Specialist LLM Agents for Real Estate Analysis

Summary & Methodology Analysis

This paper investigates whether localized numerical operations and integrative judgments of financial analysis benefit from LLM specialization, specifically examining prompt-level specialist decomposition and task-aligned reinforcement learning post-training in European listed real estate. The methodology maps a 16-lens European listed-real-estate analysis framework to eight lens-aligned specialists using Larix, routes each benchmark task to the corresponding lens-aligned specialist using a deterministic router, and evaluates three frozen-template frontier model prompting conditions. These conditions include a monolithic financial-analysis prompt, a monolithic prompt with the complete 16-lens framework, and lens-aligned specialist prompting under identical conditions. Additionally, the approach post-trains Qwen3.5-9B using Group Relative Policy Optimization, which is a reinforcement learning technique that optimizes policy models using relative group rewards, with task-aligned structured rewards based on the deterministic benchmark score.

The key results demonstrate that specialist decomposition improves the numerical-task aggregate by 15.8 percentage points across 19 firms spanning seven regulatory wrappers. Furthermore, the reinforcement learning post-training raises the development-split score by 12.0 points and the judgment aggregate by 14.2 points on Qwen3.5-9B. These post-training gains transfer to unseen firms with an overall increase of 15.2 points and an increase of 40.4 points on covenant stress. The models and tools utilized in this study include Larix, Qwen3.5-9B, Claude Opus 4.8, and veRL.

Despite these positive findings, the paper acknowledges two major limitations. The production Larix system contains a downstream synthesizer that is not invoked or scored, meaning no empirical claim is made about cross-agent synthesis, conviction calibration, or position sizing. Additionally, the 30-step training run evaluates one preserved checkpoint to establish feasibility but does not characterize convergence. The paper does not specify hardware requirements, latency numbers, or token costs.