Navalia framework improves long-horizon agent performance via verifiable task synthesis and post-training

Researchers propose Navalia, a framework designed to advance LLM-based cowork agents by combining verifiable task synthesis with long-horizon post-training. The approach targets financial analysis workflows, where agents must integrate cross-document evidence and perform multistep numerical reasoning. To address the high cost of expert task authoring and the unreliability of direct LLM synthesis, Navalia uses corporate financial reports as its data source to semantically select supporting evidence and construct executable computation chains. These chains generate analytical tasks with traceable evidence and actual data dependencies between steps.

Navalia further employs independent LLM review and programmatic checks to verify evidence consistency and answer correctness before using the verified tasks for trajectory collection, cold-start supervised fine-tuning (SFT), and reinforcement learning (RL). The framework is initialized from the Qwen3.6-35B-A3B model and achieves consistent gains over the base model across five benchmarks: OfficeQA Pro, FinSearchComp, Finance Agent Benchmark, APEX-Agents-AA, and Workspace-Bench-Lite. The results indicate improvements in evidence integration, tool use, and long-horizon task execution, with generalization to broader cowork settings including general office tasks.