Hello, I'mRobin Labryga

Transformer models systematically favor certain token positions, yet the architectural origins of this position bias remain poorly understood. This bias is closely connected to the Lost-in-the-Middle phenomenon, where models underutilize information placed in the middle of the context. We show that Lost-in-the-Middle-type behavior can arise from the architecture of causal Transformers itself. To do so, we develop a structural theory of position bias based on residual-aware cumulative attention rollout. At finite depth, causal masking and residual connections induce broad, often U-shaped, influence profiles. At infinite depth, our framework resolves a discrepancy between prior attention-only collapse theory and practical Transformer behavior: residual connections fundamentally change cumulative attention dynamics. Empirically, the predicted profiles closely match measured input-token influence in pretrained language models.

Line search is a fundamental part of iterative optimization methods for unconstrained and bound-constrained optimization problems to determine suitable step lengths that provide sufficient improvement in each iteration. Traditional line search methods are based on iterative interval refinement, where valuable information about function value and gradient is discarded in each iteration. We propose a line search method via Bayesian optimization, preserving and utilizing otherwise discarded information to improve step-length choices. Our approach is guaranteed to converge and shows superior performance compared to state-of-the-art methods based on empirical tests on the challenging unconstrained and bound-constrained optimization problems from the CUTEst test set.

We present a multi-head vision transformer approach for multi-label plant species prediction in vegetation plot images, addressing the PlantCLEF 2025 challenge. The task involves training models on single-species plant images while testing on multi-species quadrat images, creating a drastic domain shift. Our methodology leverages a pre-trained DINOv2 Vision Transformer Base (ViT-B/14) backbone with multiple classification heads for species, genus, and family prediction, utilizing taxonomic hierarchies. Key contributions include multi-scale tiling to capture plants at different scales, dynamic threshold optimization based on mean prediction length, and ensemble strategies through bagging and Hydra model architectures. The approach incorporates various inference techniques including image cropping to remove non-plant artifacts, top-n filtering for prediction constraints, and logit thresholding strategies. Experiments were conducted on approximately 1.4 million training images covering 7,806 plant species. Results demonstrate strong performance, making our submission 3rd best on the private leaderboard.