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

Introduction

Large language model (LLM) agents are moving beyond chat‑only assistants toward Cowork Agents—systems that can plan autonomously, invoke tools, and execute complex, multi‑step tasks. In domains such as corporate finance, these agents must pull together evidence from many documents and perform precise numerical reasoning. Creating high‑quality training tasks for such capabilities is expensive, and naïve LLM‑generated tasks often lack reliable answers. The Navalia framework tackles this gap by generating verifiable financial‑analysis tasks and then using those tasks to fine‑tune and reinforce‑learn agents for long‑horizon performance.

Verifiable Task Synthesis

Navalia starts with a corpus of corporate financial reports. It first semantically selects supporting evidence—complete fact sets that cover a target analytical question. From these evidence sets it builds executable computation chains that link data extraction, calculation, and reasoning steps. Each generated task therefore includes:

1. A clear question (e.g., “What was the YoY change in operating cash flow for Company X in Q4 2025?”).
2. The exact excerpts from the reports that serve as evidence.
3. A step‑by‑step computation plan that can be run programmatically.

To ensure correctness, Navalia employs two independent verification layers:

• LLM review – a separate language model checks that the evidence cited actually supports each reasoning step.
• Programmatic checks – the computation chain is executed; the resulting numeric answer is compared against the ground‑truth value extracted from the reports.

Only tasks that pass both checks are admitted to the training pool, guaranteeing traceable evidence and answer fidelity.

Long‑Horizon Post‑Training Pipeline

The verified tasks feed a three‑stage training regimen:

1. Trajectory collection – the base model (initialized from Qwen‑3.6‑35B‑A3B) attempts each task, and its step‑wise actions are recorded.
2. Cold‑start supervised fine‑tuning (SFT) – the recorded trajectories serve as supervised examples, teaching the model to follow the evidence‑driven reasoning flow.
3. Reinforcement learning (RL) – a reward model, derived from the verification checks, incentivizes correct evidence usage, tool invocation, and final answer accuracy across long horizons.

This pipeline equips the agent with the ability to maintain context over many intermediate steps, a critical skill for financial analysis where a single conclusion may depend on dozens of prior calculations.

Performance Highlights

Evaluated on five benchmarks—including OfficeQA Pro, FinSearchComp, and the Finance Agent Benchmark—Navalia consistently outperforms the base model. Gains are observed in:

• Evidence integration – agents retrieve and cite the correct document fragments more reliably.
• Tool use – the model more effectively calls calculators, spreadsheet APIs, and data parsers.
• Long‑horizon execution – success rates improve on tasks requiring ten or more reasoning steps.

Importantly, the improvements transfer to broader office‑task settings, suggesting that the verification‑driven synthesis approach generalizes beyond finance.

Practical Takeaways

• Task verification matters – embedding both LLM and programmatic checks creates a high‑quality training signal that mitigates hallucination.
• Evidence‑centric design – structuring tasks around explicit evidence chains teaches agents to ground their reasoning, a habit that scales to new domains.
• Iterative post‑training – collecting trajectories from the same model and then fine‑tuning it creates a feedback loop that steadily extends horizon capabilities.

Conclusion

Navalia demonstrates that coupling verifiable task synthesis with a disciplined long‑horizon post‑training regimen can substantially elevate Cowork Agents’ performance on financially intensive, evidence‑heavy workloads. By ensuring that every training example is traceable and correct, the framework not only improves immediate benchmark scores but also lays a foundation for trustworthy, autonomous agents in high‑stakes professional environments.