Navalia improves cowork agent performance over Qwen3.6-35B-A3B on finance benchmarks
Researchers present Navalia, a framework that equips cowork agents with verifiable task synthesis and long‑horizon post‑training. The system draws on corporate financial reports, selects supporting evidence, and builds executable computation chains that create multi‑step analytical tasks with traceable data dependencies. Independent LLM reviews and programmatic checks verify evidence consistency…
Key points
- Navalia creates multi‑step financial analysis tasks using verifiable evidence from corporate reports.
- Initialized from Qwen3.6-35B-A3B, Navalia outperforms the base model on five office‑task benchmarks.
- The framework combines independent LLM review, programmatic checks, SFT and RL to improve long‑horizon execution.
Navalia fine‑tunes a base Qwen3.6-35B-A3B model through supervised fine‑tuning and reinforcement learning, aiming to boost autonomous planning, tool use, and long‑horizon execution. Across five benchmarks—OfficeQA Pro, FinSearchComp, Finance Agent Benchmark, APEX‑Agents‑AA, and Workspace‑Bench‑Lite—the Navalia‑enhanced model shows consistent gains over the unmodified Qwen3.6‑35B‑A3B, demonstrating better evidence integration and numerical reasoning in financial analysis contexts.
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Researchers propose SPOKE method to improve AI poker strategy · 1 src
- Does Multi-Agent Collaboration Reduce Hallucination in Clinical Tasks? · 1 src
- LLMs struggle with planning problems as symbolic heuristics outperform learned models · 1 src
- LightCRS boosts multi-domain conversational recommendation accuracy and speed · 1 src
- Researchers introduce MAMath-Bench for multimodal math reasoning · 1 src
Comments
via GitHub Discussions