Publications

Journal articles and preprints. Selected papers are highlighted

Illustration for TriRoute: Unified Learned Routing for Joint Adaptive Attention, Experts, and KV-Cache Allocation

TriRoute: Unified Learned Routing for Joint Adaptive Attention, Experts, and KV-Cache Allocation

Andrii Balashov, Olena Ponomarova

arXiv:2607.06601, 2026

  • Introduces TriRoute, a single lightweight controller that jointly routes attention mode, expert selection, and KV-cache bit-width per token per layer, instead of optimizing each axis in isolation.
  • Trains end-to-end under a Lagrangian budget constraint and identifies a cross-axis routing-collapse cascade, fixed via per-axis normalization and a coupling-aware balancing loss.
  • Pareto-dominates the best independent MoD+MoE+KV-quantization combination at matched FLOPs and memory across 160M–1.3B parameter models, while better preserving robustness on rare entities, code, and arithmetic.
arXivHTMLPDF
Illustration for A Spatio-Temporal Graph Attention Network for Real-Time Damage Localization in Prestressed Concrete Bridge Structures under Operational and Seismic Loads

A Spatio-Temporal Graph Attention Network for Real-Time Damage Localization in Prestressed Concrete Bridge Structures under Operational and Seismic Loads

Andrii Balashov, K. Medvediev, O. Usychenko, H. Ishutina, Yuliia Balashova

Journal of Physics: Conference Series 3251 (1), 012003, 2026

  • Proposes a spatio-temporal graph attention network for real-time damage localization in prestressed concrete bridges.
  • Models sensor networks as graphs to detect and localize damage under both operational and seismic loading.
IOP ScienceDOIScholar
Illustration for Adaptive Traffic Signal Optimization with Thermal Sensors and Reinforcement Learning

Adaptive Traffic Signal Optimization with Thermal Sensors and Reinforcement Learning

Andrii Balashov, Olena Ponomarova, Yuliia Balashova, Olexandr Tregub

Results in Engineering, 2025

  • Proposes an adaptive traffic light controller using FLIR TrafiOne thermal sensors and reinforcement learning.
  • Frames intersection control as a Markov Decision Process over real-time sensor data.
  • Reports up to 38% reduction in average delay compared to fixed-time signal plans, with headroom for multi-intersection coordination.
ScienceDirectSSRN PreprintPlumX MetricsElsevier
Illustration for Reinforcement Learning Fine-Tunes a Sparse Subnetwork in Large Language Models

Reinforcement Learning Fine-Tunes a Sparse Subnetwork in Large Language Models

Andrii Balashov

arXiv:2507.17107, 2025

  • Reinforcement learning modifies only a small subnetwork of an LLM, typically 5–30% of parameters, while more than 70% of weights remain effectively unchanged during fine-tuning.
  • Fine-tuning exclusively this RL-identified subnetwork fully reproduces the performance of full-model RLHF, yielding models whose parameters match the fully trained version in more than 99.9% of weights.
arXivPDF