At the Desk
You’re off the clock.
Research lives on the work side.
Research
Continual learning.
Current work on how AI agents adapt over time, and a finished preprint in topology.
Current research
Continual learning
I’m studying how AI agents adapt as their environment changes, and how to measure what they retain, forget, and transfer over time. I’m developing a seed-replayable benchmark with evolving rules and executable ground truth, with a particular interest in agents that adapt through persistent memory and state while their model weights remain frozen.
That is a narrower question than retrieval. Looking something up isn’t continual learning; what matters is whether an agent that changes only its memory and state, not its parameters, keeps what still applies, drops what a rule change has overruled, carries what it learned to new tasks, and gives the answer that is correct at the current point in time. Each run can be replayed exactly from its seed, and answers are checked against ground truth by executing it.
This is work in progress. There is no publication yet, and I’m not reporting results.
Preprint
Coloring graphs of homologous spheres
On a surface, the curve graph joins two curves when they can be made disjoint. For the 3-manifold Mr, a connected sum of r copies of S1 × S2, the sphere graph plays the same role.
We look only at the spheres representing one primitive homology class, and show that the graph they form can be colored with r − 1 colors in exactly one way, up to relabeling the colors. For r ≥ 3, this gives a nontrivial homomorphism from the Torelli subgroup of Out(Fr) to the symmetric group on r − 1 symbols.
Earlier thread
Memory for speech agents
Separately, I’ve explored how voice agents should store and recall what they hear over long conversations. Results so far are exploratory, and I’m not reporting numbers until they hold up against matched baselines.