Amirhossein Samandar
Ph.D. Candidate in Physics · Case Western Reserve University
Cosmologist working at the intersection of cosmological data analysis, statistical inference, probabilistic machine learning, and computational methods.
My research covers Bayesian inference, information-theoretic model comparison, simulation-based (likelihood-free) inference with deep generative models, CMB statistics, and calibrated Bayesian evaluation of large language models. I am comfortable moving between analytic derivation, GPU/HPC implementation, and validation of learned inference.
Graduating Fall 2026 · applying for postdoctoral positions.
The question behind the work
Cosmology gives us one universe and one sky. Nearly every statistic we use to decide whether a model is detectable is an average over an ensemble of universes we will never observe. My research asks what those ensemble quantities actually tell us about the single measurement in hand, and how to answer that honestly — analytically where the distributions are tractable, and with learned inference where they are not.
The same question turns out to have teeth outside cosmology. A benchmark score for a language model is also one noisy draw from a distribution, reported as if it were a number. The methods transfer.
Current work
2025 – present
Detection Probability for Gaussian Model Comparison
Turning an expected information gain into a probability of detection on the one sky we actually observe.
2025 – present
Neural Likelihood Emulation: Simulation-Based Inference vs. MCMC
A direction I am moving into: replacing an intractable MCMC scan with amortized neural inference, and proving the replacement is honest.
2023 – present
CMB Statistics and Covariance Modelling for Cosmic Topology
If space is compact, the CMB is no longer described by a power spectrum. Someone has to compute what replaces it.
Selected publications
The Topology of the Universe
Nature Astronomy · Invited review, in press
Invited review of the observational status of cosmic topology and the prospects for CMB (LiteBIRD, Taurus) and large-scale-structure searches.
Cosmic topology. Part IIb. Eigenmodes, correlation matrices, and detectability of non-orientable Euclidean manifolds
Journal of Cosmology and Astroparticle Physics 07 (2026) 004
Eigenmodes, non-diagonal aℓm covariance matrices, and cosmic-variance-limited detectability for the non-orientable Euclidean topologies E7–E10, E13–E15 and E17.
Don't Pass@k: A Bayesian Framework for Large Language Model Evaluation
International Conference on Learning Representations (ICLR) 2026 · Main conference
Replaces Pass@k and avg@N with posterior estimates and credible intervals on a model's underlying success probability. Outcomes are modelled as categorical with a Dirichlet prior, giving closed-form posterior mean and uncertainty for any weighted rubric. Achieves faster convergence and greater rank stability than Pass@k at far smaller sample counts.
Cosmic topology. Part IIIb. Eigenmodes and correlation matrices of spin-2 perturbations in orientable Euclidean manifolds
Journal of Cosmology and Astroparticle Physics 08 (2025) 015
First computation of spin-2 (tensor) Laplacian eigenmodes and the corresponding CMB correlation matrices for orientable compact Euclidean manifolds.
Software
CMBtopology
GPU-parallelized Python/HPC package computing CMB covariance matrices for compact Euclidean topologies. Adopted collaboration-wide for tensor-mode likelihood analysis.
Scorio
Open-source Bayesian evaluation toolkit for large language models, implementing Dirichlet-posterior estimates and credible intervals in place of Pass@k.