HedgesFamily.com

Source-Optimal Training is Transfer-Suboptimal
C. Evans Hedges
TMLR, May 2026

This paper comes out of a desire to understand transfer learning: training a model on some task you have a ton of data on but don't particularly care about so that it might learn useful things to use when you fine tune it for a task you do care about. I explored a very simple version of transfer learning (L2-SP Ridge regression) to be able to get a handle on when transfer is beneficial, and how beneficial it can be. Typically when people train their initial model they do the simple thing of optimizing for source performance, however this paper shows that doing the optimal thing for source performance is almost never the optimal training for transfer learning.

OrthoGrad Improves Neural Calibration
C. Evans Hedges
NeurIPS Workshop on Optimization, 2025

A common problem with training neural network classifiers is that eventually the best way to decrease loss is simply to scale up confidence and send your logits to infinity. This doesn't help us learn new useful features, but instead leads to models that are highly overconfident. By orthogonalizing the gradients with respect to the model weights, we force the model to not simply scale these logits and instead further explore the loss landscape. We show this technique improves calibration without sacrificing accuracy.

A Theoretical Framework Bridging Model Validation and Loss Ratio in Insurance
C. Evans Hedges
Preprint, 2025

During my first few months working as a Data Scientist at Lemonade, I was tasked with building a framework that would allow us to connect model performance metrics directly with business value. From this exploration came the first closed form formula linking loss model correlation with loss ratio. It's a simple result that backs up practitioner intuition regarding diminishing returns of performance improvement.

Which Phases Are Thermodynamically Realizable? A Local Entropy Criterion
C. Evans Hedges
Preprint, 2026

A natural question in thermodynamic formalism is what states/phases are realizable physically. This paper fully answers this question in a large class of dynamical systems (including non-compact settings) and answers it with a tight if-and-only-if criterion involving the local entropy structure of that state. This was the main question I was interested in during grad school and I'm extremely happy to have been able to answer it.

x Plays Pokemon, for Almost-Every x
C. Evans Hedges
arXiv, 2025

This is a short write up that attempts to settle the age old question of "will pi plays pokémon ever win the game?" There is no new math here, but is supposed to be a fun framing that motivates some graph theory facts and learning what a disjunctive number is.

Some Results in Thermodynamic Formalism
C. Evans Hedges
PhD Dissertation, 2025

This is my PhD dissertation. It mostly just collects the results of the three dynamical system papers below, but the introduction provides a nice background on motivation and a quick overview of the relevant definitions/objects and main results.

On the Equivalence of Equilibrium and Freezing States in Dynamical Systems
C. Evans Hedges
Under peer review, 2025

A dynamical system freezes if you hit some equilibrium state and then it doesn't change as things get colder. It's easy to see that if you are a frozen state, you must be an equilibrium state, however the opposite direction is extremely unclear. The primary result of this paper is showing that the attainability of these two things is equivalent.

Computability of pressure for subshifts on countable amenable groups
C. Evans Hedges, Ronnie Pavlov
Nonlinearity, 2025

Here, my advisor and I explore computability theory in the highly abstract setting of thermodynamic formalism on subshifts under countable amenable groups. We show computability (and non-computability) of many functions of interest, as well as explore computability properties of ground states for a given potential.

Bounds for equilibrium states on amenable group subshifts
C. Evans Hedges
Ergodic Theory and Dynamical Systems, 2025

This paper shows that for general subshifts, equilibrium states remain somewhat Gibbs even if they are not always Gibbs. The main idea is that as your subshift gets worse, your equilibrium states might also get worse, but they still maintain some Gibbs-ness.

A New Graph Recoloring Game
Mark Anderson, Shiying Gu, Charles E. Hedges, Felipe Quiroga
Congressus Numerantium, 2017

In the summer after my freshman year of college I joined a small group of researchers to play with graph theory. We built a game and proved some facts about it. This was my introduction to mathematics research and I had a blast.

Transplants for non-lethal conditions: a case against hand transplantation in minors
C. Evans Hedges, Phillip M. Rosoff
Journal of Medical Ethics, 2018

As part of my M.A. in Bioethics and Science Policy I wrote this paper that argues, except for very specific extreme circumstances, it is not worth it to perform hand transplants on minors. I believed this was true in 2016 when I wrote the paper, but I have not followed the field closely enough to see if things have changed.

Source Credibility Signals in Fine-Tuning
Toy experiment (GPT-2 fine-tuning)

This repo contains a toy experiment evaluating how source credibility cues in training text affect which facts a language model learns when presented with conflicting information. In the sweep, low-credibility examples needed to be ~93-95% of the training data to flip the model's preference.