Pith. sign in

REVIEW 4 major objections 5 minor 58 references

Structured multi-agent underwriting beats single-pass LLMs and plain retrieval when facts must be derived or referred.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 16:33 UTC pith:ZNCBDGEA

load-bearing objection Solid applied comparison of agentic vs single-pass underwriting on a released synthetic BOP benchmark; gains are real inside that harness, transfer is the open question. the 4 major comments →

arxiv 2607.07858 v1 pith:ZNCBDGEA submitted 2026-07-08 cs.AI cs.LG

Agentic AI and Retrieval-Augmented Models in Straight-Through Underwriting

Robert Richardson , Josh Meyers , Brian Hartman , David Sandberg This is my paper
classification cs.AI cs.LG
keywords Retrieval-Augmented Generation (RAG)Agentic AIStraight-through underwritingBusiness Owner PolicyHuman-AI interactionsSynthetic data generationInsurance automationActuarial decision workflows
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that straight-through underwriting for small commercial Business Owner Policies is not mainly a pure language-generation problem. Many decisions depend on messy narratives, guidebook rules, derived quantities, and incomplete evidence, so the useful architecture is one that plans retrieval, checks completeness, uses third-party facts when available, and escalates when the evidence cannot support accept or reject. On a synthetic but human-validated set of 635 applications, a multi-agent Agentic RAG pipeline outperforms a single-LLM baseline and a naive RAG system, with the biggest gains on multi-step reasoning and irrecoverable missing-information cases. The practical point is triage with an audit trail: handle clear passes and fails automatically, and force human review when the model would otherwise guess.

Core claim

On a controlled BOP underwriting benchmark, multi-agent Agentic RAG reaches higher overall decision accuracy than single-LLM and naive RAG baselines, and the advantage is concentrated where underwriting is fragile: multi-step rule application and cases that should be referred rather than forced into accept or reject. Structure around retrieval and reflection improves both decisions and explanation alignment, at the cost of higher latency.

What carries the argument

Agentic RAG: a three-agent stateful workflow that routes appetite, diagnoses missing facts with targeted retrieval or third-party lookup, and reflects over multi-step underwriting rules before accepting, rejecting, or referring.

Load-bearing premise

The comparison assumes that gains on this synthetic, human-reviewed guidebook and application set will still matter in real underwriting with noise, conflicting evidence, legacy systems, and institutional judgment.

What would settle it

Run the same three pipelines, with fixed prompts and retrieval setup, on a large sample of production BOP applications labeled by underwriters for accept, reject, or refer; if Agentic RAG no longer wins on multi-step and missing-information cases, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper argues that multi-agent “Agentic RAG” architectures can better support actuarial priorities of transparency, auditability, and human-in-the-loop governance in straight-through underwriting than single-pass LLMs or naïve RAG. Concretely, it builds a synthetic BOP underwriting environment (143-page guidebook, 127 business types, 635 human-validated applications across five scenario templates) and compares three pipelines: single-LLM, naïve RAG, and a three-agent Agentic RAG system with targeted retrieval, third-party checks, multi-step rule evaluation, and explicit referral. Under gpt-5.2, Agentic RAG attains the highest overall accuracy (86.5% vs 77.6%/76.9%), with the largest gains on multi-step reasoning (78.0%) and irrecoverable missing-information referral (84.3% vs 56.7%/53.5%), plus higher rationale cosine similarity at higher latency. The authors frame the work as a controlled proof-of-concept for structured triage rather than production validation, and release data and code.

Significance. If the comparative result holds under fairer controls, the paper is a useful contribution to applied actuarial AI: it maps agentic design choices onto regulated accept/reject/refer workflows, emphasizes escalation and audit trails, and provides a reproducible synthetic benchmark with human label review and public artifacts. Strengths include explicit governance framing (§3.2), label-leakage controls (§5.2–5.3), per-scenario decomposition (Table 2), secondary rationale/latency diagnostics (Table 3), and candid limitations on synthetic transfer (§5.7, §6). The work is more an engineering and evaluation study than a theoretical advance, but it is timely for insurers exploring LLM-based STP with human oversight.

major comments (4)
  1. [Tables 1–2; §4.3–4.4; §5.1.3; Appendix A] Tables 1–2 and §5.5–5.7: the headline gains are concentrated in multi-step reasoning and irrecoverable missing-information, which Appendix A and §4.3–4.4 construct so that the correct path is diagnose missing/derived facts, retrieve third-party or targeted guidebook evidence, then escalate if incomplete—precisely the Agentic RAG state machine in §5.1.3. Baselines receive one pass with static or single-query context and no comparable explicit completeness/referral policy. This risks measuring process match to the evaluation harness rather than general underwriting skill. Please either (i) equip baselines with the same accept/reject/refer decision space, completeness checks, and third-party access under a single-agent or non-orchestrated prompt, or (ii) reframe claims as “structured workflow vs unstructured single-pass” rather than “agentic multi-agent superiority,” and report ablations th
  2. [§5.2; Table 2 (Missing info irrecoverable)] §5.2 states that unparseable agentic responses default to human review. That policy mechanically improves accuracy on the irrecoverable missing-information class (Table 2: 84.3% vs ~55%), where referral is the ground truth. Report accuracy with and without this default, and apply an analogous safe-default policy to baselines so the comparison does not confound architecture with a more conservative failure mode.
  3. [Tables 1–2; §5.4–5.5] Tables 1–2 report point accuracies on n=635 without confidence intervals, McNemar/bootstrap tests, or paired error analysis. Differences of ~9 pp overall and ~28 pp on irrecoverable missing-info are large, but without uncertainty quantification the comparative claim is under-supported for a journal result. Add CIs and appropriate paired tests (overall and per scenario), and discuss effect sizes relative to scenario-class imbalance (accept 20%, reject 54.3%, refer 25.7%).
  4. [Abstract; §5.7; §6] §5.7 and §6 correctly note that the corpus is synthetic and stylized, yet the abstract and discussion present agentic structure as most valuable for fragile STP cases as if the synthetic comparative result generalizes. Keep the internal ranking claim, but tighten abstract/conclusion language to “within this controlled synthetic benchmark” and, if possible, add a small sensitivity check (e.g., noisy narratives, conflicting third-party signals, or rule-threshold perturbations) so the result is less tied to clean scenario templates.
minor comments (5)
  1. [Table 2; §5.7] Table 2: single-issue violation is the one category where Agentic RAG underperforms the single-LLM baseline (81.9% vs 84.3%). Discuss this briefly—possible over-escalation or multi-agent noise—so readers do not over-read “best overall” as uniformly better.
  2. [Table 3; §5.6] Rationale similarity (Table 3) uses embedding cosine similarity as a secondary diagnostic; §5.6 already cautions appropriately. Still, state the embedding model used and whether similarity was computed only on free-text rationales after stripping decision labels.
  3. [§5.1; Figures 1–3] §5.1.1–5.1.3 refer to Figures 1–3 for pipeline diagrams; ensure figure captions state model family, retrieval settings (chunk 1000/overlap 200, FAISS), and temperature=0 evaluation so the figures are self-contained.
  4. [Throughout; References] Minor consistency: “naïve/na¨ıve/naive RAG” and “gpt-5.2” naming should be standardized; also fix occasional encoding artifacts (e.g., W¨ uthrich) in the reference list.
  5. [§3.1] §3 use-case scenarios (pricing, reserving, CAT) are illustrative and lengthy relative to the empirical core. Consider shortening or moving some material to an appendix to keep focus on the BOP experiment.

Circularity Check

0 steps flagged

Empirical pipeline comparison on a designed synthetic benchmark; no derivation reduces to its inputs by construction.

full rationale

This paper is an empirical systems comparison, not a first-principles derivation. The load-bearing claim is measured decision accuracy of three underwriting pipelines on a human-validated synthetic BOP corpus (n=635), with ground-truth labels and rationales withheld from the models (§5.2–5.4; Tables 1–2). Scenario templates (compliant, single-issue, multi-step, recoverable/irrecoverable missing info) are intentionally constructed to stress multi-step composition and referral (§4.3–4.4, Appendix A), which is standard targeted evaluation design, not a circular identity: accuracy is still an independent observation of model behavior, and agentic accuracy remains imperfect (e.g., 78.0% multi-step, 84.3% irrecoverable under gpt-5.2). There is no fitted parameter renamed as a prediction, no self-definitional quantity, and no load-bearing uniqueness or ansatz imported via overlapping-author citation. Human review of synthetic labels further breaks pure self-label circularity. Concerns about transfer to production or evaluation harness alignment with agentic workflows are external-validity / fairness-of-baseline issues, not circular reduction of a claimed derivation.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 3 invented entities

The central comparative claim rests on a synthetic experimental world and standard LLM/RAG tooling assumptions rather than free physical parameters. Load-bearing premises are domain modeling choices (synthetic guidebook/rules, scenario taxonomy, referral as correct when facts are irrecoverable) and implementation choices (prompt/JSON schema, retrieval chunking, foundation models). No new physical entities; invented constructs are the synthetic corpus and the three-agent workflow.

free parameters (4)
  • FAISS chunk size / overlap
    Guidebook split into 1000-character chunks with 200-character overlap; retrieval quality and thus RAG/agentic performance depend on this hand-chosen setup.
  • Prompt and escalation thresholds
    Routing, missing-info diagnosis, and default-to-refer-on-parse-failure are prompt/policy choices that define accept/reject/refer boundaries; paper notes these are adjustable policy choices.
  • Foundation model choice (gpt-4o-mini, gpt-5.2)
    Reported accuracies are conditioned on specific proprietary models and temperature=0; absolute performance is not model-agnostic.
  • Scenario design thresholds (e.g., square-footage/revenue cutoffs)
    Guidebook eligibility thresholds and multi-step templates (50% of 14,000 sq ft, alcohol revenue share, etc.) are author-constructed rules that define ground truth.
axioms (5)
  • domain assumption Synthetic applications, guidebook clauses, and third-party records, after human review, are a valid controlled proxy for comparing underwriting workflow designs.
    Stated throughout §4 and Limitations §5.7/§6; without this, accuracy rankings do not support claims about straight-through underwriting value.
  • domain assumption When decisive information is absent from application and third-party data, the correct system action is human referral rather than forced accept/reject.
    Defines ground truth for irrecoverable missing-info scenarios (§4.3 category 5); drives the largest reported agentic gain.
  • domain assumption LLM outputs under temperature=0 plus structured JSON prompts are sufficiently stable for decision-accuracy comparison.
    Implementation details §5.2; residual API non-determinism and model updates are unquantified.
  • ad hoc to paper Embedding cosine similarity between model rationales and human-validated reasons is a useful secondary diagnostic of explanation quality.
    §5.6 explicitly cautions it is not actuarial correctness; still used as supporting evidence of agentic benefit.
  • standard math Standard transformer LLM + vector retrieval + multi-agent orchestration can implement completeness, evidence, and rule-application checks.
    Background from Lewis et al. RAG and multi-agent literature; used as engineering premise, not proved.
invented entities (3)
  • Synthetic BOP underwriting guidebook (143 pages, 127 business types) no independent evidence
    purpose: Primary retrieval corpus and rule source defining accept/reject/refer ground truth.
    LLM-generated then used as authoritative policy text; no independent insurer manual correspondence claimed.
  • 635-application five-scenario synthetic dataset no independent evidence
    purpose: Benchmark for comparing Single-LLM, Naive RAG, and Agentic RAG pipelines.
    Author-constructed with human validation; released publicly but not drawn from production books of business.
  • Three-agent Agentic RAG underwriting pipeline (routing, clarification, multi-step reflection) no independent evidence
    purpose: Structured workflow claimed to improve multi-step and missing-info decisions and auditability.
    Architectural invention of the paper; evaluated only in the synthetic environment.

pith-pipeline@v1.1.0-grok45 · 21589 in / 3528 out tokens · 39981 ms · 2026-07-10T16:33:04.584292+00:00 · methodology

0 comments
read the original abstract

Artificial intelligence (AI) is beginning to reshape actuarial practice, particularly in domains that require reasoning over unstructured documents, heterogeneous data sources, and regulated decision workflows. Actuaries now face a design space that ranges from traditional rule-based automation to large language models (LLMs), retrieval-augmented generation (RAG), and multi-agent ``agentic'' systems that plan, retrieve, call tools, and reflect. This paper examines how these emerging architectures can support actuarial priorities such as transparency, auditability, and human-in-the-loop governance, with a focus on straight-through decision processes. To make these ideas concrete, we develop and analyze an agentic AI framework for straight-through underwriting of small commercial Business Owner Policies (BOPs). We construct a synthetic but realistic experimental environment and compare three underwriting pipelines: (i) a single-LLM baseline, (ii) a naive RAG system, and (iii) a multi-agent ``Agentic RAG'' pipeline that combines targeted retrieval, third-party data checks, and explicit multi-step rule evaluation. The agentic system performs best overall, with the largest gains in multi-step and missing-information scenarios, where structured retrieval and reflection help the model avoid unsupported straight-through decisions.

Figures

Figures reproduced from arXiv: 2607.07858 by Brian Hartman, David Sandberg, Josh Meyers, Robert Richardson.

Figure 1
Figure 1. Figure 1: shows the simplest underwriting workflow, in which a single large language model performs end-to-end decision-making based on the application data and static policy￾guide context. We use the term “Determine Appetite” because it is the underwriting￾appropriate term for the task, but again emphasize that the LLM is a parameter-based model and does not reason or think in the human sense [PITH_FULL_IMAGE:figu… view at source ↗
Figure 2
Figure 2. Figure 2: extends the baseline by adding a retrieval-augmented generation layer that supplements each application with relevant passages from the synthetic underwriting guidebook [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Agentic Multi-Agent Pipeline. A coordinated three-agent system with targeted retrieval and reflection loops. Agent 1 performs initial appetite routing, Agent 2 diagnoses missing or ambiguous information and retrieves third-party data when avail￾able, and Agent 3 checks multi-step logic before returning the case for reevaluation. Resolved cases are accepted or rejected; unresolved cases are referred for hum… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

58 extracted references · 58 canonical work pages · 12 internal anchors

  1. [1]

    Advances in Neural Information Processing Systems (NeurIPS) 33 , year=

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks , author=. Advances in Neural Information Processing Systems (NeurIPS) 33 , year=

    work page

  2. [4]

    Artificial Intelligence: A Modern Approach , author=

    work page

  3. [5]

    Advances in Neural Information Processing Systems , year=

    Attention is All You Need , author=. Advances in Neural Information Processing Systems , year=

    work page

  4. [6]

    ACM Computing Surveys , year=

    Survey on Hallucination in Natural Language Generation , author=. ACM Computing Surveys , year=

    work page

  5. [7]

    Advances in Neural Information Processing Systems (NeurIPS) , year=

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks , author=. Advances in Neural Information Processing Systems (NeurIPS) , year=

    work page

  6. [8]

    AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation , author=

    work page

  7. [9]

    2019 , url =

    Yeo, Nicholas and Lai, Raymond and Ooi, Min Jyeh and Liew, Jie Yin , title =. 2019 , url =

    work page 2019

  8. [10]

    2024 , institution =

    Carlin, Stephen and Mathys, Stephan , title =. 2024 , institution =

    work page 2024

  9. [11]

    British Actuarial Journal , volume=

    Stochastic Claims Reserving in General Insurance , author=. British Actuarial Journal , volume=. 2002 , publisher=

    work page 2002

  10. [12]

    2008 , publisher=

    Stochastic Claims Reserving Methods in Insurance , author=. 2008 , publisher=

    work page 2008

  11. [13]

    2023 , howpublished =

    Balona, Caesar , title =. 2023 , howpublished =

    work page 2023

  12. [14]

    and Chimedza, C

    Mahohoho, B. and Chimedza, C. and Matarise, F. and Munyira, S. , title =. Open Journal of Statistics , volume =. 2024 , doi =

    work page 2024

  13. [16]

    2025 , institution =

    Balona, Caesar , title =. 2025 , institution =

    work page 2025

  14. [17]

    e-forum of the Casualty Actuarial Society , year =

    Richman, Ronald , title =. e-forum of the Casualty Actuarial Society , year =

    work page

  15. [18]

    2005 , isbn =

    Catastrophe Modeling: A New Approach to Managing Risk , editor =. 2005 , isbn =

    work page 2005

  16. [19]

    The Insurer , year =

    Roberts, Kimberly and Schmidt, Jeffrey and Peterson, Shannon , title =. The Insurer , year =

    work page

  17. [20]

    Statement of Principles on Ratemaking , year =

    work page

  18. [21]

    ASTIN Bulletin , volume =

    Mack, Thomas , title =. ASTIN Bulletin , volume =

    work page

  19. [22]

    Code of Professional Conduct , year =

    work page

  20. [23]

    Principles on Artificial Intelligence , year =

    work page

  21. [24]

    OECD Principles on Artificial Intelligence , year =

    work page

  22. [25]

    Nature Machine Intelligence , volume =

    Jobin, Anna and Ienca, Marcello and Vayena, Effy , title =. Nature Machine Intelligence , volume =

    work page

  23. [26]

    Minds and Machines , volume =

    Floridi, Luciano and Cowls, Josh and Beltrametti, Monica and Chatila, Raja and Chazerand, Patrice and Dignum, Virginia and Luetge, Christoph and Madelin, Robert and Pagallo, Ugo and Rossi, Francesca and Schafer, Burkhard and Valcke, Peggy and Vayena, Effy , title =. Minds and Machines , volume =

    work page

  24. [27]

    Model Risk Management Handbook , year =

    work page

  25. [28]

    ACM Computing Surveys , volume =

    Ji, Ziwei and Lee, Nayeon and Frieske, Rebecca and others , title =. ACM Computing Surveys , volume =

    work page

  26. [29]

    Transactions on Machine Learning Research , year =

    Few-Shot Learning with Retrieval Augmented Language Models , author =. Transactions on Machine Learning Research , year =

    work page

  27. [30]

    International Conference on Learning Representations (ICLR) , year =

    Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection , author =. International Conference on Learning Representations (ICLR) , year =

    work page

  28. [38]

    Asai, A., He, Y., Chen, X., and Hajishirzi, H. (2024). Self-rag: Learning to retrieve, generate, and critique through self-reflection. In International Conference on Learning Representations (ICLR)

    work page 2024

  29. [39]

    Balona, C. (2023). Actuarygpt: Applications of large language models to insurance and actuarial work. SSRN pre-print

    work page 2023

  30. [40]

    Botti, V. (2025). Agentic ai and multiagentic: Are we reinventing the wheel? arXiv preprint arXiv:2506.01463

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [41]

    and Mathys, S

    Carlin, S. and Mathys, S. (2024). A primer on generative ai for actuaries. Technical report, Society of Actuaries Research Institute

    work page 2024

  32. [42]

    Statement of principles on ratemaking

    Casualty Actuarial Society (2019). Statement of principles on ratemaking. CAS Statement. Available from the Casualty Actuarial Society website

    work page 2019

  33. [43]

    Chen, M., Wang, J., Qiu, Y., et al. (2024). Mindagent: Emergent dynamics and autonomy in llm-based multi-agent systems. arXiv preprint arXiv:2403.13890

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  34. [44]

    England, P. D. and Verrall, R. J. (2002). Stochastic claims reserving in general insurance. British Actuarial Journal , 8(3):443--518

    work page 2002

  35. [45]

    Floridi, L., Cowls, J., Beltrametti, M., Chatila, R., Chazerand, P., Dignum, V., Luetge, C., Madelin, R., Pagallo, U., Rossi, F., Schafer, B., Valcke, P., and Vayena, E. (2018). Ai4people—an ethical framework for a good ai society: Opportunities, risks, principles, and recommendations. Minds and Machines , 28(4):689--707

    work page 2018

  36. [46]

    and Kunreuther, H., editors (2005)

    Grossi, P. and Kunreuther, H., editors (2005). Catastrophe Modeling: A New Approach to Managing Risk . Huebner International Series on Risk, Insurance and Economic Security. Springer, New York, NY

    work page 2005

  37. [47]

    Gupta, S., Ranjan, R., and Singh, S. N. (2024). A comprehensive survey of retrieval-augmented generation (rag): Evolution, current landscape and future directions. arXiv preprint arXiv:2410.12837

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  38. [48]

    Advanced Applications of Generative AI in Actuarial Science: Case Studies Beyond ChatGPT

    Hatzesberger, S. and Nonneman, I. (2025). Advanced applications of generative ai in actuarial science: Case studies beyond chatgpt. arXiv pre-print arXiv:2506.18942

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [49]

    and Grave, E

    Izacard, G. and Grave, E. (2022). Few-shot learning with retrieval augmented language models. In Transactions on Machine Learning Research

    work page 2022

  40. [50]

    Ji, Z., Lee, N., Frieske, R., et al. (2023a). Survey on hallucination in natural language generation. ACM Computing Surveys

    work page

  41. [51]

    Ji, Z., Lee, N., Frieske, R., et al. (2023b). Survey on hallucination in natural language generation. ACM Computing Surveys , 55(12)

    work page

  42. [52]

    Jobin, A., Ienca, M., and Vayena, E. (2019). The global landscape of ai ethics guidelines. Nature Machine Intelligence , 1(9):389--399

    work page 2019

  43. [53]

    Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.-t., Rockt \"a schel, T., Riedel, S., and Kiela, D. (2020). Retrieval-augmented generation for knowledge-intensive nlp tasks. In Advances in Neural Information Processing Systems (NeurIPS) 33 . arXiv:2005.11401

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  44. [54]

    Li, Z., Yang, Y., Zhang, H., et al. (2023). Autogen: Enabling next-gen llm applications via multi-agent conversation. arXiv preprint arXiv:2308.08155

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  45. [55]

    Liu, H., Zhao, H., Ren, X., et al. (2024). Agentbench: Evaluating llms as agents. arXiv preprint arXiv:2308.03688

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  46. [56]

    Mack, T. (1993). Distribution‐free calculation of the standard error of chain ladder reserve estimates. ASTIN Bulletin , 23(2):375--393

    work page 1993

  47. [57]

    Madaan, A., Yazdanbakhsh, A., et al. (2024). Self-refine: Iterative refinement with self-feedback in large language models. arXiv preprint arXiv:2303.17651

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  48. [58]

    Mahohoho, B., Chimedza, C., Matarise, F., and Munyira, S. (2024). Artificial intelligence-based automated actuarial pricing and underwriting model for the general insurance sector. Open Journal of Statistics , 14(3):294--340

    work page 2024

  49. [59]

    Park, J., Ouyang, L., Chen, B., et al. (2023). Generative agents: Interactive simulacra of human behavior. arXiv preprint arXiv:2304.03442

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  50. [60]

    Richman, R. (2024). An ai vision for the actuarial profession. e-forum of the Casualty Actuarial Society

    work page 2024

  51. [61]

    Roberts, K., Schmidt, J., and Peterson, S. (2023). Informed decision making with catastrophe risk insights. The Insurer . Guy Carpenter viewpoint article

    work page 2023

  52. [62]

    and Norvig, P

    Russell, S. and Norvig, P. (2016). Artificial Intelligence: A Modern Approach . Pearson, 3rd edition

    work page 2016

  53. [63]

    Shuster, K., Xu, J., Smith, E., et al. (2022). Blenderbot 3: A 175b-parameter, open-domain chatbot that learns to retrieve and refine. In arXiv preprint arXiv:2208.03188

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  54. [64]

    Trivedi, H., Karpukhin, V., Lewis, M., et al. (2023). Rethinking retrieval-augmented generation from first principles. arXiv preprint arXiv:2309.01603

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  55. [65]

    Vaswani, A., Shazeer, N., Parmar, N., et al. (2017). Attention is all you need. In Advances in Neural Information Processing Systems

    work page 2017

  56. [66]

    Wang, Y., Chen, Y., Liang, H., et al. (2024). A survey on multi-agent systems for large language models. arXiv preprint arXiv:2402.01680

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  57. [67]

    W \"u thrich, M. V. and Merz, M. (2008). Stochastic Claims Reserving Methods in Insurance . Wiley, Chichester

    work page 2008

  58. [68]

    J., and Liew, J

    Yeo, N., Lai, R., Ooi, M. J., and Liew, J. Y. (2019). Literature review: Artificial intelligence and its use in actuarial work. Technical report, Society of Actuaries

    work page 2019