SWE-Synth uses LLMs to simulate debugging workflows, synthesizing verifiable bug-fix data, improving open-source model training by 2.3%.

Minh V. T. Pham, Huy N. Phan, Hoang N. Phan et al.

OmniV-Med unifies 2D/3D images and videos in a medical vision-language model, using rotary position encoding and token pruning to achieve SOTA performance across 7 benchmarks.

Songtao Jiang, Yuan Wang, Sibo Song et al.

RoboOcc integrates Opacity-guided Self-Encoder and Geometry-aware Cross-Encoder to improve 3D scene understanding, outperforming SOTA with 8.47% IoU gain.

Zhang Zhang, Qiang Zhang, Wei Cui et al.

Proposes AIOS server for decentralized AgentSite network, enabling scalable peer-to-peer communication and discovery.

Xiang Zhang, Yongfeng Zhang

This study critically examines RLVR's ability to induce novel reasoning in LLMs, revealing paths are bounded by base models, with distillation offering better expansion.

Yang Yue, Zhiqi Chen, Rui Lu et al.

CheXWorld employs a self-supervised world model with multi-task learning, capturing local structures, global layouts, and domain shifts, achieving state-of-the-art results on 8 benchmarks.

Yang Yue, Yulin Wang, Chenxin Tao et al.

CodeVisionary employs a two-stage multi-agent framework combining requirement-driven context distillation and collaborative scoring for complex code evaluation.

Xinchen Wang, Pengfei Gao, Chao Peng et al.

PerceptionLM employs multi-stage training with 2.8M human-annotated video QA and spatio-temporal captions, advancing transparent vision-language modeling.

Jang Hyun Cho, Andrea Madotto, Effrosyni Mavroudi et al.

Miras framework unifies sequence models via associative memory, diverse attentional biases, and retention gates, outperforming some SOTA models.

Ali Behrouz, Meisam Razaviyayn, Peilin Zhong et al.

Introduces sleep-time compute to reduce test-time compute requirements by pre-computing context, improving accuracy.

Kevin Lin, Charlie Snell, Yu Wang et al.

Deep neural networks excel in representation learning for tabular data, enhancing accuracy in classification and regression tasks.

Jun-Peng Jiang, Si-Yang Liu, Hao-Run Cai et al.

Nemotron-CLIMB employs semantic embedding and iterative search to optimize data mixtures, boosting language model pre-training performance.

Shizhe Diao, Yu Yang, Yonggan Fu et al.

Proposes Uni-Inv and Uni-Edit for accurate inversion and region-aware editing in flow models, achieving low error rates and natural edits.

Guanlong Jiao, Biqing Huang, Kuan-Chieh Wang et al.

This study assesses international AI safety cooperation risks, highlighting verification mechanisms and shared protocols as promising areas for secure collaboration.

Ben Bucknall, Saad Siddiqui, Lara Thurnherr et al.

MLRBench evaluates LLMs' reasoning over multilingual long contexts, revealing resource gaps.

Amey Hengle, Prasoon Bajpai, Soham Dan et al.

Persona-judge achieves personalized alignment via self-judgment, enhancing alignment efficiency by 98%.

Xiaotian Zhang, Ruizhe Chen, Yang Feng et al.

Proposes ForgetMe dataset and Entangled metric for selective forgetting in diffusion models, using prompt-layered editing and training-free local feature removal.

Zhenyu Yu, Mohd Yamani Inda Idris, Pei Wang

Proposes an effective semantic diversity metric for LLM outputs, revealing preference tuning enhances high-quality content diversity.

Alexander Shypula, Shuo Li, Botong Zhang et al.

BrowseComp is a challenging benchmark for AI web browsing, comprising 1266 questions requiring persistent, creative search for hard-to-find information.

Jason Wei, Zhiqing Sun, Spencer Papay et al.

RLHF uses preference models and PPO to align language models with human preferences, improving safety and style consistency.

Nathan Lambert