قيّم القرار قبل شراء القدرة

Liang et al.; TMLR · 2023-10-01

Peer-reviewed research · Abstract, methodology overview

Historical multi-model benchmark; its 30 models and 16 core scenarios do not represent current systems or enterprise outcomes. Seven metrics were evaluated where feasible, not for every scenario-model combination.

Bean et al.; NeurIPS · 2025-11-03

Peer-reviewed research · Abstract

A systematic review of benchmark validity, not an experiment proving that all benchmark results are unusable or predicting a deployment failure rate.

Mitchell et al.; FAT*2019 · 2019-01-14

Peer-reviewed research · Abstract

A documentation proposal; completing a model card does not validate the complete deployed service.

Zheng et al.; NeurIPS · 2023-12-24

Peer-reviewed research · Abstract

Agreement concerns human preferences in the studied settings and model generation; it does not establish factual, legal or clinical correctness.

Kapoor et al.; arXiv · 2024-07-01

Research preprint · Abstract

Research preprint on evaluation design and agent costs; it supplies no universal enterprise return-on-investment multiplier.

Sinha et al.; ICLR 2026 · 2026-03-13

Peer-reviewed research · Abstract, version3

Controlled long-horizon tasks and specified models; findings cannot be extrapolated directly to a business agent or treated as a current model ranking.

NIST · 2024-07-26

Voluntary guidance · Publication abstract

Voluntary lifecycle guidance, not a certification, legal opinion or assurance that a particular deployment is safe.

Rabanser et al.; arXiv · 18 February 2026, v1; rendered text marked preprint as of 24 August 2026

Empirical research preprint · Abstract; Sections 3–4; Section 6 limitations

Two benchmarks, scaffold dependence, subjective metric choices, LLM safety judges; operational rather than adversarial robustness. Abstract and conclusion differ on sub-metric count; no count is asserted here.