We introduce KV Cache Steering, a novel approach to enhance reasoning capabilities in small language models by manipulating their key-value cache representations. Our method demonstrates significant improvements in logical reasoning tasks while maintaining computational efficiency. Through extensive experiments on various benchmarks, we show that our approach can induce reasoning behaviors typically observed only in much larger models. The technique opens new avenues for deploying capable language models in resource-constrained environments.
@article{belitsky2025kv,title={KV Cache Steering for Controlling Frozen LLMs},author={Belitsky, Max and Kopiczko, Dawid J and Dorkenwald, Michael and Mirza, M Jehanzeb and Glass, James R and Snoek, Cees GM and Asano, Yuki M},journal={arXiv preprint arXiv:2507.08799},year={2025},github={https://github.com/MaxBelitsky/cache-steering},}
[Re] Learning Fair Graph Representations via Automated Data Augmentations
Max Belitsky, Filipe Laitenberger, Denys Sheremet, and 1 more author
@article{Belitsky:2025,author={Belitsky, Max and Laitenberger, Filipe and Sheremet, Denys and Belkacemi, Nordin},title={{[Re] Learning Fair Graph Representations via Automated Data Augmentations}},journal={ReScience C},year={2025},volume={10},number={1},pages={{#6}},doi={10.5281/zenodo.16374814},url={https://zenodo.org/record/16374814/files/article.pdf},github={https://github.com/Thiggel/FACT},}
2024
Exploring Monotonicity in Early-Exiting Language Models
Filipe Laitenberger, Max Belitsky, and Denys Sheremet
In Workshop on Efficient Systems for Foundation Models II @ ICML, 2024
@inproceedings{laitenberger2024exploring,title={Exploring Monotonicity in Early-Exiting Language Models},author={Laitenberger, Filipe and Belitsky, Max and Sheremet, Denys},booktitle={Workshop on Efficient Systems for Foundation Models II @ ICML},year={2024},}
2021
Examining team interaction using dynamic complexity and network visualizations
Travis J Wiltshire, Dan Hudson, Max Belitsky, and 3 more authors
In 2021 IEEE 2nd International Conference on Human-Machine Systems (ICHMS), 2021
This paper presents a novel approach to examining team interaction patterns through dynamic complexity measures and network visualizations. We analyze real-time team communication data to identify key interaction patterns that correlate with team performance. Our methodology combines network analysis techniques with complexity measures to provide insights into effective team dynamics. The results demonstrate significant correlations between specific interaction patterns and successful team outcomes.
@inproceedings{wiltshire2021examining,title={Examining team interaction using dynamic complexity and network visualizations},author={Wiltshire, Travis J and Hudson, Dan and Belitsky, Max and Lijdsman, Philia and Wever, Stijn and Atzmueller, Martin},booktitle={2021 IEEE 2nd International Conference on Human-Machine Systems (ICHMS)},pages={1--6},year={2021},organization={IEEE},doi={10.1109/ICHMS53169.2021.9582670},github={https://github.com/example/network-analysis},}
