Navalia: verifiable synthetic tasks to train financial agents

An open mixture-of-experts model with 35 billion total parameters and around 3 billion active (that is what the A3B suffix indicates), trained on synthetic financial tasks that are verified before use. That is the idea behind Navalia, a paper published on October 21 on arXiv (2610.26887). Its ten authors start from Qwen3.6-35B-A3B and report consistent gains over the base model on five benchmarks: OfficeQA Pro, FinSearchComp, Finance Agent Benchmark, APEX-Agents-AA and Workspace-Bench-Lite.

The paper is framed around what the authors call Cowork Agents, agents that no longer just converse but plan, use tools and carry long tasks from start to finish. Financial analysis is their test case, and the choice makes sense.

The problem it tries to solve

An analyst working with corporate reports has to cross-reference data from several documents and chain calculations together. They take a margin from the income statement, compare it with another fiscal year and adjust it with a figure from the balance sheet. Training an agent to do that requires tasks with the same structure and reliable answers. The abstract points to two approaches and their limits: having experts write tasks by hand is expensive, and asking an LLM to produce them directly tends to yield undemanding exercises or answers nobody can guarantee.

How Navalia works

Navalia builds tasks from the data, rather than starting from the question. The process has four steps:

1. It takes corporate financial reports as its source and semantically selects supporting evidence from complete fact sets.
2. It builds executable computation chains. Each task therefore has traceable evidence and real dependencies between steps: step two needs the result of step one.
3. It verifies each task through an independent review by another LLM and programmatic checks that the evidence is consistent and the answer is correct.
4. With the verified tasks, it collects trajectories, applies cold-start supervised fine-tuning (SFT) and finishes with reinforcement learning (RL) aimed at long-horizon tasks.

The decisive part is the second step. If the answer comes from executing a computation chain, it can be checked without relying on another model's judgment, and the RL reward becomes verifiable. It is the same logic that has worked in math and code, carried over to a setting with long documents and scattered data. And in long-horizon tasks an early mistake carries through to the end, so having a checkable answer at every link is worth even more.

What we still don't know

The abstract mentions consistent gains on all five benchmarks, but it gives no figures and no comparison with proprietary frontier models. We will need to read the full paper to learn how much it improves, on which kinds of task and at what training cost. The authors also claim the agent generalizes to general office tasks. That fits with OfficeQA Pro and Workspace-Bench-Lite being among the benchmarks, but it needs confirming with the numbers in hand.

Another reasonable concern is format bias. If every task comes from corporate reports with a similar structure, the agent may end up learning to navigate that specific format rather than to reason about finance in general.

Who this is useful for

For teams adapting open models to finance, auditing or back-office work, Navalia offers a concrete recipe for generating training data without relying on an army of annotators. Its value goes beyond fine-tuning, because the same procedure can be used to build evaluation suites with checkable answers. Anyone working with agents in Claude Code (subagents, skills or MCP servers that read internal documentation) can apply the idea without training anything. It is enough to generate tasks from the data, compute the answers with code and add a hook that checks the result before accepting it.

At ElephantPink, the verification method interests us more than the model improvement itself. Tasks with checkable answers are what separate an agent that seems to work from one you can actually measure.