Large language model (LLM) agents are shifting from conversational assistants to Cowork Agents that can plan autonomously, use tools, and carry out complex tasks. Financial analysis demands that these agents integrate evidence across documents and perform multi‑step numerical reasoning. Expert task authoring is expensive, and direct LLM synthesis often fails to produce challenging tasks with reliable answers.
The Navalia framework addresses this gap by coupling verifiable task synthesis with long‑horizon post‑training. Using corporate financial reports as the data source, Navalia semantically selects supporting evidence from complete fact sets and builds executable computation chains, creating analytical tasks that contain traceable evidence and real data dependencies between steps. Independent LLM review and programmatic checks are then applied to verify evidence consistency and answer correctness.
The verified tasks feed into trajectory collection, cold‑start supervised fine‑tuning (SFT), and reinforcement learning (RL), improving the model’s ability to execute long‑horizon tasks. Initialized from Qwen3.6-35B-A3B, Navalia consistently outperforms the base model on five benchmarks—OfficeQA Pro, FinSearchComp, Finance Agent Benchmark, APEX-Agents-AA, and Workspace‑Bench‑Lite—demonstrating gains in evidence integration, tool use, and long‑horizon execution, and showing generalization to broader office cowork settings.
Review: Navalia’s verifiable task generation pipeline provides a reliable data backbone for high‑quality financial analysis tasks, and its post‑training strategy markedly enhances LLM performance in complex, cross‑document reasoning scenarios.