The paper investigates how current image denoisers handle textures, finding that generic denoising models often over‑smooth fine patterns and that improving texture performance requires fine‑tuning on texture‑specific datasets. However, such fine‑tuning causes the models to lose their ability to denoise the original data, indicating a trade‑off. The authors argue that dedicated, texture‑aware denoisers are necessary for accurate texture reconstruction and for downstream tasks like generative modeling.
By Saptarshi Mandal, Oindrila Haldar, Chirayata Bhattacharya, Chandra Sekhar Seelamantula, Zhiwei Xu, Thrasyvoulos N. Pappas, Gaurav Sharma
HonestFace is a one‑step diffusion model designed for honest face restoration, preserving identity and realistic textures. It introduces an identity embedder that captures key identity features from low‑quality inputs and reference faces, and a masked face alignment method to enhance fine‑grained details. The authors also release a new real‑world dataset, MultiRefCeleb‑Test, and demonstrate that HonestFace outperforms state‑of‑the‑art methods in both visual quality and quantitative metrics.
By Jingkai Wang, Wu Miao, Jue Gong, Zheng Chen, Xing Liu, Hong Gu, Yutong Liu, Yulun Zhang
The paper introduces KoreaGEO, a fine‑grained, multimodal, privacy‑aware benchmark for image‑based geolocation built from Korean street views. It covers four socio‑spatial clusters, nine place types, and includes two captioning styles that mimic real‑world privacy exposure. Using a three‑path protocol (image‑only, functional‑caption, high‑risk‑caption), the study shows that input modality has a stronger effect on localization accuracy and privacy risk than model size or architecture, and reveals structural prediction biases toward core cities.
By Xiaonan Wang, Bo Shao, Hansaem Kim
The paper introduces Numerical Visual Chain-of-Thought (NV‑CoT), a framework that expands multimodal large language models (MLLMs) to reason over images using continuous numerical coordinates instead of discrete tokens or fixed patches. NV‑CoT allows MLLMs to generate bounding‑box coordinates directly as actions with minimal architectural changes, supporting both supervised fine‑tuning and reinforcement learning via a Gaussian or Laplace policy. Experiments on three benchmarks show that NV‑CoT improves localization precision, answer accuracy, and training convergence compared to 11 visual reasoning baselines.
By Kesen Zhao, Beier Zhu, Junbao Zhou, Xingyu Zhu, Zhongqi Yue, Hanwang Zhang
The paper introduces a benchmark for evaluating Structure-from-Motion (SfM) methods used in novel view synthesis (NVS) such as NeRFs and 3DGS. It compares classical SfM, which relies on local feature matches and bundle adjustment, with recent neural network approaches that regress camera parameters directly for faster but less accurate results. The study finds that reducing feature counts speeds up classical SfM without sacrificing pose accuracy, and that combining neural network initial estimates with classical refinement yields the best balance of efficiency and effectiveness.
By Jhacson Meza, Martin R. Oswald, Torsten Sattler
The paper introduces V‑GIFT, a lightweight method that enhances visual instruction tuning for multimodal large language models by incorporating a small set of visually grounded self‑supervised tasks expressed as natural language instructions. These tasks—such as rotation prediction, color matching, and cross‑view correspondence—are reformulated into image‑instruction‑response triplets, providing supervision that cannot be solved without visual evidence. The approach requires no human annotations, architectural changes, or extra training stages, and consistently improves vision‑centric performance across various models, regimes, and benchmarks when only 3‑10% of the training data is augmented with these instructions.
By Sophia Sirko-Galouchenko, Monika Wysoczanska, Andrei Bursuc, Nicolas Thome, Spyros Gidaris
EPIC-Bench is a perception‑centric benchmark for fine‑grained embodied visual grounding in vision‑language models. It contains 6.6k human‑annotated tuples of Image, Text, and Mask across the three core stages of embodied interaction—Target Localization, Navigation, and Manipulation. The benchmark offers 23 fine‑grained tasks, multi‑dimensional settings, and evaluates over 80 leading VLMs to provide diagnostic insights into their vision‑language grounding capabilities.
By Haozhe Shan, Xiancong Ren, Han Dong, Yingji Zhang, Haoyuan Shi, Jiayu Hu, Qinfan Zhang, Yi Zhang, Yong Dai, Bin Shen, Lizhen Qu, Zenglin Xu, Xiaozhu Ju
Fine-grained Human Motion Understanding with Language Models introduces FiGMo, an LLM-based model that represents motion as timestamp‑grounded skeletal poses, enabling reasoning over action order and duration. The model uses a unified pose encoder for both 2D and 3D skeletons, optionally incorporating RGB context, and relies on newly constructed pose‑to‑motion supervision to address limited pose‑level grounding. Experiments demonstrate that timestamps improve temporal localization and preserve physical timing under subsampling, leading FiGMo to achieve state‑of‑the‑art results on motion question answering, dense motion captioning, and action recognition with detected poses.
By Thomas Markhorst, Zhi-Yi Lin, Jouh Yeong Chew, Jan van Gemert, Xucong Zhang
WHTMix replaces the global self‑attention stage in stereo transformers with a data‑independent Walsh‑Hadamard token mixer, achieving the same end‑point error on synthetic driving data while cutting model compute by 2.46× and single‑image inference latency by 2.65×. The log‑linear cost of the mixer is especially advantageous for high‑resolution stereo matching, as shown by a complexity analysis that links benefits to the ratio of sequence length to channel width. Additionally, a hybrid log‑disparity loss is introduced to up‑weight small‑disparity pixels, improving accuracy on distant objects without extra computational cost.
By Prathyush Sajith, Emadeldeen Hamdan, Ahmet Enis Cetin
The paper introduces Locked Image Multiple Instance PreTraining (LIMIT), a method that aligns slide‑level pathology reports with patch‑level visual features by freezing a vision encoder and fine‑tuning a text encoder. LIMIT uses report embeddings as cross‑attention queries over bags of patch features, optimizing slide representations via InfoNCE. Applying LIMIT, the authors create CONCH‑Z, which achieves state‑of‑the‑art zero‑shot performance on 22 pathology tasks, outperforming vision‑language encoders trained on curated caption datasets and narrowing the gap with larger slide‑level foundation models.
By Daniel Shao, Dongmin Bang, Luca L. Weishaupt, Sophia J. Wagner, Nic G. Reitsam, Ming Y. Lu, Long P. Le, Richard J. Chen, Faisal Mahmood
TwinPrune introduces a unified spatial‑density signal to score token importance for both inference phases of vision‑language models: a bipartite merge during prefill and KV eviction during decode. By separating benchmarks into prefill‑only and full‑pipeline probes, the authors show that decode‑phase pruning must be evaluated on full‑pipeline tasks, as prefill probes can be misleading. Experiments on seven model variants demonstrate that density‑based prefill compression preserves accuracy while scaling linearly with visual token count, and that density‑top‑k KV eviction outperforms strong baselines on TextVQA and MM‑Vet across multiple retention ratios, achieving a Pareto frontier in compute, memory, and accuracy.
By Kai Liu, Anqi Li, Junxian Li, Zhixin Wang, Zhikai Chen, Renjing Pei, Yulun Zhang
DuplexBridge is a unified image‑editing model that coordinates its own understanding and generation capabilities without relying on an external module. It uses a source‑grounded bridge to specify editing targets, final‑state changes, and preservation constraints, and employs reciprocal trust checks to selectively share global task feedback while maintaining independent local supervision. The system, evaluated on the new CE‑Bench and other benchmarks, achieves 21.75–25.75 percentage points higher complete compliance than direct editors and demonstrates the effectiveness of distinct local responsibilities with selective shared feedback.
By Yidi Liu, Jie Xiao, Chaojie Mao, Jingfeng Zhang, Xueyang Fu
L2G-SR is a framework that enhances large language model (LLM)-based symbolic regression by incorporating a learned structural verifier. The verifier uses tree edit distance (TED) to assess structural similarity between candidate equations and the target, complementing numerical error during search. Experiments on LLM-SRBench show that L2G-SR achieves state‑of‑the‑art performance in numerical fit, equation‑level accuracy, and TED across various LLM backbones and domains.
By Letian Huang, Xiangyu Wang, Haowei Lin, Di He
GRADS is a graph-guided framework for generating synthetic data that explicitly controls difficulty by maintaining a graph of evidence and inference dependencies. It iteratively selects seeds, expands evidence, constructs question–answer pairs, and verifies them, ensuring that each generated task contains only indispensable inference steps. Experiments across security, software engineering, and scientific domains show that GRADS produces verified tasks with distinct difficulty levels, and training models on this data yields significant performance gains, including high solution-termination rates and reduced tool usage.
By Gaocheng Yang, Ruijie Tao, Kai Ye, Senhua Zhang, Zhengqin Zhu, Ziyang Li
The paper introduces RADAR, a framework for repairing and auditing confidence scores in large language models (LLMs). It separates confidence repair from deployment decisions and from error auditing, proposing Probe‑Anchored Isotonic Calibration (PAIC) for monotone mapping and a prompt‑clustered kernel test for detecting residual errors. Experiments across checkpoints and benchmarks show that richer predictors can improve confidence estimation, but projecting back to a scalar score can lose gains, and lower prediction risk does not always lead to better review decisions.
By Arijit Dey, Bhartendu Kumar, Ramya Hebbalaguppe, Khuushi Maheshwari
EmbodiedMidtrain bridges the gap between Vision‑Language Models (VLMs) and Vision‑Language‑Action Models (VLAs) by first analyzing the distribution mismatch between their data. It introduces a lightweight proximity estimator that selects VLA‑aligned samples from a large VLM pool, creating the EmbodiedMidtrain‑1M dataset used to mid‑train VLMs before VLA fine‑tuning. Experiments on simulated and real‑world robot tasks show consistent performance gains across different VLM backbones and VLA action heads, with early training steps already benefiting from the improved initialization.
By Yiyang Du, Zhanqiu Guo, Haochen Zhang, Annabella Macaluso, Xin Ye, Liu Ren, Yonatan Bisk, Chenyan Xiong
The paper introduces a domain‑specific debugging agent for repairing near‑miss accelerator kernels, arguing that debugging is more efficient than regenerating kernels from scratch. It tackles knowledge scarcity, integrity, and cost control through pattern retrieval, anti‑cheat detection, and convergence safeguards. Empirical results show the debug agent achieves a 66.7% Pass@1 rate and uses 92.8% fewer tokens per success compared to traditional regeneration methods.
By Yansong Sun, Shenxiu Wu, Siyuan Chen, Runlin Hou, Junhao Qiu, Junming Cao, Shudi Shao, Zhichao Lu, Qingfu Zhang
The paper investigates whether vision‑language‑action (VLA) models truly follow How‑level instructions that detail how to perform a task. The authors find that current VLAs do not benefit from such instructions and that task success actually decreases. They propose training with segment‑level enriched instructions and a new benchmark, SIMPLER‑IN, to better assess How‑level instruction following, but conclude that this remains an open challenge.
By Sangwu Park, Wonjoong Kim, Yeonjun In, Chanyoung Park
ARC‑SORT is a real‑time, training‑free multi‑object tracker designed for fast, direction‑changing athletes in sports broadcasts. It uses an Interacting Multiple Model Kalman filter to generate per‑track uncertainty‑adaptive search regions (EIoU) and a lightweight corrective identity swap mechanism for track crossings, achieving high HOTA scores on SportsMOT and DanceTrack. The tracker also drops the motion gate for long‑absent players, linking tracklets solely by appearance, and runs at ~50 fps on an RTX 3060 GPU.
By Girish Kumar Adari
KoViDoRe is a new benchmark for Korean visual document retrieval that addresses gaps in existing English‑centric datasets by including Korean documents with complex layouts such as tables, figures, and multi‑column structures. The dataset is built through a multi‑stage pipeline involving structured document parsing, synthetic query generation (summary‑based and context‑based), and human‑verified relevance mapping. Evaluation on KoViDoRe shows that current multimodal retrieval models struggle with Korean documents, especially when handling structured content and diverse query types, prompting the creation of a large‑scale training set, Ko‑VDR Train Public, to support model development.
By Yongbin Choi, Yongwoo Song, Mujeen Sung