Soheil Sayah Varg
B.Sc. Computer Engineering, Sharif University of Technology
Tehran, Iran
I work on inference and approximation under uncertainty: when the thing you care about cannot be computed exactly — a posterior, a value function, an expectation over an intractable distribution — you compute an approximation instead, and the question that interests me is what licenses trusting it.
01Research
Exact controlled generation from tilted autoregressive targets
Steering a language model toward a reward implicitly defines a tilted target distribution, and the samplers used in practice do not draw from it — they draw from something nearby, and the gap is rarely measured. This work samples from the tilted target exactly in distribution, so the remaining cost is compute rather than a silent bias. Manuscript in preparation.
Model-based reinforcement learning in confounded POMDPs
Completed and presented an empirical implementation of a theory-only ICML 2024 result on offline RL under unobserved confounding. At full benchmark scale the confounding-blind baseline beat the corrected estimator in every cell of the sweep — identification is not estimation. I drove the implementation.
Bandit literature review
Reading in batched bandits, privacy-constrained linear bandits, and misspecified kernelized bandit optimization. No project or manuscript yet.
02Selected projects
model-based-rl-confounded-pomdps
End-to-end empirical implementation of model-based RL for confounded POMDPs, with exact-oracle checks and a documented pessimism failure mode.
war-of-attrition-rl
Bayesian type filtering and potential-based reward shaping for a multi-stage war of attrition with incomplete information.
modern-information-retrieval
Dense + sparse retrieval fused by RRF, with cross-encoder reranking.
convex-optimization-algorithms
Chambolle-Pock and ADMM written from their update rules, checked against a solver rather than delegated to one.
market-predictability-study
A backtest of mine returned several hundred percent. This is the study of whether any of the reasons I gave for it hold up. They do not.
segmentation-and-deep-rl
Attention U-Net variants for segmentation, and Soft Actor-Critic built from scratch in PyTorch.
neogit
A version control system written from scratch in C — staging, commits, branches, checkout, revert, tags, and a pre-commit hook framework. No libraries.
imdb-postgresql-optimization
~190M IMDb rows into PostgreSQL via a SKIP LOCKED queue, then eight queries indexed with real before/after measurements.
03Teaching
Teaching Assistant, Sharif University of Technology
Head TA for Signals & Systems (Dr. Manzuri). Also Stochastic Processes and Probability & Statistics (Dr. Najafi), Machine Learning and Probability & Statistics (Dr. Sharifi-Zarchi), Linear Algebra twice (Dr. Rabiee & Dr. Ramezani), Logic Design (Dr. Hessabi & Dr. Arshadi), and Data Structures & Algorithms (Dr. Abam).
Co-founder, Resonance
An online Physics Olympiad school, started as a small study group, for students who have no selective high school near them. I teach the mechanics, electromagnetics and laboratory tracks and write problem sets for the national selection rounds.
04Education and honors
B.Sc. Computer Engineering, Sharif University of Technology
GPA 19.38 / 20.0, ranked 14th of 186 — the departmental average for this cohort is 17.24. Stochastic Processes and Reinforcement Learning taken as graduate courses.
Gold Medal, Iranian National Physics Olympiad — 2023
7th nationwide among more than 10,000 participants, and 2nd nationwide in the experimental examination. Admitted to Sharif through the olympiad medalist track.
Juror, 13th Persian Young Naturalists’ Tournament — 2025
An IYPT-format research tournament under the World Federation of Physics Competitions. Judged team defenses of open-ended physics research problems.