Fast and Easy Infinite Neural Networks in Python

CVPR 2024-Improved Implicit Neural Representation with Fourier Reparameterized Training

ICML2025-Inductive Gradient Adjustment for Spectral Bias in Implicit Neural Representations

Existing literature about training-data analysis.

A unified framework for attributing model behavior to model components, training data, and training dynamics.

Official repository for "FOCUS: First Order Concentrated Updating Scheme"

Code for "What Happens During the Loss Plateau? Understanding Abrupt Learning in Transformers" (NeurIPS 2025)

Code for 'Towards a Theoretical Understanding of the 'Reversal Curse' via Training Dynamics'

Source code for <Probability Consistency in Large Language Models: Theoretical Foundations Meet Empirical Discrepancies>

Code for "Effect of equivariance on training dynamics"

Official repository for the EMNLP 2024 paper "How Hard is this Test Set? NLI Characterization by Exploiting Training Dynamics"

TMLR 2026 | Mechanistic interpretability: attention-head binding (EB*) as a marker of concept emergence. 7 models, 5 architectures (Pythia 160M–2.8B, OLMo-1B, CRFM GPT-2, SmolLM3-3B, Qwen2.5-1.5B), 41 terms.

[Ongoing] Post-training dynamics for SFT/RL checkpoint trajectories, concept directions, and structural analysis.

External LLM intelligence monitors & diagnoses MoE expert ecology during training — preventing routing collapse without auxiliary loss engineering. 16 Experts, 3 Tiers, Top-2 Gating, Claude-in-the-Loop.

Cross-Family Convergence of Neural Network Weight Skeletons. Companion to Zenodo paper (10.5281/zenodo.19652706).

A two-parameter Weibull lens on transformer weights — diagnose weight-magnitude distributions (shape k, scale λ) across 7 model families, and explain how λ evolves under AdamW training. Library (npm-weibull-py) + benchmark database + companion code for arXiv:2605.18898 and 2606.19367.

Code and data for: Three Phases of Expert Routing — How Load Balance Evolves During MoE Training

A plug-in debugger and visualizer for RL reward functions. Detects reward hacking, tracks training health, and renders a live terminal dashboard.

Code for "Abrupt Learning in Transformers: A Case Study on Matrix Completion" (NeurIPS 2024)

σFlow-PDE: A drop-in H-Bar training engine that escapes the σ-trap in neural PDE solvers via live σ/δ/α ODE integration, autonomous phase curriculum, and auto-falsification.