Larger and more instructable language models become less reliable.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Current large language model (LLM) development focuses on scaling (increasing size, data, computation) and shaping (fine-tuning, human feedback).
• Despite advancements, larger and more "instructable" LLMs may exhibit reduced reliability and unpredictable error patterns.

Purpose of the Study:

• To investigate the relationship between task difficulty, model avoidance, and prompting stability in various LLM families.
• To assess how scaling and shaping impact LLM reliability and error predictability, particularly in high-stakes applications.

Main Methods:

• Analysis of difficulty concordance between human participants and LLMs.
• Evaluation of task avoidance and prompting stability across different LLM families.
• Comparison of error types and detectability between early and scaled-up/shaped-up LLMs.

Main Results:

• LLMs find easy tasks easy, but scaled models do not guarantee error-free or easily supervised low-difficulty zones.
• Scaled LLMs, unlike earlier models, frequently provide plausible but incorrect answers, often on difficult questions missed by human supervisors.
• While scaling and shaping improve response stability to varied phrasing, unpredictable errors persist across difficulty levels.

Conclusions:

• Scaling and shaping LLMs do not inherently improve reliability or predictability of errors.
• A paradigm shift in AI design is needed, focusing on predictable error distributions for critical applications.
• Further research is required to ensure AI safety and trustworthiness, especially in high-stakes domains.