Visual prompting aims to enhance the performance of vision-language models (VLMs), which, despite their remarkable capabilities, often struggle with dense, detailed images, leading to incorrect answers or hallucinations. We propose Visual Prompting Through Image Mines , a novel algorithm that leverages attention patterns from a base VLM to generate image crops for improved visual grounding. Specifically, we extract attention values from output text tokens in the LLaVA-8B model, overlay them onto image patches to create an attention graph, and apply a modified breadth-first search (BFS) to identify key regions as image crops. Using the SigLIP model, we refine these regions into Image Mines, retaining only the most relevant crops. Our approach supports both single-image and multi-image inference setups, delivering superior performance. It consistently outperforms existing visual prompting methods in nearly all scenarios.

We propose Step-by-Step Coding (SBSC): a multi-turn math reasoning framework that enables Large Language Models (LLMs) to generate sequence of programs for solving Olympiad level math problems. After each turn/step, by leveraging the code execution outputs and programs of previous steps, the model generates the next sub-task and the corresponding program to complete it. This way, SBSC, sequentially navigates to reach the final answer. SBSC allows more granular, flexible and precise approach to problem-solving compared to existing methods. Extensive experiments highlight the effectiveness of SBSC in tackling competition and Olympiad-level math problems. For Claude-3.5-Sonnet, we observe SBSC (greedy decoding) surpasses existing state-of-the-art (SOTA) program generation based reasoning strategies by absolute 10.7% on AMC12, 8% on AIME and 12.6% on MathOdyssey. Given SBSC is multi-turn in nature, we also benchmark SBSC’s greedy decoding against self- consistency decoding results of existing SOTA math reasoning strategies and observe performance gain by absolute 6.2% on AMC, 6.7% on AIME and 7.4% on MathOdyssey.

Video Object Detection (VOD) is one of the fundamental problems in video understanding with applications ranging from surveillance to autonomous driving. But many such real-world applications are unable to leverage the existing VOD models owing to their higher computational complexity which reduces inference speed. Single-stage still-image object detection models are naively used without any use of video information. In this paper, we present YOLOX based VOD model, YOLO-MaxVOD, which provides a better trade-off between accuracy and inference time than the current real-time VOD solutions. Specifically, we propose a temporal fusion module that integrates within the YOLOX architecture to take advantage of the high speed that the YOLOX model offers. In our experimentation on the Imagenet-VID dataset, we show that YOLO-MaxVOD shows 4.4-5.6% AP50 improvement over the baseline YOLOX, across different versions, with just a 1-2 ms increase in latency on NVIDIA 1080Ti GPU.

Tiling is a key technique for data locality optimization and is widely used in high-performance implementations of dense matrix-matrix multiplication for multicore/manycore CPUs and GPUs. However, the irregular and matrix-dependent data access pattern of sparse matrix multiplication makes it challenging to use tiling to enhance data reuse. In this paper, we devise an adaptive tiling strategy and apply it to enhance the performance of two primitives: SpMM (product of sparse matrix and dense matrix) and SDDMM (sampled dense-dense matrix multiplication). In contrast to studies that have resorted to non-standard sparse-matrix representations to enhance performance, we use the standard Compressed Sparse Row (CSR) representation, within which intra-row reordering is performed to enable adaptive tiling. Experimental evaluation using an extensive set of matrices from the Sparse Suite collection demonstrates significant performance improvement over currently available state-of-the-art alternatives.