Response drift across frontier large language models

2026-07-25 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

A study involving 47 participants evaluated ten frontier large language models (LLMs) across 62 multi-domain questions, yielding 29,140 human assessments. It reveals that all LLMs exhibit "response drift," producing outputs deviating from expert-validated references. While two models, Claude (47.0% deviation) and Gemini (49.4% deviation), performed significantly better, eight others converged on a statistically indistinguishable ceiling (78–81% deviation). This drift is domain- and question-dependent, and automated similarity metrics explain less than 2% of human judgments, underscoring the need for human-centered evaluation. The evaluation period was December 2025 to February 2026.

Key takeaway

For AI Scientists and ML Engineers evaluating LLMs for knowledge-intensive applications, you must move beyond automated benchmarks. This study demonstrates that human-centered, reference-based fidelity evaluations are critical to accurately assess model performance and identify significant response drift. Relying solely on automated metrics will obscure substantial quality differences, leading to suboptimal model selection and deployment decisions.

Key insights

All frontier LLMs exhibit significant response drift, which automated metrics fail to detect, necessitating human evaluation.

Principles

• Response drift is universal across LLMs.
• Automated metrics do not capture human-perceived fidelity.
• Model choice primarily determines drift magnitude.

Method

A fully crossed human evaluation with 47 participants assessing 10 LLMs on 62 multi-domain questions against expert references using a 5-point Likert scale.

In practice

• Implement reference-based fidelity assessments.
• Consider domain-specific LLM routing.
• Avoid sole reliance on automated NLP metrics.

Topics

• Large Language Models
• Human Evaluation
• Response Drift
• Model Fidelity
• Automated Benchmarks
• Evaluation Metrics

Best for: AI Architect, AI Engineer, NLP Engineer, AI Scientist, Director of AI/ML, Machine Learning Engineer

Related on AIssential

• Human Evaluation of Large Language Models: How Good is Hugging Face’s BLOOM? — Human evaluation is crucial for assessing large language models' real-world performance beyond flawed academic benchmarks.
• Accuracy, Stability, and Repeated-Run Reliability of Large Language Models on Deterministic Programming Tasks — LLM evaluation for deterministic tasks requires repeated-run stability analysis to complement conventional single-run accuracy reporting.
• LLMs Can Better Capture Human Judgments--With the Right Prompts — Asking better questions to LLMs through specific prompting strategies improves their ability to capture human judgments and response distributions.
• Cross-Lingual Response Consistency in Large Language Models: An ILR-Informed Evaluation of Claude Across Six Languages — ILR-informed expert judgment reveals interpretable, domain-dependent cross-lingual variation in LLM outputs.
• Comparing performance of LLMs is not very interesting — Quantitative LLM performance comparisons quickly become obsolete; qualitative insights offer lasting value.
• # I Got Tired of Saying “This Response Is Better” Without Being Able to Prove It — Combining human domain expertise with automated pattern detection improves LLM evaluation accuracy and provides objective vocabulary for subjective judgments.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.