AI Language Models Show Hidden Phase Transition in Truthfulness
A new study from arXiv reveals a phase transition in large language models where reasoning and truthfulness shift from anticorrelation to cooperation at a critical scale of approximately 3.5 billion parameters. Researchers measured coupling across 63 base models from 16 families, finding that below the critical threshold, capabilities conflict, while above it they align. The transition point varies by architecture, data curation, and training recipe. Notably, curated training eliminated the coupling dip in Qwen models, Gemma-4 achieved high coupling through distillation, and Phi matched larger models via data curation alone. Width normalization removed anticorrelation across all tested families.
Key facts
- Phase transition in LLMs at ~3.5B parameters
- 63 base models from 16 families tested
- Reasoning and truthfulness anticorrelate below critical scale
- Architecture, data curation, training recipe shift critical scale
- Curated training improved Qwen coupling from 0.025 to 0.830
- Gemma-4 at 4B achieves coupling of 0.871
- Phi at 1B matches web-trained coupling at 10B
- Width normalization eliminates anticorrelation
Entities
Institutions
- arXiv