Lying Is Just a Phase: The Hidden Alignment Transition in Language Model Scaling

Researchers discover that language models exhibit a phase transition between reasoning and truthfulness capabilities at around 3.5B parameters, where smaller models show anticorrelated capabilities while larger ones show cooperation. This hidden alignment transition is invisible to standard loss curves but can be diagnosed from public benchmarks alone, and a proof-of-concept intervention demonstrates that adding a truth-direction vector can correct misaligned outputs without retraining.

This research identifies a fundamental structural phenomenon in large language model scaling that challenges conventional understanding of how capabilities emerge. Rather than capabilities developing smoothly with scale, the study reveals a sharp regime change where below approximately 3.5B parameters, reasoning ability and truthfulness work against each other (r = -0.989), but above that threshold they cooperate positively (r = +0.72). This phase transition persists across 63 models from 16 different families, suggesting it reflects something inherent to language model architecture rather than implementation details.

The discovery carries profound implications for model development and alignment strategy. The researchers identify an output-projection bottleneck as the mechanism driving this phenomenon, validated through width normalization experiments that eliminate the anticorrelation entirely. Critically, the phase transition can be shifted through architecture, data curation, and training methodology—Gemma-4 at 4B parameters achieves large-model-scale coupling through distillation, while Phi-1B reaches 10B-equivalent performance through curated data alone. This reveals that parameter count is not destiny for model quality.

The practical intervention is particularly significant: injecting a single truth-direction vector at the identified layer corrects 60% of misaligned outputs without weight modification or retraining. This surgical, per-inference approach suggests that model misalignment may be partially separable from core capabilities—a finding that could accelerate safe AI development by enabling rapid alignment adjustments post-deployment.

For AI practitioners and safety researchers, this work provides both diagnostic and interventional tools released as open-source software. The ability to predict phase transitions from public benchmarks alone democratizes alignment research across resource-constrained teams. Future research should focus on understanding whether this bottleneck architecture is optimal or an artifact worth redesigning.

• →Language models undergo a sharp phase transition around 3.5B parameters where reasoning and truthfulness shift from anticorrelated to cooperative behaviors.
• →The phase transition can be diagnosed using only public benchmark scores without access to model internals, enabling widespread analysis.
• →Model architecture, data curation, and training methods independently shift the phase transition point, making parameter count insufficient to predict capability coupling.
• →A single truth-direction vector injection corrects 60% of misaligned outputs without retraining, demonstrating exploitable structural vulnerabilities in the bottleneck.
• →Open-source steering tools and diagnostic dashboards are now available to identify and potentially correct alignment issues across any open-weight model.