We show that vision-language models are selectively blind: linguistic framing alters visual attention, degrading grounding and accuracy. We introduce a prompt-tuning method that realigns attention and restores performance.
Wan-Cyuan Fan, Jiayun Luo, Declan Kutscher, Leonid Sigal, Ritwik Gupta. (2026). "Tinted Frames: Question Framing Blinds Vision-Language Models." EMNLP 2026.
We proposed a reinforcement learning technique using a Plackett-Luce policy to find the optimal ordering of patches for long-sequence vision transformers.
Declan Kutscher, David M. Chan, Yutong Bai, Trevor Darrell, Ritwik Gupta. (2025). "REOrdering Patches Improves Vision Models." Neural Information Processing Systems (NeurIPS) 2025.
This work introduces a graph neural network method to predict stream water temperature and reduce model bias across locations with different socioeconomic attributes.
Erhu He, Declan Kutscher, Yiqun Xie, Jacob Zwart, Zhe Jiang, Huaxiu Yao, Xiaowei Jia. (2025). "Physics-Guided Fair Graph Sampling for Water Temperature Prediction in River Networks." AAAI Conference on Artificial Intelligence.
M.S. thesis analyzing when ImageNet or satellite pretraining helps remote sensing, and how representation quality drives transfer.
Declan Kutscher (2025). On the Effectiveness of Pretrained Models for Remote Sensing. M.S. thesis, Department of Computer Science, University of Pittsburgh.
In this paper, we investigate the usefulness and design methodology of applying large language models (LLMs) to improve and automate the process of coding case data. We introduce tools to guide LLMs to assist in this coding process.
D. Kutscher, A. Whisnant, "Leveraging LLMs for Data Coding," Carnegie Mellon University, Software Engineering Institute Digital Library. Software Engineering Institute, White Paper, 04-Nov-2024