CAS Funds Research to Fine-Tune LLMs for P&C Actuarial Reasoning: What the Dual RFP Means for the Profession

The CAS issued two Requests for Proposals in February 2026: up to $80,000 for "Adapting Large Language Models for Specialized P&C Actuarial Reasoning," and up to $40,000 for "Leveraging LLMs in Unstructured Claims Data."

It is the first time a major actuarial body has funded the construction of domain-specific AI rather than the evaluation of off-the-shelf tools. The distinction is written into the RFP, and it is the whole point of the exercise.

Key Takeaways

• Up to $80,000 and $40,000 fund two tracks: reasoning that reflects professional standards, and extraction that turns adjuster notes, medical records and call transcripts into modelling variables.
• The CAS explicitly excludes prompt-testing research. It wants "the mechanisms by which LLM behavior is shaped for actuarial use," with proposals measured against a clearly articulated baseline.
• Reserve estimation error fell from 6.5% to 4.0% in the first funded output, a severity-segmented analysis extracting 36 actuarial variables from claims documents.
• 11.3% baseline hallucination in underwriting decision support, reduced to 3.8% by an adversarial self-critique architecture. Both figures are far too high for a signed statement of actuarial opinion.
• Mozilla Public License 2.0 governs the deliverables, which means the profession is funding a commons that any carrier can fold into a proprietary stack.

What the CAS Is Actually Funding

The first RFP draws a line most published work on actuarial AI sits on the wrong side of. Testing whether a general model can solve exam problems or generate a triangle is not eligible. Investigating how model behavior is shaped for actuarial use is.

The eligible domains cover the core of casualty practice: pricing and ratemaking, reserving and loss development, capital modelling and stress testing, reinsurance and portfolio risk, and emerging risk assessment.

The named technical approaches show the working group knows the state of the art. Parameter-efficient fine-tuning through LoRA and QLoRA sits alongside instruction tuning, retrieval-augmented generation, training domain models from scratch, hybrid systems pairing an LLM with a deterministic actuarial model, and reinforcement learning from human feedback where the feedback comes from credentialed actuaries. That last one has no published precedent.

Deliverables are systems, not papers alone: a peer-reviewed paper, demonstrated use cases, a reproducible architecture description, and a GitHub repository of code, pipelines and artifacts under the Mozilla Public License 2.0.

The evaluation criteria carry the requirement that matters most. Proposals must beat a clearly articulated baseline and must demonstrate stability of outputs, which addresses the failure mode that keeps generative models out of filings: the same prompt producing materially different reserve estimates on consecutive runs.

The 6.5% to 4.0% Is Where This Becomes a Reserving Question

The claims data track has already produced a result, and it is specific enough to price.

A June 2026 arXiv paper by Lieberthal et al., funded by the CAS AI Working Group, describes a two-stage architecture extracting 36 actuarial variables from synthetic FHIR-based claims data and real claims documents. Two independent clinical reviewers scored 14 core variables above 4.0 on a five-point scale. The severity-segmented analysis built on those variables cut reserve estimation error from 6.5% to 4.0%.

That is the mechanism the funding is aimed at. Loss development currently runs on transaction codes and aggregate severity. Litigation involvement signalled in an adjuster note, treatment trajectory in a medical record, and settlement posture in a recorded call are all known to the claim file and invisible to the triangle. Moving estimation error from 6.5% to 4.0% is what closing part of that gap was worth on this data.

The barrier is the failure rate on the other side. Roy and Singh measured 11.3% baseline hallucination in underwriting decision support, reduced to 3.8% by an adversarial self-critique layer. An appointed actuary signing a Statement of Actuarial Opinion cannot append a note that the underlying model fabricates data 3.8% of the time, so 3.8% is not a partial success here, it is a disqualification.

Mario DiCaro, FCAS, who chairs the working group and runs capital modelling and analytics at Tokio Marine, described the behavior directly: "It turns out it's extremely good at reading. Unfortunately, when it doesn't find what it's looking for, it would make stuff up."

Domain adaptation attacks that architecturally rather than through prompting. Fine-tuning grounds the probability distribution in actuarial terminology. Retrieval anchors outputs in CAS papers and filed rate manuals. Hybrid routing sends the arithmetic to a deterministic model and leaves the LLM the language. The published ceiling is INS-S1, a March 2026 insurance model family reporting a 0.6% hallucination rate against a 39,000-sample benchmark.

Open Research Against Proprietary Scale

The combined budget is $80,000 and $40,000. That sits against carrier programmes running orders of magnitude larger, and the asymmetry cuts in two directions.

AIG's multi-agent system runs 30-hour autonomous underwriting cycles inside a domain ontology that constrains agent behavior. State Farm's Navi reaches 19,200 agent offices on the mutual's own policy and claims data. Allstate's ALLIE cut billing escalations 50% on a stack built rather than bought. INS-S1's synthesis system is likewise controlled by its authors.

Every one of those is proprietary. The CAS deliverables are not, which is the intended difference and also the unresolved one. MPL 2.0 lets any carrier incorporate CAS-funded work into a closed system without contributing anything back at the architecture level. The profession funds the shared layer; the parties with the capital to build on it capture the returns.

Mike McGavick, formerly of XL Group, argued for small insurance-focused models at the CAS Seminar on Reinsurance and put the dependency plainly: "If I use the word model, it's the actuary in the end that's going to be telling the rest of us what can be trusted and what cannot." That is a validation role, and it is available whether or not the profession builds anything.

The scale gap shapes what the money can buy. It cannot fund a competitive model. It can fund the evaluation apparatus, the baselines, the stability tests and the reproducible pipelines that let an actuary say which vendor implementation clears a professional threshold. Whether that capacity arrives in time is a question of the profession's own throughput: the SOA and CAS competence ladder covers using these tools, and neither exam pathway yet covers validating, monitoring or governing a domain-adapted one.

Sources

• CAS, "2026 Request for Proposals: Adapting Large Language Models (LLMs) for Specialized P&C Actuarial Reasoning," February 2026 - casact.org
• CAS, "Deadline Extended! 2026 Request for Proposals: Adapting LLMs for Specialized P&C Actuarial Reasoning" - casact.org
• CAS, "Leveraging LLMs in Unstructured Claims Data: The CAS Issues a Research RFP for an Actuarial Solution" - casact.org
• CAS, "Using Artificial Intelligence as Actuarial Tools Spurs CAS Research" - casact.org
• CAS, "AI Tools and Resources" - casact.org
• Actuarial Review, "Behind the Scenes at the CAS Artificial Intelligence Working Group," November 2025 - ar.casact.org
• Lieberthal, R. et al., "Leveraging LLMs for Unstructured Claims Data Analysis," arXiv 2606.06089, June 2026 - arxiv.org
• Roy, J. and Singh, S., "Agentic AI for Commercial Insurance Underwriting with Adversarial Self-Critique," arXiv 2602.13213, January 2026 - arxiv.org
• "An Industrial-Scale Insurance LLM Achieving Verifiable Domain Mastery and Hallucination Control without Competence Trade-offs" (INS-S1), arXiv 2603.14463, March 2026 - arxiv.org
• Balona, C., "ActuaryGPT: Applications of Large Language Models to Insurance and Actuarial Work," British Actuarial Journal, vol. 29, 2024 - cambridge.org
• Carrier Management, "Exclude It, Harness It, Get Greedy: McGavick's Take on Insurers' AI Playbook," June 2026 - carriermanagement.com
• ASOP No. 56, Modeling - actuarialstandardsboard.org
• American Academy of Actuaries, "Actuarial Professionalism Considerations for Generative AI," 2026 - actuary.org