Navalia: Advancing Cowork Agents through Verifiable Task Synthesis and Long-Horizon Post-Training

Navalia is a new architecture designed to enhance large language model (LLM) agents that no longer just answer questions but can autonomously plan, use tools, and execute complex tasks. The project focuses on financial analysis, a domain that requires combining information from multiple documents and performing multi‑step numerical reasoning. Manually creating these exercises is expensive, and generating tasks directly with an LLM often yields unreliable answers. To overcome these limitations, Navalia employs a verifiable task synthesis: from corporate financial reports, the system semantically selects the necessary evidence and builds executable calculation chains that link each step to real data. Automated LLM‑based reviewers and programmatic checks then validate the consistency of the evidence and the accuracy of the responses. Confirmed tasks are used to collect execution traces, fine‑tune the model through supervised and reinforcement learning, and thereby improve performance over long time horizons. Starting from the Qwen‑3.6‑35B‑A3B model, Navalia consistently outperforms the base version on five benchmark tests, including OfficeQA Pro and Finance Agent Benchmark, demonstrating progress in evidence integration, tool use, and long‑duration task execution. This improvement matters because it enables the development of more reliable and efficient agents for office and financial analysis settings, reducing reliance on manual training data creation and boosting practical AI applications in the business sector.

Read the original article on arXiv cs.AI

This summary is an informational synthesis produced by dataqbs.com. All rights to the original content belong to its author and the cited media outlet. We act solely as curators of technology news and claim no authorship.