Morning Edition · Wednesday, October 21, 2026Published at 2:08 AM EDT · New York

Navalia Post-Trains Qwen3.6-35B-A3B on Machine-Verified Finance Tasks as Vendors Build Rubric Graders on GPT-6.1 Sol

The open-weight agent improves on its base model across five finance and office benchmarks by synthesizing tasks from corporate filings with executable calculation chains, while Terminal X reports its rubric-based grader matches senior investors' judgment 88.6% of the time.

Long-horizon finance agents are limited by training data as much as by model capability. Hiring experts to write tasks is expensive, and tasks synthesized directly by an LLM often have unreliable answers. Navalia, a new arXiv paper, builds…