VideoAgent uses a large language model as an agent for multi-round retrieval, achieving 54.1% zero-shot accuracy with only 8.4 frames on average.

Xiaohan Wang, Yuhui Zhang, Orr Zohar et al.

SuperM2M integrates weakly-supervised mixture-to-mixture training with simulated data to improve real-world speech enhancement and noise-robust ASR.

Zhong-Qiu Wang

Introduces MultiGripperGrasp, a dataset with 11 grippers, 345 objects, 30.4M grasps, enabling cross-gripper transfer and ranking.

Luis Felipe Casas, Ninad Khargonkar, Balakrishnan Prabhakaran et al.

Proposes MM1, a multimodal pretraining framework combining contrastive and reconstructive losses, achieving SOTA with 30B parameters.

Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier et al.

OpenGraph introduces a hierarchical 3D graph framework using vision-language models for large-scale outdoor scene understanding, achieving 73.4% IoU without fine-tuning.

Yinan Deng, Jiahui Wang, Jingyu Zhao et al.

BEHAVIOR-1K: A benchmark with 1,000 everyday activities using OMNIGIBSON for realistic physics simulation.

Chengshu Li, Ruohan Zhang, Josiah Wong et al.

CodeUltraFeedback uses LLM-as-a-Judge to evaluate LLM alignment with coding preferences, enhancing model alignment.

Martin Weyssow, Aton Kamanda, Xin Zhou et al.

Proposes TRUMANS dataset and diffusion-based autoregressive model for arbitrary-length human-scene interaction motion synthesis.

Nan Jiang, Zhiyuan Zhang, Hongjie Li et al.

MoMa-LLM integrates dynamic scene graphs with open-vocabulary reasoning, boosting interactive object search efficiency by 30% in large indoor environments.

Daniel Honerkamp, Martin Büchner, Fabien Despinoy et al.

Proposes an incremental STL decomposition method that reduces complexity, improving trajectory planning efficiency by 40% and success rate to 95%.

Parv Kapoor, Eunsuk Kang, Romulo Meira-Goes

TaskCLIP leverages large vision-language models with a transformer aligner to improve task-oriented object detection by 3.5%, requiring only one RTX4090.

Hanning Chen, Wenjun Huang, Yang Ni et al.

Hierarchical 3D scene graph-based Task and Motion Planning (TAMP) with sparse problem instances and incremental object relevance detection enables scalable planning in large environments.

Aaron Ray, Christopher Bradley, Luca Carlone et al.

LiveCodeBench introduces a dynamic, multi-task, contamination-free benchmark with 400+ problems, evaluating 18 base and 34 tuned LLMs across time-segmented datasets.

Naman Jain, King Han, Alex Gu et al.

DexCap is a portable hand motion capture system combined with DexIL imitation learning, enabling scalable, high-precision dexterous robot skill acquisition.

Chen Wang, Haochen Shi, Weizhuo Wang et al.

SHROOM detects hallucinations in NLG outputs using a 4000-sample dataset, leveraging fine-tuning and zero-shot prompting strategies.

Timothee Mickus, Elaine Zosa, Raúl Vázquez et al.

DriveDreamer-2 uses LLMs and a Unified Multi-View Model to generate customized driving videos; FID 11.2, FVD 55.7.

Guosheng Zhao, Xiaofeng Wang, Zheng Zhu et al.

NeuPAN directly maps raw point clouds to latent distance features for real-time, map-free robot navigation using end-to-end model-based learning.

Ruihua Han, Shuai Wang, Shuaijun Wang et al.

SPLADE-v3 achieves over 40 MR R@10 on MS MARCO, improves out-of-domain results by 2%, surpassing BM25 and SPLADE++ via training innovations.

Carlos Lassance, Hervé Déjean, Thibault Formal et al.

FastV learns adaptive attention and prunes visual tokens, reducing FLOPs by up to 45% in LVLMs like LLaVA-1.5-13B, with minimal performance loss.

Liang Chen, Haozhe Zhao, Tianyu Liu et al.

ToolRerank combines adaptive truncation and hierarchy-aware reranking, significantly improving tool retrieval accuracy.

Yuanhang Zheng, Peng Li, Wei Liu et al.