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

## Alignment Phase Transition: What 63 Models Reveal About Small-Scale Deception

### 1. The Core Phenomenon

Loss curves see nothing. That is the starting point of arXiv:2605.18838, which analyzes 63 base models from 16 families and detects a regime invisible to standard metrics: below a critical threshold N_c ≈ 3.5B parameters (bootstrap 95% CI: [2.9B, 13.4B]), reasoning and truthfulness anticorrelate. A model that reasons better lies more — or more precisely, its reasoning capability and its tendency to produce true statements evolve in opposite directions. Above the threshold, the relationship inverts: both capabilities cooperate.

This anticorrelation is not a benchmark artifact. The authors reproduce it across 38 of 40 internally tested models, with zero competing attention heads detected, pointing to a structural bottleneck at the output-projection layer. Width normalization eliminates the anticorrelation across all tested families — a surgical intervention that confirms the architectural hypothesis.

### 2. N_c Is Not a Constant — It Is a Design Variable

The 3.5B figure is a median, not a law. The study demonstrates that three levers shift N_c independently:

**Data curation**: Phi at 1B parameters matches the coupling of a web-trained model at 10B — a 10:1 efficiency ratio in favor of curation. Across Qwen generations at matched scale, curation shifts coupling from 0.025 to 0.830, a 33× increase.

**Architecture + distillation**: Gemma-4 at 4B achieves coupling of 0.871, characteristic of standard-trained models at 13B+. Gemma-4 compresses roughly 3× the scale required to exit the anticorrelated regime.

**Training recipe**: independent effect from the two above, quantified but not detailed in the abstract.

These three levers are orthogonal. A practitioner optimizing only model size can remain stuck in the anticorrelated regime well beyond 3.5B if data or architecture do not follow.

### 3. The Diagnostic Tool and Operational Implications

The diagnostic requires no access to model internals — only public benchmark scores across a model family. This is a critical point: any team can audit its own model family without interpretability infrastructure. The dashboard (zehenlabs.com/cape/) provides: - Coupling phase diagnosis - Concrete intervention suggestions (curation, width, benchmark rotation) - ODE scaling predictions (validated on Llama-2 at 5.6% error) - Eigenstructure analysis - Frontier diagnostics (r = +0.72 across 34 models, 10 labs)

The sparse-regression ODE cross-predicts held-out Llama-2 at 5.6% error — sufficient precision for compute allocation decisions on unseen families.

### 4. Potential Losers and Blind Spots

**Teams scaling without curating**: if N_c can be pushed down to ~1B through curation (Phi case), organizations investing heavily in 3–7B pretraining runs on raw web data may be operating in the anticorrelated regime unknowingly. Their models reason better but become less factually reliable — exactly the profile that produces confident hallucinations.

**Evaluators using isolated benchmarks**: the study recommends benchmark rotation as an intervention. Single-benchmark evaluations do not capture inter-capability coupling. A model can score high on reasoning and low on truthfulness without standard evaluation flagging it as problematic.

**Pure-scaling emergence assumptions**: the demonstration that Gemma-4 at 4B reaches the coupling of 13B+ standard models empirically invalidates the idea that reasoning/truthfulness cooperation is an emergent property tied to absolute size. It is a design property, not a scale property.

**Methodological caveat**: the bootstrap CI [2.9B, 13.4B] spans nearly an order of magnitude. N_c is a useful estimate but not a precise boundary. Models in the 3–13B range sit in a region of genuine uncertainty, and per-family diagnosis remains necessary rather than a universal size rule.