2026 Request for Proposals: Adapting Large Language Models (LLMs) for Specialized P&C Actuarial Reasoning

The Casualty Actuarial Society’s Artificial Intelligence Working Group is seeking research proposals that examine how Large Language Models (LLMs) can be deliberately adapted to support core actuarial reasoning in property-casualty insurance. While prior research has explored the use of general-purpose LLMs as actuarial tools—capable of performing calculations or responding to prompts—this RFP focuses on the mechanisms by which LLM behavior is shaped for actuarial use.

Proposals Due: Friday, March 27, 2026

We are interested in research that moves beyond “out-of-the-box” or prompt-driven applications to investigate how models can be structured, trained, or constrained to reflect actuarial logic, data structures, and judgment, and that results in a practical, reproducible system or workflow suitable for adoption by actuarial practitioners. Proposed approaches may include fine-tuning existing models, structured context engineering, retrieval-augmented or modular architectures, training domain-specific models from scratch, or hybrid systems that combine LLMs with traditional actuarial models. Regardless of approach, proposals should emphasize interpretability, auditability, and stability of actuarial outputs, and clearly articulate how the chosen adaptation method differs from generic LLM usage.

Research Problem

Large Language Models have rapidly entered actuarial workflows, most commonly as tools for code generation, document parsing, data extraction, or general analytical assistance. While these applications provide immediate productivity gains, they largely rely on general-purpose models that are not adapted to actuarial data structures, methodologies, or decision contexts.

This RFP seeks research that investigates how LLMs can be adapted or otherwise specialized for actuarial work. The CAS is particularly interested in research that demonstrates how such adaptation enables new or improved actuarial capabilities relative to a clearly articulated baseline. Improvements may include, but are not limited to, enhanced interpretability, greater consistency with actuarial reasoning, or deeper integration with established actuarial models and processes.

Proposals should explain the chosen baseline—such as an out-of-the-box LLM, a prompt-based approach, or an existing actuarial workflow—and describe how the proposed adaptation provides meaningful added value. The proposed research should address both (i) the practical feasibility of adapting and deploying LLM-based systems in actuarial workflows and (ii) the extent to which such adaptation results in meaningful learning or behavioral change, while also discussing interpretability, validation, and appropriate use in actuarial practice.

Scope of Research

Proposals are open to any major P&C actuarial domain. The objective is to produce a system in which an LLM is meaningfully adapted to perform better on a specific actuarial task relative to a clearly articulated baseline. Within these domains, the CAS is particularly interested in research that focuses on how LLM behavior is adapted to actuarial structure and judgment, rather than on isolated data extraction or automation tasks.

Potential areas of focus include, but are not limited to:

• Pricing and ratemaking
• Reserving and loss development analysis
• Capital modeling, stress testing, or scenario analysis
• Reinsurance analysis and portfolio risk management
• Emerging risks assessment frameworks
• Other actuarial analytics relevant to P&C insurance

Researchers are encouraged to define a clear actuarial problem and demonstrate how an adapted LLM contributes meaningfully to its solution. Solutions may involve LLMs producing actuarial outputs directly, generating structured inputs to traditional actuarial models, or operating within hybrid modeling frameworks.

The CAS does not prescribe specific fine-tuning techniques, model architectures, or platforms. Proposals may explore a range of approaches, provided the role of the LLM and its actuarial relevance are clearly articulated.

Proposal and Work Product Requirements

Researchers will develop a paper and accompanying materials documenting the design, implementation, and evaluation of an LLM-based system adapted for an actuarial use case. The research should:

• Clearly define the actuarial problem being addressed and explain why adaptation of an LLM is appropriate.
• Describe the overall solution architecture, including how the LLM interacts with data, actuarial methods, and other models.
• Explain the adaptation approach at a conceptual level. Adaptation may include, but is not limited to, parameter-based methods (e.g., full fine-tuning, LoRA, QLoRA, instruction tuning), prompt- or context-based approaches, retrieval-augmented generation (RAG), full model training from scratch, or hybrid architectures. The choice of approach should be justified based on data availability, actuarial relevance, and governance considerations.
• Identify the potential for any reinforcement learning from human feedback (RLHF) where the human feedback is an actuary.
• Use a benchmark dataset to demonstrate the success, or lack thereof, from fine-tuning.
• Address interpretability and explainability of model outputs in an actuarial context.
• Describe the evaluation and validation framework, including performance metrics appropriate to the actuarial task (e.g., predictive accuracy, calibration, stability, or operational efficiency). Proposals should specify a transparent benchmarking approach against a relevant baseline and, where feasible, distinguish performance gains attributable to the LLM component from those arising from other elements of the solution.
• Include a discussion of governance, appropriate use, and practical implementation considerations.

Researchers may use real, anonymized, or synthetic data, provided the data is sufficient to demonstrate the methodology and findings. The CAS may support researchers by facilitating connections with industry partners or identifying appropriate public or synthetic data sources.

Final Deliverables

The goal of this research is to provide the CAS membership with an implementable and reproducible framework for adapting LLMs to actuarial use cases. Final deliverables should include:

• A research paper documenting data preparation, the actuarial use case, the adaptation approach, results, limitations, and lessons learned.
• Demonstration of actuarial use cases illustrating both the effectiveness and limitations of the adapted approach.
• A clear description of the system architecture and implementation workflow, sufficient for a technically proficient actuary to reproduce and adapt the approach using comparable data.
• Code and supporting materials hosted in a GitHub repository documenting the system architecture, data pipelines, adaptation approach, and evaluation framework. Where applicable, researchers are expected to release relevant model adaptation artifacts (e.g., fine-tuned weights, parameter-efficient adapters, prompts, configuration files, or equivalent components) under an open-source license.
• An executive summary (1–2 pages) suitable for a blog post or magazine article, written for a non-technical audience and highlighting key insights and implications for actuarial practice.

Submission, Selection, and Compensation

Interested researchers should submit proposals by Friday, March 27, 2026, including:

• A detailed outline of the proposed research and deliverables
• A description of the actuarial problem and proposed approach
• Estimated out-of-pocket expenses (e.g., cloud computing, model usage, software licenses)
• Expected compensation for labor
• Resumes of the researcher(s), demonstrating relevant qualifications

Proposals and questions should be submitted to Elizabeth Smith, CAS Director of Publications and Research, at esmith@casact.org and Heather Davis, CAS Research Manager, at hdavis@casact.org with “LLM AI Research Proposal” in the subject line.

Receipt of proposals will be acknowledged. Respondents who are not selected will be informed shortly thereafter. A CAS contract will be awarded to the proposal that, in the judgment of the CAS Artificial Intelligence Working Group, best meets the objectives of this RFP. If no proposal meets the requirements, no contract will be awarded.

Compensation will be commensurate with the scope of work. Total costs should not exceed $80,000. (Note: A portion of the funding should be used to cover usual and customary travel expenses to present the paper at a CAS-sponsored seminar or meeting if such a presentation is anticipated.) 

Presentation, Ownership, and Publication of Report

As a condition of selection, all rights, title, and interest in the final work product will be owned by the CAS. Authors will receive appropriate authorship credit in any publication. The CAS may publish the work in whole or in part across its publications and digital platforms. Publishing outside the CAS requires prior permission and acknowledgement of original publication.

To support adoption, code and data will be hosted in the CAS GitHub repository under the MPL 2.0 license. Authors must upload the final paper through the CAS ScholarOne system and make reasonable efforts to present the work at a CAS meeting or seminar.

Timeline and Deadlines

• Proposal Deadline: Friday, March 27, 2026
• Researchers Notified: Friday, April 27, 2026
• Interim Progress Report & Executive Summary: Friday, July 26, 2026
• Final Paper & Deliverables: Friday, August 31, 2026

About the Casualty Actuarial Society and the Artificial Intelligence Working Group

The Casualty Actuarial Society (CAS) is a leading international organization for credentialing and professional education. Founded in 1914, the CAS is the world’s only actuarial organization focused exclusively on property and casualty risks and serves over 11,000 members worldwide. CAS members are experts in property and casualty insurance, reinsurance, finance, risk management, and enterprise risk management. Professionals educated by the CAS empower business and government to make well-informed strategic, financial and operational decisions. 

The CAS Artificial Intelligence Working Group was established to fulfill the CAS mission to “advance the body of knowledge” on a technology that is transforming actuarial practice. Its objective is to encourage the exploration of Artificial Intelligence in actuarial practice through research to help educate members, build knowledge, provide practical insight and establish CAS as thought leaders.