Publication Program

28+ papers mapped from one GL framework. Each opens a new domain while pointing back to the same equation. Contact adil@zehenlabs.com for preprints.

Submitted
Paper 3A NeurIPS 2026
“Lying Is Just a Phase”
The Hidden Alignment Transition in Language Model Scaling
We discover a phase transition at Nc = 3.5B parameters where the coupling between reasoning (HellaSwag) and truthfulness (TruthfulQA) flips sign. Below Nc: alignment taxes capabilities (r = -0.989). Above: they cooperate. The same coupled ODE cross-predicts Llama-2 at 5.6% MAE. An algebraic classifier, the isocline of the ODE, separates standard-trained from curated families. Engineering guidelines: data curation below Nc, free scaling above.
63base models 16families r = -0.989pre-Nc 5.6%ODE MAE
PDF Code + Data Dashboard
Paper 3B NeurIPS 2026
“The Growing Pains of Frontier Models”
Capability Coupling Analysis at Frontier Scale
At frontier scale (SWE-bench vs GPQA Diamond, 34+5 models, 10 labs), capabilities remain cooperative (r = +0.72, slope 0.513). The h-field diagnostic — deviation from the cooperation trend — reveals each lab’s training philosophy: Google is reasoning-specialist (h = +5.7), Anthropic is coding-rich (h = -9.1). Per-lab coupling slopes span 5x (Google 1.15 vs DeepSeek 0.23). Tax excursions (Sonnet 4.6, GPT-5.4) are temporary, recovering at the next release. Seven falsifiable predictions with timestamped deadlines.
39frontier models 10labs 0.513slope 7predictions
PDF Code + Data Dashboard
Writing Now
Paper 3I EMNLP ARR · May 25
Basin Memory
Physics-Informed Agent Memory with Dynamic Free Energy Learning
Energy-based retrieval with Boltzmann selection, temperature-controlled exploration, and offline sleep consolidation. 9 physics signals. 100% Judge accuracy on LoCoMo (1986/1986), 158ms latency. Beats Mem0 (49.7%), OpenClaw (72.5%), PropMem (82.3%).
100%LoCoMo Judge 158mslatency 9physics signals
Near-Ready
Paper 1 85% Written
“A Calorimeter Is All You Need”
Inverse Gor’kov Framework for Superconductor Classification · Nature / PRB
Classifies superconductor pairing symmetry from bulk thermodynamics alone. 33/33 known materials classified correctly. Leave-one-out error 3.9%. Det(a) sign-change discriminator separates s-wave from sign-changing. L1/L0 boosting ratio fingerprints: 0.2x conventional, 3-5x s±, 11x d-wave.
33/33classified 3.9%LOO error
Paper 2 85% Written
“Thirty-Three Times Too Heavy”
Mass Enhancement and Pairing in Heavy Fermions · Nature Physics / PRL
Three falsifiable predictions from one framework: FeSe 8-pressure trajectory (β ~ 0.12), La3Ni2O7 ΔC/γTc = 1.7±0.3, UTe2 Leggett mode at 209 GHz (unmeasured).
209 GHzUTe2 prediction 3falsifiable predictions
Paper 3E Data Complete
SFEE Universality
From CeRh2As2 to AI Scaling · PRL
R² = 0.855, Bayes factor 1049.6. The same free-energy structure governs superconductor phase boundaries and AI scaling transitions.
R² = 0.855 BF = 1049.6
Sleep 75% Written
GL Dynamics for EEG Sleep Stage Transitions
Nature (target)
Critical slowing down 2.60x, susceptibility 4.6x, Kramers escape rate within 1.4x of observation. Sleep consolidation connects to Basin Memory offline processing.
2.60xcritical slowing 4.6xsusceptibility
Planned & In Progress
3C
GL Phase Theory — RG Flow + Beyond-GL + Feynman Rules
From 125-page monograph · Nature Physics / ICLR
3F
Alignment Engineering — Self-Aligning via CAPE
EMNLP / ICML Safety
3G
Microscopic Thermodynamics — Per-Layer SAE Feature Coupling
NeurIPS Interpretability
3K
Architecture Dynamics — fk Library (Transformer, Mamba, CNN, MoE)
NeurIPS / ICML
3L
Crown Jewel Intervention — 5-Arm Coupling Modification Suite
Standalone