Researcher, Reasoning Post-Training Current
Reasoning enhancement and post-training (SFT, RLHF, RLVR) for large language models: reward and evaluation design, reasoning-failure analysis, and large-scale experiments.
Vladislav Smirnov
Reliable and controllable reasoning in language models and their agents
I work on the safety and reliability of large language models and the agents built on them — using uncertainty quantification to ask whether the training that makes them reason better also keeps them honest and calibrated, and to control them at inference time. MSc researcher at MBZUAI (advised by Timothy Baldwin and Artem Shelmanov); reasoning post-training in the Yandex Alice AI team; recently a research visit at the UKP Lab, TU Darmstadt with Iryna Gurevych. Background in applied mathematics at MIPT.
Open to PhD positions in technical AI safety · 2027/28 entry
News
- 2026Bayesian Control for Coding Agents accepted to EMNLP 2026 (main conference).
- 2026ThinkBooster accepted to ACL 2026 (System Demonstrations).
- 2026Joined the Yandex Alice AI team — reasoning post-training (SFT / RLHF / RLVR).
- 2026Completed a research visit at the UKP Lab, TU Darmstadt (with Iryna Gurevych).
- 2026Papers under review at EMNLP / NeurIPS — UQ for Diffusion LLMs, PRPO (personalized RLHF).
- 2026New work in preparation for ICLR 2027.
Research
My aim is large language models — and the agents built on them — that are safe and reliable. I approach this through the two levers I've worked on most: strengthening how LLMs reason, and applying uncertainty quantification to make their behavior more trustworthy, at both training time and inference time.
01
Uncertainty-aware test-time control
When should a model think longer, stop, verify, abstain, or escalate? I allocate inference-time compute by uncertainty rather than fixed heuristics, and study controllers that trade off accuracy, calibration, cost, and risk.
02
Safe post-training
Does reasoning-oriented RL (RLHF, RLVR) buy real reliability, or reward-hack the verifier? I study how post-training reshapes calibration, and how to preserve honest uncertainty through aggressive fine-tuning.
03
Agentic safety & monitoring
Uncertainty compounds across long-horizon tool use. I work on propagating it across steps and tools, catching cascading errors before irreversible actions, and monitoring reasoning and coding agents.
Underpinned by uncertainty-quantification methods for LLMs (contributor to lm-polygraph) and UQ for diffusion language models.
Looking ahead. For a PhD I want to unify these threads into one question: how do we keep language models honest about their uncertainty as they are trained to reason and act — and use that uncertainty to control them at inference time? A concrete first step: study whether reasoning-oriented RL (RLHF / RLVR) preserves or erodes calibration, then build uncertainty-aware controllers that decide when a model should think longer, verify, abstain, or hand off — so that more capable reasoning also becomes more trustworthy and monitorable.
Publications
* equal contribution · full list on Google Scholar →
Bayesian Control for Coding Agents
Uncertainty Quantification for Large Language Diffusion Models
PRPO: Collaborative Online Policy Learning in Personalized RLHF
In preparation — ICLR 2027
- Uncertainty quantification for LLM agents
- Preserving calibrated uncertainty through reasoning-oriented RL
- Uncertainty-aware control for safer, more reliable reasoning
Experience
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.
Education
MBZUAI
MSc, Natural Language Processing. Fully funded UAE scholarship, 4.0/4.0. Advisors: Timothy Baldwin, Artem Shelmanov.
HSE University
MSc, Finance (concurrent). Quantitative finance and statistical / mathematical methods for investment.
MIPT
BSc, Applied Mathematics & Informatics. Thesis: Federated RLHF (BRAIn Lab, adv. A. Beznosikov), 4.6/5.0.
Yandex School of Data Analysis
Graduate program in ML, NLP, and computer science.
Service
Reviewer
Workshop on Uncertainty-Aware NLP (co-located with EMNLP 2026) — peer review of submissions on uncertainty estimation and calibration for NLP.
Get in touch
Based in Abu Dhabi, UAE. Open to PhD positions in technical AI safety (2027/28 entry) and to research collaborations on uncertainty, reasoning, and reliable LLM agents.