acceptodds
Under review as a conference paper at ICLR 2027

RERA: A Multi-Agent Framework with Textual Gradient Optimization for Real Estate Enterprise Risk Attribution

Abstract

Real estate markets are highly susceptible to policy shifts and macroeconomic fluctuations, making accurate risk prediction and attribution essential for enterprise decision-making. Although large language models have shown considerable promise in processing multi-source data, integrating disparate information such as company financial reports, policy documents, and macroeconomic indicators into coherent risk transmission chains remains a persistent challenge. To bridge this gap, we introduce RERA, a multi-agent framework with textual gradient optimization for real estate enterprise risk attribution through large language models (LLMs). RERA employs an analyst-attribution agent architecture, in which analysts review policy, macroeconomic, and financial information to provide structured semantic interpretations and summarized analyses. The attribution agent consolidates insights from multiple analyst agents to identify external factors relevant to the company, establish causal chains, and generate comprehensive risk attribution reports. Its attribution optimization component evaluates generated reports against human expert reports for logical consistency, factual alignment, and quantitative reasonableness. These evaluation results are treated as textual gradients, and the attribution prompt is optimized accordingly to guide the agent toward more expert-aligned and evidence-grounded reports. Experiments conducted on a dataset comprising data from multiple real estate enterprises show that RERA-generated reports outperform both general and finance LLMs in terms of semantic features, attribution accuracy, and readability.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Real Estate Enterprise Risk Attribution, Large Language Models, Multi-Agent Framework, Textual Gradient Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_S6eeOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper focuses on real estate enterprise risk attribution: how policy and macroeconomic changes may lead to financial risks for real estate.
The proposed solution is an LLM-based pipeline, which separate agents analyze different sources of evidence, an attribution agent combines their outputs, and feedback is used to revise its prompt. My main concern is that the methodological contribution remains limited. Also, I find the description of prompt revision procedure is incomplete.

Critical strengths
  • This paper focuses on a practical problem;

  • The authors proposed a technically-reasonable method;

  • The authors conducted many experiments to prove it.

Critical weaknesses
  • The proposed RERA is a LLM-based pipline. The paper does not clearly identify a methodological insight beyond integration of existing techniques. I think the authors overstated ``textual gradient optimization''. Earlier papers, including Automatic Prompt Optimization with “Gradient Descent” and Beam Search and TextGrad: Automatic “Differentiation” via Text, make clear that “gradient” is an analogy. RERA uses the term as if it were a literal part of the method, but it does not calculate gradients or change the model’s weights. It uses feedback to rewrite a prompt. The authors should say this more plainly and avoid making the method sound more technical than it is.

  • Every metric in Table 1 measures similarity to expert reference reports; these metrics cannot independently establish factual correctness or the validity of the proposed causal chains. RERA uses multiple GPT‑4o agents and several rounds of prompt optimization, whereas the reported GPT‑4o baseline appears to be a single model run. A baseline with comparable inference and optimization budgets is needed. In Table 3, removing prompt optimization reduces BERTScore from 0.9226 to 0.8583, below the standalone GPT‑4o baseline of 0.8853 in Table 1. The reported improvement may depend heavily on prompt optimization, and the authors must explain the real contribution of multi-agent systems. Also, the appendix provides representative prompt templates, but important implementation details remain unspecified. For example, after Eq. (5), the paper says that the average score affects how many words can be changed, but it does not explain how that number is chosen. The prompt in the appendix simply takes this number as an input, without saying where it comes from.

Questions for authors

Please refer to the ``Critical Weaknesses'' part.

Reviewer_iXw8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents RERA, a multi-agent framework that explains how policy changes and macroeconomic conditions drive the financial risks of real estate companies. Three specialized agents analyze policy documents, macroeconomic data, and company financials. An attribution agent then combines their findings, traces how the external factors affect the firm’s operations and financial performance, and writes a risk attribution report. The authors also add a prompt optimization loop. An evaluation agent compares generated reports with expert-written ones and flags weaknesses in logical consistency, factual alignment, and quantitative reasonableness. A revision agent then uses this feedback to update the attribution prompt.

RERA is evaluated on 528 firm-year observations from 88 listed real estate companies (2018–2023). It outperforms general-purpose and finance-specific LLMs, including GPT-4o used directly, on BERTScore, SMART, and ROUGE. In a human evaluation, three senior analysts rate RERA’s reports 4.30/5, compared with 4.60 for expert-written reports. In the ablations, both the staged attribution and the prompt optimization help, with prompt optimization having the largest effect. A stronger evaluation-agent LLM generally gives better results.

Critical strengths
  1. The paper goes beyond predicting whether a firm is risky and tries to explain why. It links policy and macroeconomic factors to firm-level financial outcomes. The recognition–reasoning–attribution pipeline makes that chain explicit, which is what analysts and regulators actually need.
  2. The human evaluation by three senior analysts is a useful anchor beyond the automatic metrics. It also shows a real gap to expert reports (4.30 vs. 4.60) rather than claiming parity.
Critical weaknesses
  1. The evaluation does not test whether the attributions are correct. Most improvements are measured with text similarity (BERTScore, SMART, ROUGE). A report can closely resemble the expert reference without identifying the right risk factors or the right relationships between them. The human evaluation gives overall ratings but does not check individual risk transmission links against the evidence. Since the paper’s contribution is attribution, this leaves the main claim largely unverified.
  2. It is unclear where the gains come from. Prompt optimization has the largest ablation effect, but there is no single-agent baseline using the same optimization. We therefore cannot tell how much the multi-agent architecture contributes. In addition, most agents, including the evaluator and reviser, run on GPT-4o. A loop in which the same model generates, critiques, and revises may reinforce its own errors, and no experiment varies the backbone beyond the evaluation agent.
Questions for authors

Could change my assessment:

  1. How were the training and test sets split? Were the final test cases and their expert reference reports kept entirely separate from prompt optimization and candidate selection?
  2. How does RERA compare with a single-agent system using the same optimization procedure?
  3. Did the analysts verify whether each identified risk factor and transmission path is supported by the original evidence? If so, how was this assessed, and what was inter-rater agreement?

Suggestions (not affecting score): The framework covers policy, macro, and firm financials but not counterparties such as competitors, suppliers, and contractors. It also relies on fixed data sources, so how agents would obtain new information when evidence is insufficient is unclear. Discussing these as limitations would help readers judge real-world applicability.

Reviewer_JwdTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

RERA generates real-estate enterprise risk-attribution reports using policy, macroeconomic, and financial analyst agents followed by an attribution agent. An evaluator compares reports with expert references and analyst evidence, and a revision agent updates the attribution prompt under a decaying edit budget. The evaluation covers 528 firm-year observations and reports improvements in reference-based text metrics, with a small human study comparing RERA and expert reports.

Critical strengths

The application is useful, and separating evidence extraction from report construction is a sensible design. The prompts explicitly encourage source preservation, numerical grounding, and withholding unsupported attribution. The domain dataset could be valuable. The paper also includes human evaluation and acknowledges that text similarity alone does not establish factual or causal correctness.

Critical weaknesses
  1. The evaluation does not establish the central improvement in attribution correctness. BERTScore, SMART, and ROUGE measure reference alignment. The human study compares RERA with expert-written reports, but not with competing models, so it cannot validate improvements over those models in factual grounding or risk reasoning. The numerical and causal claims in Table 6 also lack source-to-claim mappings in the example. A blinded, paired comparison with GPT-4o and unoptimized RERA, reporting numerical errors and evidence-supported claims, would be more decisive. I am not asserting that the example numbers are false; their provenance and attribution cannot be audited from the presented evidence.

  2. The generalization protocol is underspecified. Section 4.2 gives a 70% training split without adequately explaining company/year grouping, validation for prompt selection, or the exact separation between optimization and final evaluation examples. Repeated companies, shared annual policies, and report templates make this consequential. A clear split manifest and held-out-company/forward-year tests are needed to distinguish generalization from dependence on familiar entities and templates. This is an unresolved dependence/leakage risk, not a claim that test references were used for training.

Questions for authors
  1. How are companies and years partitioned, which examples select candidate prompts, and is the human-evaluation subset drawn exclusively from the final test set?
  2. Can RERA be compared directly with GPT-4o and unoptimized RERA on blinded numerical-correctness and source-support judgments? What original evidence supports the numerical and causal claims in Table 6?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.