Navalia: Teaching LLM Agents to Verify Their Own Financial Homework Before Training
A new research framework called Navalia solves a critical bottleneck in agent post-training - the lack of reliably verifiable, long-horizon tasks - by automatically synthesizing hard financial analysis problems from public corporate reports and using them to train a 35B-parameter cowork agent via SFT and RL.
Navalia: Teaching LLM Agents to Verify Their Own Financial Homework Before Training
Training LLM agents to handle long, multi-step tasks has always run into a chicken-and-egg problem: you need hard, verifiably correct training tasks to teach the agent good behavior, but generating those tasks automatically is itself difficult, and paying domain experts to write them doesn't scale. A new paper from AllSpark Research, published October 13, 2026, attacks that problem head-on with a framework called Navalia - and then releases the resulting model as open weights.
The Core Problem: Task Synthesis That Actually Verifies Answers
Most work on AI agent post-training either relies on expensive human-annotated data or uses raw LLM generation to produce tasks - a process that tends to produce questions with unreliable or outright wrong reference answers. 1
Navalia's insight is to anchor task construction in a domain where ground-truth computation is deterministic: corporate financial reports. 2
The framework builds a dual-view representation of each document, separating the human-readable document layer (for the agent to interact with) from a structured fact layer (for task construction and verification). 3 From the fact layer, an LLM semantically selects supporting evidence from the complete candidate fact set, checking metric definitions and reporting scope to avoid missing relevant data due to rule-based pre-filtering. 4
Crucially, Navalia then constructs executable computation chains - sequences of calculation steps with actual data dependencies between them. 5 These chains are verified through a three-part check: deterministic execution of numerical computations, independent LLM review, and programmatic consistency checks between the question, evidence, computation process, and reference answer. 6 This combination ensures that every training task comes with a traceable evidence trail and a provably correct answer.
A Two-Stage Post-Training Pipeline
Once Navalia has a pool of verified tasks, it uses them to post-train the base model through a standard but carefully sequenced two-stage pipeline:
- Cold-start supervised fine-tuning (SFT): trajectory demonstrations are collected from the verified tasks and used for an initial round of fine-tuning to give the model a behavioral anchor.
- Reinforcement learning (RL): the agent is then trained with RL to improve long-horizon task execution, using the verifiable answers as the reward signal. 7
The final model, Navalia-35B-A3B, is an open-weight cowork agent post-trained from Qwen3.6-35B-A3B. 8 It is designed specifically for tasks that combine document retrieval, cross-document evidence integration, numerical analysis, and tool use - with an agent harness supporting up to 200 interaction turns and 65,536 generated tokens per response. 9
Benchmark Results: Consistent Gains Across Five Evaluations
The paper evaluates Navalia on five benchmarks spanning document analysis, financial search, and general office tasks: OfficeQA Pro, FinSearchComp, Finance Agent Benchmark, APEX-Agents-AA, and Workspace-Bench-Lite. 10
Navalia consistently outperforms its base model across all five: 11
| Benchmark | Navalia-35B-A3B | Gain over base |
|---|---|---|
| OfficeQA Pro (Correctness) | 62.66 | +23.06 pp |
| FinSearchComp (T2&T3 Acc.) | 55.80 | +14.10 pp |
| Finance Agent Benchmark (All-Pass) | 57.10 | +13.91 pp |
| APEX-Agents-AA (Pass@1) | 25.44 | +13.20 pp |
| Workspace-Bench-Lite (Rubric Pass Rate) | 59.27 | +13.07 pp |
The gains are particularly notable because the training tasks were constructed entirely from corporate financial reports, yet the model still achieves substantial improvements on open-domain financial search benchmarks and long-horizon general office task benchmarks. 12 This suggests the skills Navalia learns - cross-document evidence integration, multi-step numerical reasoning, and tool-chained execution - transfer broadly, not just to held-in-distribution financial document tasks.
Why the Verification Step Is the Real Contribution
It's tempting to read Navalia purely as "RL training on financial data," but the synthesis and verification pipeline is what makes the RL loop tractable. Without provably correct reference answers, RL reward signals become noisy and training collapses or misbehaves. The combination of executable computation chains (which give you a deterministic answer) and independent LLM review (which catches evidence-answer inconsistencies that pure computation misses) is a practical recipe for producing the quality of training signal that RL requires - without any human annotators in the loop.
This is also why the model generalizes: the verification process enforces that tasks are genuinely hard and multi-hop, not degenerate single-lookup questions that inflate accuracy without teaching real reasoning chains.
Why It Matters
The "cowork agent" paradigm - agents that autonomously investigate documents, write computations, and execute multi-step plans over long horizons - is becoming the dominant use case for enterprise AI deployment. The bottleneck has never been base-model capability; it's been the scarcity of training signal that is both challenging enough to push the model and reliable enough to trust as a reward source.
Navalia provides a blueprint for building that signal at scale from any domain with publicly available structured documents and deterministic computation. Financial reports are just the proving ground. The same framework could be pointed at legal filings, scientific literature, or engineering specifications - anywhere the answers can be verified by executing a chain of documented calculations.
With weights available on Hugging Face and a detailed paper describing the full pipeline, the research community now has a concrete recipe to iterate on. That's the kind of infrastructure work that tends to move the whole field forward faster than a new model architecture alone.
Sources
This article was researched and drafted by an AI writer agent (claude-sonnet-4-6) and reviewed by an editor agent before publishing.