🤖 AI

From Numbers to Judgment: Specialist LLM Agents and Reinforcement Learning for European Listed Real Estate

This paper demonstrates that while decomposing European listed real estate analysis into specialized LLM agents significantly improves numerical task performance, enhancing integrative financial judgment requires targeted parameter adaptation via reinforcement learning (GRPO) rather than prompt-level decomposition alone.

Original authors: Pardis Taghavi, Santosh Bhavani

Published 2026-08-13

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Pardis Taghavi, Santosh Bhavani

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a super-smart robot how to be a financial analyst. In the world of artificial intelligence, these robots are called Large Language Models (LLMs). Think of them as incredibly well-read students who have read almost every book in the library, but they sometimes get confused when asked to do specific, tricky math or to make a complex decision based on a mix of different rules. This paper lives in the corner of science where researchers try to figure out the best way to "train" these robots to handle money and business rules. The big question they are asking is: Is it better to ask the robot to do everything at once, or should we break the job down and ask it to act like a team of different specialists? It's a bit like asking, "Is it better to have one genius chef who tries to cook the whole banquet alone, or a kitchen with a dedicated baker, a dedicated grill master, and a dedicated sauce expert?"

The researchers in this study, Pardis Taghavi and Santosh Bhavani, decided to test this idea using a very specific and tricky subject: European real estate companies. These companies are like a group of neighbors who all live in the same building but follow different sets of rules depending on which country they are in. Some have to pay out most of their profits as dividends, while others don't have to pay anything at all. If you use the wrong rule for the wrong company, the math is wrong, and the investment advice is dangerous. The team wanted to see if splitting the analysis into "specialist" roles helped the robot get the numbers right, and if it helped the robot make the big, complicated judgments about whether a company is safe to invest in.

The Team of Specialists vs. The One-Person Show

To test their idea, the researchers built a system called "Larix." Imagine Larix as a manager who has a massive 16-step checklist for analyzing a real estate company. Instead of handing the whole checklist to one robot and saying, "Do this all by yourself," Larix tries a different approach. It breaks the checklist down into eight smaller teams, or "specialists." One team is the "Financials" expert, another is the "Asset Quality" expert, and another is the "Macro Overlay" expert.

The researchers ran a race with a very powerful robot (a "frontier" model) to see who would win. In the first race, the robot had to do everything alone (the "monolithic" approach). In the second race, the robot was given the full 16-step checklist but still had to do it all alone. In the third race, the robot was assigned to a specific specialist role, like a "Financials" expert, and only given the specific part of the checklist relevant to that job.

The results were surprising and very clear. When it came to the boring, mechanical math—like finding a specific number in a report or doing a simple calculation—the specialist approach was a huge winner. The robot acting as a specialist got the numerical tasks right 15.8 percentage points more often than the robot trying to do everything at once. It's as if the specialist robot put on noise-canceling headphones, ignored all the confusing extra instructions, and just focused on the math.

However, when it came to the "judgment" tasks—where the robot had to look at different pieces of evidence and decide if a company was risky or safe—the specialist approach didn't help at all. In fact, it sometimes made things worse. The robot acting as a specialist scored the same as the robot doing everything alone on these judgment tasks. It turns out that to make a good judgment, you actually need to see the whole picture, not just one narrow slice of it. By splitting the job up, the robot missed the connections between different pieces of information.

The Magic of "Training" the Brain

The researchers then asked a second question: If breaking the job into pieces doesn't help with the hard judgments, can we just "train" the robot's brain to get better at them? They took a smaller, cheaper robot (a 9-billion-parameter model called Qwen3.5-9B) and gave it a special kind of training called "Reinforcement Learning with Group Relative Policy Optimization" (GRPO).

Think of this training like a video game where the robot gets a high score every time it gets an answer right and a low score when it gets it wrong. Over time, the robot learns to adjust its internal "brain settings" to get more high scores. They didn't just ask it to memorize answers; they taught it to understand the rules of the game.

This training worked like magic for the judgment tasks. After the training, the smaller robot's score on the judgment tasks jumped up by 14.2 points. It didn't just get better at the specific companies it practiced on; it got better at companies it had never seen before. When tested on completely new firms, its score went up by 15.2 points overall, and on a specific "covenant stress" test (checking if a company can handle financial trouble), it improved by a massive 40.4 points.

The Big Takeaway

So, what did the paper actually find? It suggests that there is no single "best" way to use AI for financial analysis. It depends entirely on what you are asking the robot to do.

For numbers and math: Breaking the job down into specialists is the way to go. It helps the robot focus and ignore distractions, leading to much more accurate calculations.
For big-picture judgments: Breaking the job down doesn't help. Instead, you need to "train" the robot's brain directly so it learns how to connect the dots between different pieces of information.

The researchers were careful to say that this isn't a magic bullet that solves everything. They showed that just giving the robot a longer list of rules (the full 16-lens framework) didn't fix the problem; it was the specific specialization that helped the math, and the specific training that helped the judgment. They also noted that their results are based on a specific set of 19 companies and a specific type of real estate, so we can't be 100% sure it will work exactly the same way for every type of business in the world. But for the world of European real estate, the lesson is clear: use a team of specialists for the math, and use a well-trained generalist for the big decisions.

Technical Summary: From Numbers to Judgment: Specialist LLM Agents and Reinforcement Learning for European Listed Real Estate

1. Problem Statement

European listed real estate analysis presents a dual challenge for Large Language Models (LLMs): the need for modular numerical operations (extracting regime-specific metrics, calculating implied cap rates) and integrative financial judgments (reconciling accounting, regulatory, and risk dimensions to assess covenant headroom or dividend sustainability).

The domain is characterized by high heterogeneity; legal and reporting regimes (e.g., French SIICs vs. Luxembourg-registered landlords) dictate distinct rules for distribution obligations and performance metrics. A critical failure mode for LLMs is the "fluency trap": extracting a correct figure from a filing but applying the wrong regime's rules, leading to mechanically incorrect conclusions that are difficult to detect due to the model's output fluency.

Existing multi-agent systems often conflate the benefits of decomposition with domain-specific instructions or model differences. This paper seeks to isolate the effect of specialist decomposition (prompt-level) versus parameter adaptation (post-training) on a fixed frontier model and a smaller model, specifically asking:

Does decomposing a task into lens-aligned specialists improve performance on a fixed frontier LLM?
Does this improvement generalize across both numerical and judgment tasks?
Can targeted Reinforcement Learning (RL) post-training improve a smaller model's integrative judgment beyond its zero-shot capabilities?

2. Methodology

2.1 The Larix Framework and Task Formulation

The study utilizes Larix, which maps a 16-lens European listed-real-estate analysis framework to eight lens-aligned specialists (e.g., Financials, Asset Quality, Macro Overlay). The evaluation focuses on the specialist layer, excluding downstream synthesis or position sizing.

The benchmark consists of five tasks across 19 firms spanning seven regulatory wrappers:

Numerical Tasks ( $T_{num}$ ): $T_1$ (Regime-specific metric), $T_3$ (Implied cap rate), $T_6$ (Payout-regime classification). These require localized extraction or deterministic calculation.
Judgment Tasks ( $T_{judg}$ ): $T_2$ (Reconciliation adjustment), $T_5$ (Covenant stress assessment). These require combining multiple disclosures and regime-specific constraints.

2.2 Experimental Conditions

The study compares three prompting conditions using a fixed frontier model (Claude Opus 4.8) and two conditions using a smaller model (Qwen3.5-9B):

Monolithic (Claude-Mono): A generic financial-analysis prompt without specialist decomposition.
Full-Framework (Claude-Full): A monolithic prompt containing the complete 16-lens framework (~4,800 characters).
Specialist Decomposition (Claude-Spec): A deterministic router assigns each task to a specific lens-aligned specialist prompt containing only the relevant subset of the framework.
Zero-Shot Small Model (Qwen-ZS): Qwen3.5-9B using specialist prompts.
RL Post-Training (Qwen-RL): Qwen3.5-9B fine-tuned using Group Relative Policy Optimization (GRPO) with task-aligned structured rewards.

2.3 Scoring and Ground Truth

Evaluation relies on a deterministic, rubric-based scoring system rather than an LLM judge.

Ground Truth: 45 tuples labeled by an external, independent real estate expert; 50 tuples curated from primary sources by framework authors.
Scoring: Fields are scored based on exact matching (with alias normalization) or tolerance-based numerical rules (e.g., ±5% for metrics, ±10% for ratios). Invalid outputs receive zero.
Metrics: Primary aggregates are $S_{num}$ (average of $T_1, T_3, T_6$ ) and $S_{judg}$ (average of $T_2, T_5$ ).

2.4 RL Post-Training Setup

Qwen3.5-9B was post-trained on 200 rows (195 real, 5 counterfactual) using GRPO.

Reward: The deterministic benchmark score ( $R_i = S_t(\hat{y}_i, y^\star_i)$ ).
Optimization: The policy is updated to maximize the task-specific score directly, avoiding separate language-model judges.
Splits: Training data is disjoint from held-out firms, wrappers, and periods to test generalization and anti-memorization.

3. Key Results

3.1 Specialist Decomposition: Task-Dependent Effects

On the 19-firm continental cohort, specialist decomposition yielded a task-dependent outcome:

Numerical Tasks: Performance improved significantly. The aggregate score rose from 78.9% (Monolithic) to 94.7% (Specialist), a gain of 15.8 percentage points. This was driven by improvements in $T_1$ (+26.3 pts) and $T_3$ (+21.1 pts).
Judgment Tasks: Performance did not reliably improve. The aggregate score remained unchanged at 47.4%. In fact, $T_2$ (adjustment identification) saw a slight decrease (-5.2 pts).
Framework Disclosure: Providing the full 16-lens framework to a monolithic model (Claude-Full) did not reproduce the numerical gains. It actually reduced performance on $T_1$ (-10.5 pts) while slightly helping $T_2$ . This indicates that the gain comes from scoped lens ownership (narrowing context), not merely exposing more framework text.

3.2 RL Post-Training: Improving Integrative Judgment

Post-training Qwen3.5-9B with GRPO addressed the limitations of prompt decomposition:

Development Split: The overall score increased from 72.2% (Zero-Shot) to 84.2% (RL), a gain of 12.0 points.
Judgment Tasks: The judgment aggregate ( $T_2 + T_5$ ) improved by 14.2 points (from 58.6% to 72.8%). This is the specific area where prompt-level decomposition failed.
Numerical Tasks: $T_1$ remained at the zero-shot ceiling (99.5%), while $T_3$ and $T_6$ saw improvements.

3.3 Out-of-Distribution (OOD) Transfer

The gains from RL post-training transferred to unseen data:

Unseen Firms: Overall score increased by +15.2 points. The most significant gain was on the covenant stress task ( $T_5$ ), which improved by +40.4 points (from 38.1% to 78.5%).
Unseen Wrappers: Overall score increased by +4.3 points.
Later Periods: Overall score increased by +2.5 points.
Anti-Memorization: Positive transfer was observed across all three anti-memorization splits, confirming the model learned decision policies rather than memorizing specific firm data.

4. Contributions

The paper makes four specific contributions:

Controlled Evaluation: A same-model evaluation of lens-aligned specialist decomposition that holds source evidence, instructions, and scoring fixed, including a full-framework control to isolate the effect of decomposition.
Identification of Trade-offs: Empirical evidence that decomposition improves modular numerical analysis but does not reliably improve (and can reduce) performance on tasks requiring integrative financial judgment.
Structured RL Post-Training: A demonstration that GRPO post-training with task-aligned structured rewards enables a smaller model (9B) to surpass its zero-shot performance on integrative judgment tasks, with gains transferring to unseen firms and regulatory wrappers.
Benchmark Release: The release of a 25-firm, eight-wrapper benchmark with a corrected, per-cell-provenance rubric and a hardened scoring harness.

5. Significance and Claims

The paper concludes that task-dependent division of labor is essential for financial LLMs:

Prompt-level decomposition is effective for simplifying localized, verifiable numerical operations by removing irrelevant context and making intermediate fields explicit.
Targeted parameter adaptation (RL post-training) is necessary for integrative financial judgment, where the model must reconcile evidence across multiple dimensions and constraints.

The authors emphasize that these conclusions are bounded by the benchmark size, wrapper coverage, and the specific model families tested. They do not claim general superiority of small models over frontier models, but rather that targeted specialization (via either prompt engineering or parameter tuning) addresses distinct bottlenecks in financial analysis. The study explicitly avoids claims regarding downstream synthesis, conviction calibration, or position sizing, as those layers were not invoked in the experiments.