Skip to content
Mansurbek Satarov

Where model behavior forms

I am a Ph.D. researcher at the University of Cincinnati, advised by Prof. Justin Zhan. My work is on how behavior emerges inside preference-based training, how to attribute it back to the updates that produced it, and how to keep tool-using agents inside their permissions once they are deployed.
Research areas
19 in 4 groups
Lifecycle stages
7 instrumented
Reading list
6 by others
Own papers
None yet

Research areas

19 areas across 4 groups — what forms behavior during training, what keeps it in bounds afterwards, what the model is made of, and what happens once it starts using tools.

Training & learning dynamics

How behavior forms while a model is being trained, and how to trace it back.

04 areas

  • 01

    RLHF & preference optimization

    How preference data and policy optimization shape model behavior.

  • 02

    Fine-tuning & adaptation

    Supervised fine-tuning across staged conditions, and what each stage does to behavior.

  • 03

    Learning dynamics & influence analysis

    Per-update gradient tracing to attribute behavior to training data.

  • 04

    Tokenization

    How text becomes tokens, and what that choice costs everything downstream of it.

Safety, security & governance

Keeping trained behavior predictable, permitted, and auditable once it is deployed.

06 areas

  • 05

    Trustworthy machine learning

    Making model behavior predictable, auditable, and safe to depend on.

  • 06

    AI alignment & model safety

    Training-time attacks and safety evaluation for aligned models.

  • 07

    LLM security

    Jailbreaks, data poisoning, and securing third-party agent extensions.

  • 08

    Guardrails

    Runtime constraints that keep a deployed model inside the behavior it is permitted.

  • 09

    Safety evaluation & benchmarking

    Building the evaluations that decide whether a model is safe enough to ship.

  • 10

    Privacy & data governance

    Treating data-governance, privacy, and confidentiality as requirements, not paperwork.

Models, modality & efficiency

What a model is made of, what it can perceive, and what it costs to run.

05 areas

  • 11

    Quantization & compression

    Making models smaller and cheaper to serve without losing the behavior that mattered.

  • 12

    Diffusion methods

    Diffusion models for generation, and the training dynamics underneath them.

  • 13

    Vision-language models

    Models that read an image and a prompt together and answer about both.

  • 14

    Multimodal learning

    Models and pipelines that cross text, image, and video.

  • 15

    Computer vision

    Detection and pose pipelines (YOLO, ViT-Pose) in research and production.

Agentic & retrieval systems

Models put to work: tools, retrieval, grounding, and serving.

04 areas

  • 16

    Agentic AI systems

    Tool-using agents, orchestration graphs, and safe third-party extensions.

  • 17

    Retrieval-augmented generation

    Grounded assistants with vector search and honest citations.

  • 18

    Question answering

    Answering with a language model: what makes an answer correct, complete, and checkable.

  • 19

    Scalable model deployment

    GPU serving, containerized inference, and reproducible training infra.

The lifecycle, instrumented

I study the whole lifecycle of an aligned model: how behavior forms during preference-based training, how training-time attacks can corrupt it, how influence analysis can audit it, and what changes once a model acts through tools and third-party extensions. These are open problems across the field — the published work below frames the questions.

Aligned-model lifecycle pipelineSeven connected stages: Data feeds supervised fine-tuning, then a reward model, then PPO reinforcement learning, then evaluation, safety analysis, and finally deployment as tool-using agents with third-party skills. The reward-model, PPO, and agent stages are highlighted as attack surfaces documented in the published literature; influence tracing feeds back from safety analysis to the training stages.Dataprompts + preferencesSFTsupervised fine-tuneReward Modelpreference learningPPO / RLHFpolicy optimizationEvaluationsafety benchmarksSafety Analysisinfluence tracingAgents & Toolsthird-party skillspoisoned data enters hereuntrusted extensionsinfluence tracing → attribute model behavior to training data
The lifecycle my research instruments — from training data to deployed, tool-using agents. The highlighted attack surfaces are documented in the published literature cited below.

Framed by the open literature

These papers are by other researchers, listed as the context this work builds on — never as his own. Every arXiv identifier was checked against the arXiv API.

Publications

Papers will be listed here as they become citable. Until then, the research areas above and the systems in the work index are the accurate picture of the work.

A Google Scholar profile will be linked alongside the first listed paper.

Looking for the full record? See the experience record →