Evaluating Foundation Models Beyond Benchmarks: A Conceptual Framework for Context-Sensitive, Risk-Weighted Assessment of AI Systems in Deployment
DOI:
https://doi.org/10.68854/znqp3982Keywords:
foundation models, evaluation, benchmarks, construct validity, data contamination, reliability, risk-weighted assessment, deployment, linguistic equity, AI governanceAbstract
Foundation models are selected, procured, regulated and trusted on the strength of benchmark scores, yet the survey literature on evaluation now documents at length why those scores are a poor guide to how a model will behave in a particular deployment. Benchmarks are contaminated by training data, optimised against by developers, aggregated in ways that hide variance, and constructed in languages and contexts far from those in which many systems are used. This paper develops a conceptual framework for assessing foundation models in terms of fitness for a deployment rather than rank on a leaderboard. Drawing on the recent survey literature on evaluation, on domain studies in healthcare and software engineering, on psychometric treatments of model assessment, on work on reliability and hallucination detection, and on the governance literature on risk-based regulation of general-purpose AI, it identifies four gaps between benchmark performance and deployed behaviour, the construct gap, the population gap, the context gap and the consequence gap, and proposes a framework in which capability measured on context-representative tasks is adjusted for reliability and population coverage and weighted by the cost, exposure and reversibility of failure in the deployment. Six propositions follow, together with a procedure for producing a deployment-fitness assessment and a note on who should produce it. The framework is offered as a tool for institutions and regulators, particularly in African settings where the distance between benchmark conditions and deployment conditions is greatest, and as a research agenda for an evaluation science that measures what deployment requires. It concludes that fitness for a specific deployment, rather than rank on a general leaderboard, is the appropriate basis for selecting, procuring and regulating a foundation model, and that this is most consequential in settings where the distance between benchmark conditions and deployment conditions is greatest.
Downloads
References
Agrawal, S., & Agbeyangi, A. O. (2026). Bridging the AI language divide: A systematic review of NMT and LLMs in low-resource translation. Technologies, 14(8), 518. https://doi.org/10.3390/technologies14080518
Bedi, S., Liu, Y., Orr-Ewing, L., Dash, D., Koyejo, S., Callahan, A., Fries, J. A., Wornow, M., Swaminathan, A., Lehmann, L. S., Hong, H. J., Kashyap, M., Chaurasia, A. R., Shah, N. R., Singh, K., Tazbaz, T., Milstein, A., Pfeffer, M. A., & Shah, N. H. (2025). Testing and evaluation of health care applications of large language models: A systematic review. JAMA, 333(4), 319–328. https://doi.org/10.1001/jama.2024.21700
Carey, S. (2026). Regulating uncertainty: Governing general-purpose AI models and systemic risk. European Journal of Risk Regulation, 17(1), 123– 139. https://doi.org/10.1017/err.2025.10040
Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., Chen, H., Yi, X., Wang, C., Wang, Y., Ye, W., Zhang, Y., Chang, Y., Yu, P. S., Yang, Q., & Xie, X. (2024). A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3), 1–45. https://doi.org/10.1145/3641289
Chen, W., Wang, J., Zhang, Y., Shi, Y., Jiang, T., Zhou, S., Wu, C., Zhang, A., Zhou, C., Wang, M., Liu, X., Hao, X., Wu, Y., Li, Y., Hu, Y., Cao, Z., Lu, Y., Li, M., Ma, Y., … Guo, Y. (2026). On-device large language models: A survey of model compression and system optimization. Artificial Intelligence Review, 59(9), 191. https://doi.org/10.1007/s10462-026-11538-1
Dritsas, E., & Trigka, M. (2026). Deployment-aware compression of large language models for edge intelligence: A system-level survey. IEEE Transactions on Artificial Intelligence, 7(7), 3703– 3718. https://doi.org/10.1109/tai.2026.3656155
Effoduh, J. O. (2026). Decolonizing the governance of artificial intelligence in Africa: From normative mimicry to epistemic sovereignty. Science and Public Policy, 53(2), 245–257. https://doi.org/10.1093/scipol/scag005
Farquhar, S., Kossen, J., Kuhn, L., & Gal, Y. (2024). Detecting hallucinations in large language models using semantic entropy. Nature, 630(8017), 625– 630. https://doi.org/10.1038/s41586-024-07421-0
Gabauer, C. (2026). LLM-based agent systems: A systematic literature review and position on hybrid architectures and evaluation frameworks. Applied Intelligence, 56(10), 325. https://doi.org/10.1007/s10489-026-07340-9
Gao, L. (2026). Colonialism in code: Linguistic hegemony and epistemic exclusion in large language models. AI & Society. Advance online publication. https://doi.org/10.1007/s00146-026-03184-6
Hasan, M. M., Waseem, M., Kemell, K.-K., Rasku, J., Ala-Rantala, J., & Abrahamsson, P. (2026). Assessing small language models for code generation: An empirical study with benchmarks. Journal of Systems and Software, 236, 112815. https://doi.org/10.1016/j.jss.2026.112815
Helm, P., Bella, G., Koch, G., & Giunchiglia, F. (2024). Diversity and language technology: How language modeling bias causes epistemic injustice. Ethics and Information Technology, 26(1), 8. https://doi.org/10.1007/s10676-023-09742-6
Hu, X., Niu, F., Chen, J., Zhou, X., Zhang, J., He, J., Xia, X., & Lo, D. (2025). Assessing and advancing benchmarks for evaluating large language models in software engineering tasks. ACM Transactions on Software Engineering and Methodology. Advance online publication. https://doi.org/10.1145/3786771
Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., & Liu, T. (2025). A survey on hallucination in large language models: Principles, taxonomy, challenges, and open questions. ACM Transactions on Information Systems, 43(2), 1–55. https://doi.org/10.1145/3703155 İbrahimoğlu, N., Demirezen, M. U., & Yıldız, F. (2026). Systematic survey of large language model evaluation: Methods, benchmarks, and comprehensive taxonomic framework. International Journal of Computational Intelligence Systems. Advance online publication. https://doi.org/10.1007/s44196-026-01490-x
Knollmeyer, S., Caymazer, O., Koval, L., Akmal, M. U., Asif, S., Mathias, S. G., & Großmann, D. (2026). Evaluating retrieval augmented generation: A comprehensive review of evaluation dimensions, question types, and application. SN Computer Science, 7(5), 576. https://doi.org/10.1007/s42979-026-05134-x
Li, Y., Liu, P., Wang, H., Chu, J., & Wong, W. E. (2025). Evaluating large language models for software testing. Computer Standards & Interfaces, 93, 103942. https://doi.org/10.1016/j.csi.2024.103942
Mabokela, K. R., Primus, M., & Celik, T. (2025). Advancing sentiment analysis for low-resourced African languages using pre-trained language models. PLOS One, 20(6), e0325102. https://doi.org/10.1371/journal.pone.0325102
Ndomba, G. E., Mswahili, M. E., & Jeong, Y.-S. (2025). Tokenizers for African languages. IEEE Access, 13, 1046–1054. https://doi.org/10.1109/access.2024.3522285
Sheikhi, S., Lovén, L., & Kostakos, P. (2026). Beyond the leaderboard: A survey of the science of evaluation, benchmarking, and methodologies for large language models. IEEE Access, 14, 66493– 66515. https://doi.org/10.1109/access.2026.3686088
Szadeczky, T., & Bederna, Z. (2025). Risk, regulation, and governance: Evaluating artificial intelligence across diverse application scenarios. Security Journal, 38(1), 35. https://doi.org/10.1057/s41284-025-00495-z
Tripathi, S., Alkhulaifat, D., Muppuri, M., Elahi, A., & Dako, F. (2025). Large language models for global health clinics: Opportunities and challenges. Journal of the American College of Radiology, 22(8), 917– 923. https://doi.org/10.1016/j.jacr.2025.04.007
Wang, M., Zhang, Y., Yu, B., Hao, B., Peng, C., Chen, Y., Zhou, W., Gu, J., Zhuang, C., Guo, R., Wang, W., & Zhao, X. (2026). Function calling in large language models: Industrial practices, challenges, and future directions. ACM Computing Surveys, 58(9), 1–37. https://doi.org/10.1145/3788284
Wang, Y., Huang, J., Du, L., Guo, Y., Liu, Y., & Wang, R. (2025). Evaluating large language models as raters in large-scale writing assessments: A psychometric framework for reliability and validity. Computers and Education: Artificial Intelligence, 9, 100481. https://doi.org/10.1016/j.caeai.2025.100481
Woesle, C., & Buettner, R. (2026). A systematic literature review of prompt injection attacks in LLM-integrated systems. IEEE Access, 14, 86797–86818. https://doi.org/10.1109/access.2026.3700492
Yilma, K., & Wodajo, K. (2026). Strategy as governance: The governance of AI in Africa. Science and Public Policy, 53(2), 236–244. https://doi.org/10.1093/scipol/scag011
Downloads
Published
Issue
Section
License
Copyright (c) 2026 IIAR Journal of Artificial Intelligence and Emerging Technologies
This work is licensed under a Creative Commons Attribution 4.0 International License.
