MSc, Natural Language Processing. Fully funded UAE scholarship, 3.84/4.0. Advisors: Timothy Baldwin, Artem Shelmanov.
2:44
Vladislav Smirnov
Reliable reasoning in large language models and agents
I study how to make large language models more reliable, particularly in reasoning and multi-step agent systems. My research combines uncertainty quantification, belief learning, and adaptive computation: how models assess evidence, recognize uncertainty, and decide when to reason further, verify an answer, or stop.
I am an MSc researcher at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI), advised by Timothy Baldwin and Artem Shelmanov. Previously, I worked on reasoning post-training in the Yandex Alice AI alignment team and visited the UKP Lab at TU Darmstadt with Iryna Gurevych. My background is in applied mathematics at Moscow Institute of Physics and Technology (MIPT).
Applying for PhD positions in LLM reliability and reasoning · Fall 2027 entry
News
- Oct 2026Awarded an OpenMined research grant (Ensembles & Routing program) to study routing and combining LLMs for the best accuracy at a given cost.
- Sep 2026Uncertainty Quantification for Diffusion LLMs accepted to NeurIPS 2026.
- Sep 2026Submitted three papers to ICLR 2027 (under review).
- Aug 2026Bayesian Control for Coding Agents accepted to EMNLP 2026 (main conference, oral).
- Apr 2026ThinkBooster accepted to ACL 2026 (System Demonstrations).
Research
I work on making LLM reasoning more reliable through both training and inference. My research spans how models learn from evidence, how they use additional computation, and how uncertainty can inform their decisions.
01
Training and belief learning
How can models learn to weigh evidence and revise their beliefs? I study posterior distillation for belief learning, alongside experience in reinforcement learning and reasoning post-training.
02
Reasoning and test-time computation
How can additional computation improve reasoning, and when is it worth the cost? I develop tools and methods for test-time scaling and adaptive stopping, studying the tradeoffs between answer quality and computational cost. With an OpenMined research grant starting in November 2026, I am extending this question to routing and combining multiple LLMs: what is the best accuracy reachable at a given cost?
03
Uncertainty and decision-making
How can estimates of correctness guide what a model does next? I study uncertainty estimation and Bayesian control for decisions such as verification, regeneration, and stopping, including in multi-step agent systems.
I also contribute to lm-polygraph, a library for uncertainty quantification in language models.
Looking ahead. I want to develop methods that make LLM reasoning more reliable through both training and inference. My interests span how models learn from evidence, how they use additional computation, and how they decide when to verify or stop. I am interested in the connections between these problems and in identifying where improvements to one can benefit the others.
Publications
* equal contribution · full list on Google Scholar →
2:44
8:32
Under review — ICLR 2027
- Training language models to hold calibrated beliefs over answers and source reliability · equal-contribution first co-author
- Quantifying an AI agent's uncertainty by tracking how its beliefs evolve across a task
Working descriptions — exact titles and preprint links will be posted once the papers are public.
Education
BSc, Applied Mathematics and Computer Science. Thesis: Federated RLHF (BRAIn Lab, adv. A. Beznosikov), 3.57/4.0.
Intensive two-year Master’s-level program in data science. Graduated.
Experience
Researcher, Reasoning Post-Training
Reasoning-focused post-training of large language models (RLHF, RLVR, and rubric-reward RL): reasoning-effort control for Mixture-of-Experts models, rubric-based rewards integrated into the RLVR pipeline, and SFT-data distillation from state-of-the-art models.
Visiting Researcher
Uncertainty quantification and calibration for LLM agents (with Iryna Gurevych): scaling an SFT + RL pipeline with reward and evaluation infrastructure toward calibrated, trustworthy behavior.
Research Engineer, LLM Agents
Built TargetOS, a B2B LLM-agent platform: a cooperating multi-agent system (Creator, Corrector, Analyst, Superagent), LLM-as-a-judge evaluation, RAG optimization, and agent memory.
Machine Learning Engineer, RecSys & Search
GPT-based shopping assistant, ranking models (CatBoost), and BERT search and product scoring distilled into production (NDCG 0.44 → 0.72).
C++ Developer
Native desktop-client features and network-performance work.
Get in touch
Based in Abu Dhabi, UAE. Applying for PhD positions in LLM reliability and reasoning (Fall 2027 entry) and open to research collaborations on uncertainty, reasoning, and reliable LLM agents.