A unified pizza–sandwich–lasagna framework shows that 2× KV-cache reduction usually preserves quality, while aggressive compression favors sandwich-middle.

You Wu, Haoyi Wu, Kewei Tu

FSBM leverages limited pre-aligned pairs to improve distribution matching efficiency and generalization in semi-supervised Schrödinger bridge framework.

Panagiotis Theodoropoulos, Nikolaos Komianos, Vincent Pacelli et al.

MemTree employs a dynamic tree-structured memory, significantly improving long-term memory management for LLMs, excelling in multi-turn dialogue and document reasoning.

Alireza Rezazadeh, Zichao Li, Wei Wei et al.

Proposed a biologically motivated evaluation framework; found scVI and PCA outperform emerging models in perturbation analysis.

Ihab Bendidi, Shawn Whitfield, Kian Kenyon-Dean et al.

DreamVideo-2 achieves zero-shot subject-driven video customization using reference attention and mask-guided motion, without test-time fine-tuning, outperforming SOTA on a new dataset.

Yujie Wei, Shiwei Zhang, Hangjie Yuan et al.

V-GPS leverages offline RL-trained value functions for test-time action re-ranking, boosting five state-of-the-art robotic policies by 82.8% across 12 tasks, with no fine-tuning needed.

Mitsuhiko Nakamoto, Oier Mees, Aviral Kumar et al.

First performance guarantee for score-mismatched diffusion models with explicit dimension dependence; analyzes zero-shot conditional sampling bias and convergence.

Yuchen Liang, Peizhong Ju, Yingbin Liang et al.

This study evaluates 24 multilingual LLMs within a real-world Indian healthcare chatbot setting using a unified RAG framework, focusing on factual correctness and cultural relevance.

Varun Gumma, Ananditha Raghunath, Mohit Jain et al.

OpenAI's o1 model excels in reasoning tasks like math and coding using Test-time Compute methods.

Siwei Wu, Zhongyuan Peng, Xinrun Du et al.

Deep model merging links model merging with neural network training phenomena via loss landscape geometry, revealing new directions for model interpretability.

Arham Khan, Todd Nief, Nathaniel Hudson et al.

This study evaluates 10 top software development agents on 500 real GitHub issues, analyzing patch patterns, code quality, and complexity with detailed metrics.

Zhi Chen, Lingxiao Jiang

Introduces MC-Bench, a benchmark for evaluating multi-image visual grounding by 20+ models, highlighting performance gaps and future directions.

Yunqiu Xu, Linchao Zhu, Yi Yang

FlashAudio uses rectified flows for fast, high-fidelity text-to-audio generation, achieving 400x speedup.

Huadai Liu, Jialei Wang, Rongjie Huang et al.

DeCo dynamically integrates preceding-layer knowledge into decoding, reducing hallucination rates by 10.8% in multimodal large language models.

Chenxi Wang, Xiang Chen, Ningyu Zhang et al.

Using ASP to generate global and local explanations for tree-ensemble learning methods, enhancing model interpretability.

Akihiro Takemura, Katsumi Inoue

Li and Zhou reveal that MoE models' routing weights (RW) serve as effective embeddings without fine-tuning, outperforming hidden states (HS) when combined.

Ziyue Li, Tianyi Zhou

SANA employs deep compression autoencoder and linear DiT to generate 4096×4096 images efficiently on a single GPU, achieving 20× smaller size and 100× faster inference.

Enze Xie, Junsong Chen, Junyu Chen et al.

VisRAG employs vision-language models for end-to-end multi-modal document retrieval and generation, achieving 20-40% performance gains over traditional text-based RAG.

Shi Yu, Chaoyue Tang, Bokai Xu et al.

SimBa injects simplicity bias to scale deep RL networks, improving sample efficiency and surpassing SOTA with fewer resources.

Hojoon Lee, Dongyoon Hwang, Donghu Kim et al.

FlatQuant employs learnable affine transformations with Kronecker decomposition to achieve sub-1% accuracy loss in W4A4 quantization of LLaMA-3-70B, boosting speed by 2.3x.

Yuxuan Sun, Ruikang Liu, Haoli Bai et al.