Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Jul 24·5 min read
Response drift across frontier large language models
TL;DR
Frontier large language models (LLMs) universally exhibit response drift, producing outputs that deviate from expert references. The study reveals significant variation in drift magnitude across different models and domains, highlighting the need for human-centered evaluation.
✦ Why It Matters
Evaluate your LLM outputs against expert references to identify and mitigate response drift.
Key Takeaways
- Assess LLM outputs with expert references to identify drift.
- Understand that drift varies significantly across models and domains.
- Use human evaluations for a more accurate assessment of model performance.
- Be cautious of relying solely on automated metrics for model evaluation.
Full Summary
Response drift refers to the tendency of large language models (LLMs) to generate outputs that diverge from established expert-validated references. A comprehensive evaluation involving 47 participants assessed 62 questions across ten frontier LLMs, resulting in 29,140 independent assessments.
All models demonstrated response drift, but the extent varied significantly; eight models showed a high deviation of 78-81%, while two models had lower deviations of 47-49%. Drift patterns were not uniform, differing across six domains and the specific questions posed.
Notably, pairwise correlations among the models with high drift exceeded 0.85, indicating similar drift profiles. Automated similarity metrics were ineffective, explaining less than 2% of the variance in human judgments, underscoring the importance of human evaluation in understanding model behavior.