures, then tested on four procedures it had never encountered, including one entirely held-out family of scripts, to see whether compliance behavior transfers.

arXiv cs.CL
9B-parameter fine-tuned policy model · TENESYS AI NEWS