Proposes Loss-Free Balancing, a method that dynamically adjusts expert biases to improve load balance without interference gradients, outperforming auxiliary loss methods.

Lean Wang, Huazuo Gao, Chenggang Zhao et al.

CoT-SelfEvolve combines StackOverflow retrieval and self-correction to improve DS-1000 code generation.

Thai Tang Quoc, Duc Ha Minh, Tho Quan Thanh et al.

Proposes a grand canonical diffusion model using continuous voxel representation, successfully generating crystalline phases and grain boundaries.

Bo Lei, Enze Chen, Hyuna Kwon et al.

DC² framework enhances MLLM perception of 4K&8K images, improving accuracy by 6%.

Wenbin Wang, Liang Ding, Minyan Zeng et al.

Categorizes concept models into four types—Abstractionism, Similarity, Functional, Invariance—highlighting their mathematical structures.

Jun Otsuka

Distilling large Transformers into linear RNN (Mamba) with hardware-aware decoding boosts inference speed, matching original performance.

Junxiong Wang, Daniele Paliotta, Avner May et al.

Symbolic working memory enhances language models for complex rule application, significantly improving multi-step reasoning accuracy.

Siyuan Wang, Zhongyu Wei, Yejin Choi et al.

Robust optimization-based confidence bands for stochastic processes, reducing sample size by an order of magnitude.

Timothy Chan, Jangwon Park, Vahid Sarhangian

Show-o is a unified transformer integrating autoregressive and discrete diffusion models, supporting multimodal understanding and generation with 1.3B parameters, achieving state-of-the-art performance.

Jinheng Xie, Weijia Mao, Zechen Bai et al.

RTF framework combines hierarchical ReAct with dynamic tool retrieval, significantly improving forecasting accuracy beyond baseline models and rivaling human predictions.

Elvis Hsieh, Preston Fu, Jonathan Chen

Introduces GRAB, a synthetic graph analysis benchmark with 3284 questions, evaluating 20 LMMs; top score 21%, highlighting current limitations.

Jonathan Roberts, Kai Han, Samuel Albanie

Proposes 'Sinkhorn bridge' for efficient Schrödinger bridge estimation via static entropic OT potentials, avoiding iterative diffusion simulations.

Aram-Alexandre Pooladian, Jonathan Niles-Weed

LARR integrates domain knowledge and contrastive learning to enhance real-time scene semantic understanding for CTR prediction.

Zhizhong Wan, Bin Yin, Junjie Xie et al.

TrackGo employs free masks and arrows for precise, efficient controllable video generation, integrating TrackAdapter into temporal self-attention layers for enhanced control.

Haitao Zhou, Chuang Wang, Rui Nie et al.

Quantum SVM and variational circuits enable faster, more accurate anomaly detection, demonstrating significant performance gains over classical methods.

Sebastiano Corli, Lorenzo Moro, Daniele Dragoni et al.

Train matting models without alpha labels using DDC loss, achieving excellent results on AM-2K and P3M-10K datasets.

Wenze Liu, Zixuan Ye, Hao Lu et al.

Goldfish: a suite of 1000+ monolingual models for 350 low-resource languages, trained with 125M parameters, outperforming large multilingual models in perplexity and grammaticality.

Tyler A. Chang, Catherine Arnett, Zhuowen Tu et al.

Introduced two GPU-based Boolean Matrix Logic Programming algorithms, achieving 1-4 orders of magnitude speedup in large-scale inference tasks.

Lun Ai

Proposes MOHAWK: a multi-stage distillation method transforming pretrained Transformers into efficient state-space models, boosting performance with minimal data.

Aviv Bick, Kevin Y. Li, Eric P. Xing et al.

This study introduces a physics-aware, data-free action masking reinforcement learning framework for assembly sequence planning, achieving 100% success on 250+ LEGO structures.

Ruixuan Liu, Alan Chen, Weiye Zhao et al.