Proposes multi-scale dynamics and Gaussian sequences for monocular 4D scene reconstruction, achieving state-of-the-art results with 17.07 dB PSNR.

Can Li, Jie Gu, Jingmin Chen et al.

IGOFormer adds geometry-aware decoding and momentum matching, reaching 78.00 AP50 on DOTA-V1.0 with Swin-T.

Junpeng Zhang, Zewei Yang, Jie Feng et al.

ThunderAgent employs program abstraction and dynamic scheduling to boost throughput by 1.5-3.6×, reducing KV cache thrashing and resource waste.

Hao Kang, Ziyang Li, Weili Xu et al.

OpAgent achieves a 71.6% success rate in web navigation via online RL.

Yuyu Guo, Wenjie Yang, Siyuan Yang et al.

Simplified on-policy supervised fine-tuning (SFT) with truncation-based length reward improves reasoning efficiency and training speed, outperforming complex RL methods.

Anhao Zhao, Ziyang Chen, Junlong Tong et al.

CoPE-VideoLM leverages codec primitives (motion vectors, residuals) to reduce token usage by 93% and TTFT by 86%, enabling efficient long-term video understanding.

Sayan Deb Sarkar, Rémi Pautrat, Ondrej Miksik et al.

Integrating stratified sampling with model-assisted estimation improves accuracy in costly outcome measurements.

Reagan Mozer, Nicole E. Pashley, Luke Miratrix

SciAgentGym enhances scientific tool-use via SciForge; SciAgent-8B outperforms Qwen3-VL-235B-Instruct.

Yujiong Shen, Yajie Yang, Zhiheng Xi et al.

INHerit-SG introduces hierarchical semantic scene graphs with RAG-style retrieval, significantly improving complex query handling in robot navigation.

YukTungSamuel Fang, Zhikang Shi, Jiabin Qiu et al.

Proposes a nonparametric distribution calibration method based on conditional kernel mean embeddings, significantly improving calibration accuracy.

Ádám Jung, Domokos M. Kelen, András A. Benczúr

ALOE integrates Q-chunking and conservative value aggregation for off-policy evaluation, boosting real-world VLA policy stability and performance.

Rushuai Yang, Hecheng Wang, Zhichao Wu et al.

SkillsBench benchmark shows curated skills improve task pass rate by 16.6 percentage points.

Xiangyi Li, Yimin Liu, Wenbo Chen et al.

Proposes PACED-RL, leveraging the partition function as an online accuracy signal to improve sample efficiency in distribution-matching training of LLMs.

Dohyung Kim, Minbeom Kim, Jeonghye Kim et al.

Self-EvolveRec employs LLM-guided directional feedback with user simulators and model diagnosis tools to autonomously evolve recommender systems, outperforming NAS and code evolution baselines.

Sein Kim, Sangwu Park, Hongseok Kang et al.

DiffuRank leverages diffusion language models for document reranking, enabling parallel decoding and outperforming comparable autoregressive models in efficiency and effectiveness.

Qi Liu, Kun Ai, Jiaxin Mao et al.

Proposes a modular skill-based architecture for LLM agents, emphasizing skill loading, acquisition, and security, fostering trustworthy self-improving systems.

Renjun Xu, Yang Yan

Proposes On-Policy Context Distillation (OPCD), minimizing reverse KL divergence on self-generated trajectories, boosting knowledge internalization for tasks like math reasoning and transfer.

Tianzhu Ye, Li Dong, Xun Wu et al.

Few-Step Diffusion Language Models via Trajectory Self-Distillation improve reasoning tasks.

Tunyu Zhang, Xinxi Zhang, Ligong Han et al.

Proposes Categorical Flow Maps, leveraging continuous paths for rapid few-step categorical data generation with state-of-the-art results.

Daan Roos, Oscar Davis, Floor Eijkelboom et al.

LDA-1B employs a unified multi-task framework with structured DINO latent space and multimodal diffusion transformer, scaling to 1.6B parameters with over 30k hours of heterogeneous embodied data, outperforming prior methods by up to 21%.

Jiangran Lyu, Kai Liu, Xuheng Zhang et al.