Research
We study how recursive neural models use additional inference steps, following their hidden-state trajectories and testing their response to perturbations. The models can settle along the path of computation while remaining sensitive in other directions. This gives a more precise picture of stability than assuming that every direction must contract.
I derive a forecast of disagreement between count-based text predictors trained on independent corpus samples, using a separate reference corpus. Experiments across three text datasets test the forecast and a correction for small reference samples. Comparisons with small masked diffusion language models show partial agreement in input sensitivity, alongside different scaling with corpus size.
I am developing a framework in which language-model agents learn to allocate computation and coordinate with one another. The central questions are how to assign credit for a shared answer and how to learn useful collaboration when inference resources are limited.
Manuscript in preparation · Adaptive computation · CoordinationI use representation probes and activation interventions to test explanations of language-model computation, including J- and R-lenses. I am especially interested in the difference between when a concept becomes decodable and when it begins to influence the model’s answer.
Manuscript in preparation · Causal tests · Representation probingMy vision work includes land-cover segmentation from paired optical and radar imagery. In ongoing work on self-supervised radar models, I use controlled synthetic benchmarks, matched probes, and spatial-context tests to ask whether models learn scattering structure beyond amplitude and land-cover cues.
Radar manuscript in preparation · Multimodal fusion · Controlled benchmarksAcross these projects, I use tools from dynamical systems, high-dimensional statistics, and statistical physics. My broader interests include energy-based learning and topological methods, including persistent homology, for studying learned representations.
Energy-based learning · Topological data analysis · Dynamical systemsMy background is in theoretical physics, with a Ph.D. from Stony Brook University. This work explores holography in de Sitter, flat, and anti-de Sitter spacetimes, and the connections between gravity and quantum information.
I study how bulk information is encoded in cosmological horizons, including horizon response to matter, rotating and charged configurations, shock waves, entropy, and the holographic interpretation of static-patch physics.
de Sitter · Horizons · Shock waves · Quantum gravityMy work explores celestial operator products, loop-corrected soft theorems, asymptotic symmetries, holographic stress tensors in asymptotically flat spacetime, and Carrollian/BMS descriptions of the boundary theory.
Celestial CFT · BMS · Soft theorems · Carrollian QFTI use holographic and Hamilton–Jacobi methods to study thermal RG flows, finite-temperature field theory, holographic renormalization, and questions about black-hole interiors and thermal observables.
AdS/CFT · RG flows · Hamilton–Jacobi · Black holesRelated work concerns supersymmetric indices and microstate counting in AdS₃/CFT₂, holographic entanglement entropy, quantum chaos, and the role of information-theoretic observables in gravity.
AdS₃/CFT₂ · Entanglement · Microstates · Quantum information