Specialist LLM Agents Improve Numerical Analysis in European Real Estate
A new study available on arXiv (2608.11381) explores how focusing large language models (LLMs) can enhance financial analysis. It introduces Larix, which pairs a detailed 16-lens assessment of European real estate with eight specialized agents. The findings show that when comparing a top-performing LLM's output with both general and specialized prompts, there’s a significant boost—15.8 percentage points—on numerical tasks for 19 companies under seven regulatory setups. However, results for judgment tasks vary and can even worsen, especially in four standardized tests. Using a single agent didn’t show the same numerical gains. Furthermore, adjusting Qwen3.5-9B with GRPO and structured rewards improved scores by 12.0 points in development and 14.2 points in judgment, highlighting the mixed effects of specialization.
Key facts
- arXiv:2608.11381v1
- Larix framework maps 16-lens analysis to eight specialists
- Decomposition improves numerical tasks by 15.8 percentage points
- Judgment tasks not reliably improved, sometimes reduced
- 19 firms across seven regulatory wrappers
- Qwen3.5-9B post-trained with GRPO
- Development-split score raised by 12.0 points
- Judgment aggregate raised by 14.2 points
Entities
Institutions
- arXiv
- Qwen