AlphaEvolve combines evolutionary algorithms with large models to discover efficient algorithms, achieving breakthroughs like a 48-multiplication matrix algorithm after 56 years.

Alexander Novikov, Ngân Vũ, Marvin Eisenberger et al.

Adding parallel data significantly boosts LLM multilingual capabilities, especially in translation tasks.

Muhammad Reza Qorib, Junyi Li, Hwee Tou Ng

MIDI-RWKV model enables controllable symbolic music infilling, enhancing music creation efficiency.

Christian Zhou-Zheng, Philippe Pasquier

ContextBench evaluates context modification methods for targeted latent activation, enhanced with model-assisted EPO achieving better activation-fluency balance.

Robert Graham, Edward Stevinson, Leo Richter et al.

KungfuBot employs physics-based whole-body control with adaptive reward tuning, achieving 30% lower tracking error than state-of-the-art, successfully deployed on Unitree G1.

Weiji Xie, Jinrui Han, Jiakun Zheng et al.

ERL-VLM employs absolute ratings from large VLMs to improve reward learning efficiency in RL.

Tung Minh Luu, Younghwan Lee, Donghoon Lee et al.

The BumbleBee framework employs behavior clustering and iterative sim-to-real transfer to enable robust, general humanoid whole-body control, achieving success rates over 89%.

Yuxuan Wang, Ming Yang, Ziluo Ding et al.

s(CASP) enables goal-directed commonsense reasoning, addressing 16+2 trustworthy AI desiderata.

Alexis R. Tudor, Yankai Zeng, Huaduo Wang et al.

Konooz corpus covers 16 Arabic dialects and 10 domains, revealing a 38% performance drop in NER models.

Nagham Hamad, Mohammed Khalilia, Mustafa Jarrar

Proposes MM-R5, a reinforcement learning-based multimodal reranker with explicit reasoning chains, achieving over 4% recall@1 improvement on MMDocIR.

Mingjun Xu, Jinhan Dong, Jue Hou et al.

Proposes a triplane-based multi-camera tokenization, reducing tokens by 72% and doubling inference speed, for end-to-end autonomous driving.

Boris Ivanovic, Cristiano Saltori, Yurong You et al.

DAVID-XR1 integrates fine-grained defect annotations and chain-of-thought reasoning, enabling interpretable AI-generated video detection with strong cross-generator generalization.

Yifeng Gao, Yifan Ding, Hongyu Su et al.

DeepResearch Bench offers 100 PhD-level tasks to evaluate deep research agents' report generation and information retrieval capabilities.

Mingxuan Du, Benfeng Xu, Chiwei Zhu et al.

Proposed occlusion-aware bilayer mask decoder with SAM integration achieves 94.4% mAP on prohibited item segmentation in X-ray images.

Yunhan Ren, Ruihuang Li, Lingbo Liu et al.

QuadricFormer leverages superquadrics for efficient 3D occupancy prediction, achieving 21.11 mIoU with fewer primitives and lower computational cost.

Sicheng Zuo, Wenzhao Zheng, Xiaoyong Han et al.

SEAL framework enables LLMs to self-adapt by generating finetuning data, improving SQuAD accuracy to 47.0%.

Adam Zweiger, Jyothish Pari, Han Guo et al.

VINCIE model learns in-context image editing from videos, achieving state-of-the-art results on multi-turn editing benchmarks.

Leigang Qu, Feng Cheng, Ziyan Yang et al.

CreatiPoster employs a multi-layer protocol model to generate editable graphic compositions, with background synthesis ensuring visual harmony and flexibility.

Dexiang Hong, Zhao Zhang, Weidong Chen et al.

VRBench evaluates multi-step reasoning in long videos, with 960 videos and 8,243 QA pairs.

Jiashuo Yu, Yue Wu, Meng Chu et al.

VideoExplorer employs dynamic reasoning and temporal grounding to outperform baselines in long-video understanding, achieving 54.3% accuracy.

Huaying Yuan, Zheng Liu, Junjie Zhou et al.