# Summary: 2026-08-11_19-42-15Z_FromNumberstoJudgment_SpecialistLLMAgentsandReinfo.md
Saved: 2026-08-12 22:26
Source: 2026-08-11_19-42-15Z_FromNumberstoJudgment_SpecialistLLMAgentsandReinfo.md
Model: None

---

## Summary  
The paper investigates whether the localized numerical operations and integrative judgments required in European listed‑real‑estate analysis benefit from LLM specialization. It compares a monolithic prompting approach with an eight‑lens specialist decomposition, holding all other factors constant, to see if modular execution improves performance. The study also evaluates how reinforcement learning fine‑tuning with structured rewards can boost the model’s ability to make financial judgments. Overall, the work shows that specialist decomposition enhances numerical accuracy but does not reliably aid judgment tasks, while targeted RL adaptation yields measurable gains across both domains.

## Key Contributions  
- [Finding 1] Specialist prompt decomposition improves the aggregate numerical‑task score by 15.8 percentage points across 19 firms and seven regulatory wrappers.  
- [Finding 2] The same decomposition does not reliably improve, and can even reduce, performance on judgment tasks; a single‑agent control fails to reproduce the numerical gain.  
- [Finding 3] Post‑training Qwen3.5‑9B with GRPO using task‑aligned structured rewards raises development‑split scores by 12.0 points and the judgment aggregate by 14.2 points, with positive transfer to unseen firms (+15.2 overall) and regulatory wrappers (+4.3).

## Methodology  
The authors map a 16‑lens European listed‑real‑estate analysis framework onto eight lens‑aligned LLM specialists. They evaluate two prompting strategies—monolithic versus specialist‑decomposed—while fixing the model, source evidence, task instructions, output schema, and scoring. A single‑agent control receives the full framework to serve as a baseline. Additionally, they fine‑tune Qwen3.5‑9B with reinforcement learning (GRPO) using structured rewards that encode both numerical and judgment objectives, then compare the resulting model’s scores on development splits and unseen data.

## Results  
The specialist decomposition yields a 15.8 pp increase in the aggregate numerical score but shows no consistent benefit for judgment tasks; some sub‑tasks even decline. The RL‑fine‑tuned Qwen3.5‑9B improves the development‑split numerical score by 12.0 points and the judgment aggregate by 14.2 points, with gains transferring to unseen firms (+15.2 overall) and regulatory wrappers (+4.3). All three anti‑memorization splits retain positive performance.

## Significance  
These findings reveal a clear division of labor between modular numerical execution (enhanced by specialist prompting) and integrative financial judgment (boosted by targeted RL adaptation). The results suggest that AI systems for regulated finance should combine specialized sub‑agents with fine‑tuned reasoning capabilities to meet both compliance and strategic decision‑making needs.

## Related Concepts  
LLM specialization, reinforcement learning fine‑tuning (GRPO), structured rewards, agent decomposition, numerical tasks vs. judgment tasks, European real‑estate covenant analysis, modular prompting, transfer learning, anti‑memorization evaluation.

## Original Paper Reference

- [Read the original paper](http://arxiv.org/abs/2608.11381v1)