DynamicRAG uses reinforcement learning to adaptively rerank documents, boosting QA performance with state-of-the-art results.

Jiashuo Sun, Xianrui Zhong, Sizhe Zhou et al.

SparseMeXT optimizes sparse representations for efficient HD map construction, achieving 68.9% mAP.

Anqing Jiang, Jinhao Chai, Yu Gao et al.

This study systematically analyzes how pre-training and fine-tuning influence knowledge acquisition in dense retrieval models, confirming that fine-tuning mainly adjusts neuron activation rather than reorganizing stored knowledge.

Zheng Yao, Shuai Wang, Guido Zuccon

Seed1.5-VL combines a 532M vision encoder with a 20B MoE LLM, achieving SOTA on 38 of 60 benchmarks, excelling in multimodal reasoning.

Dong Guo, Faming Wu, Feida Zhu et al.

Proposes a quantitative measure (DEq(p)) for distracting effect, using synthetic and retrieval-based hard distractors to improve RAG accuracy by 7.5%.

Chen Amiraz, Florin Cuconasu, Simone Filice et al.

Proposes Sparse Concept Layer in ViT for rule extraction, improving accuracy by 5.14%.

Parth Padalkar, Gopal Gupta

Proposes paired autoencoders for inverse problems, integrating data-driven and model-based approaches to enhance reconstruction robustness and uncertainty quantification.

Matthias Chung, Bas Peters, Michael Solomon

Filtered LLaVA dataset using Toxic-BERT and LlavaGuard, removing 7,531 toxic image-text pairs.

Karthik Reddy Kanjula, Surya Guthikonda, Nahid Alam et al.

Large-scale Anymate dataset enables Transformer-based models to automatically predict 3D skeletons and skinning weights, outperforming prior methods.

Yufan Deng, Yuhao Zhang, Chen Geng et al.

Moosbauer-Poole algorithm optimizes matrix multiplication, reducing 5x5 matrix multiplications to 93.

Manuel Kauers, Isaac Wood

GPT-Image model for image restoration enhances visual quality but lacks pixel-level structural fidelity.

Hao Yang, Yan Yang, Ruikun Zhang et al.

BrickGPT uses large-scale stable datasets and physics-aware inference to generate structurally stable brick models from text prompts.

Ava Pun, Kangle Deng, Ruixuan Liu et al.

The clem:todd framework employs self-play with large language models (LLMs) to systematically benchmark task-oriented dialogue systems across architectures, datasets, and metrics, revealing performance trade-offs.

Chalamalasetti Kranti, Sherzod Hakimov, David Schlangen

FG-CLIP leverages 1.6 billion long caption-image pairs, 12M region annotations, and 10M hard negatives to enhance fine-grained alignment.

Chunyu Xie, Bin Wang, Fanjing Kong et al.

Evaluated nine large language models' zero-shot performance on Bengali health query summarization; Mixtral outperformed others in ROUGE-1 and ROUGE-L.

Ajwad Abrar, Farzana Tabassum, Sabbir Ahmed

Integrates real-time motion-contact planning with tactile feedback tracking, boosting robustness and precision in multi-finger in-hand manipulation.

Yongpeng Jiang, Mingrui Yu, Xinghao Zhu et al.

DenseGrounding improves 3D visual grounding accuracy by 5.81% using HSSE and LSE modules to enhance visual and textual semantics.

Henry Zheng, Hao Shi, Qihang Peng et al.

Proposes a four-stage roadmap from modular to native multimodal reasoning, emphasizing the potential of N-LMRMs for scalable, autonomous AI.

Yunxin Li, Zhenyu Liu, Zitao Li et al.

VLA models integrate vision, language, and actions, enabling robots to understand and operate autonomously in complex environments.

Ranjan Sapkota, Yang Cao, Konstantinos I. Roumeliotis et al.

ZeroSearch enhances LLM search capabilities via simulated searches, with 14B model surpassing real search engines.

Hao Sun, Zile Qiao, Jiayan Guo et al.