Publications
Newest first. Workshop papers are marked as workshop papers. Each entry links to the paper and, where I have them, the code, artifacts, and data.
7 of 7 publications
Igor Bogdanov, C. Huang
ICML 2026 Workshop on Mechanistic Interpretability
Geometry-invariant sparse autoencoders and causal intervention probe whether reasoning representations are shared, and functionally interchangeable, across languages.
Igor Bogdanov, C.-H. Lung, T. Kunz, J. Gao, A. Taylor, M. Zaman
ACM Conference on AI and Agentic Systems (CAIS), 2026
3,475 episodes across five model families: structured state abstraction improves return by up to 76%, while distributed deliberation across an agent hierarchy can be up to 3.4× worse.
Igor Bogdanov, C.-H. Lung, T. Kunz, J. Gao, A. Taylor, M. Zaman
ACM Conference on AI and Agentic Systems (CAIS), 2026
A population-based protocol that turns failed trajectories into reusable prompt memory and broadcasts champion memories, improving return 1.7–7.7× over zero-shot without weight updates.
Igor Bogdanov, O. Manakina, C.-H. Lung
ICLR 2026 Workshop on LLM Logical Reasoning
84,540 trajectories across eight model families: time-to-event analysis and a seven-category rationale taxonomy expose when agents abandon goals and how they justify it.
O. Manakina, Igor Bogdanov, C.-H. Lung
NeurIPS 2025 Workshop on Multi-Turn Interactions in Large Language Models
A text-based multi-agent gym modelled as an MDP/POMDP for studying long-horizon LLM behaviour under social exposure, personas, and tool-use budgets.
O. Manakina, Igor Bogdanov
EDM 2025 Workshop on Leveraging LLMs for Innovative Educational Data Mining (WLIEDM)
A multi-armed bandit selects among four grading recipes per essay, cutting LLM calls by 78.4% while tracking cost and reliability.
Igor Bogdanov, J. Green
IEEE COMPSAC 2025, MediComp Symposium
A 4,144-video, 12-class, privacy-compliant NICU intervention dataset with TimeSformer and MotionFormer baselines reaching up to 94% top-1 accuracy.
Contact
I’m interested in research engineering, applied research, agent infrastructure, evaluation, reliability, interpretability, and research-to-production work.