We prove generalization bounds for neural networks which exploit approximate low-rank structure in the weights.

We propose SRPO, a reflection-aware RL method that significantly improves multimodal LLM reasoning by explicitly teaching self-reflection, outperforming state-of-the-art models on multiple benchmarks.

This work utilizes vision foundation models to construct a visual tokenizer, which is trained in an end-to-end manner for AR image generation, achieving state-of-the-art results on the $256\times256$ class-to-image generation task on ImageNet.

This paper explores the representation of robot actions in the frequency domain and introduces FreqPolicy, a novel frequency-domain autoregressive framework for hierarchical robotic action generation.

EAG3R synergistically integrates standard RGB video frames with asynchronous event streams to achieve robust 3D geometry estimation in dynamic and extreme-lighting scenes where conventional RGB-only approaches struggle.

The state-of-the-art open-sourced LLM RL algorithm, dataset and infrastructure

We introduce MMOral, the first large-scale multimodal instruction dataset and benchmark tailored for panoramic X-ray interpretation. MMOral-Bench is a comprehensive evaluation suite covering five key diagnostic dimensions in dentistry.

This article proposes the first refined dataset, A Fine-Grained Global Terraced Parcel and Boundary Dataset(GTPBD), covering typical terraced field distribution areas worldwide. It is innovative and supports multi task evaluation.

We propose SAS to simulate larger attention head numbe and hidden size per head for better performance, keeping the original model size.

We introduce a novel embodied VLM agent with a VLM fine-tuned by agentic data synthesis for open-world mobile manipulation, unifying scene understanding, state tracking, and action generation for state-of-the-art results.

We propose a unified framework that distills knowledge from monocular geometry models to improve the performance of feed-forward reconstruction networks.

We trained MeshLLM, an Object-to-Code inference framework that generates Blender Python scripts to reconstruct 3D meshes from point clouds in a structured and editable manner.

Our approach improves Vision Transformer Dense Representations via Self-Distillation

We provide a refined analysis on the excess risk of kernel ridge regression in structured non-i.i.d. setting and apply it to denoising score learning.

A training-free framework for high-fidelity textured 3D morphing via Flow Models.

We propose a four-stage training task for LLMs—Fast Thinking, Verification, Slow Thinking, and Summarization—separating the training of intuition, verification, reasoning, and integration.

A data construction pipeline and a DiT framework for subject-driven video customization under multimodal control conditions

we propose Seg-VAR, a novel framework that rethinks segmentation as a conditional autoregressive mask generation problem

LiteReality is an automatic pipeline that converts RGB-D scans of indoor environments into graphics-ready 3D scenes, featuring high-quality meshes with realistic PBR materials and articulated objects.

Regularize the model attention in a transformer (VLM) using human gaze to enhance egocentric activity prediction and understanding.

A multimodal human activity and pose dataset for diversify sensing including acoustic, RF-based, vision-based, inertial, and physiological modalities

We present a novel pair of benchmarks to evaluate the fundamental deductive reasoning abilities of test-time compute reasoning models on program semantics.

We propose Magical, an asymmetric LoRA architecture to meet the requirement for semantic fidelity and diverse lay-style generation in medical lay language generation task.

This study pioneers a framework to address theoretical bias in RWKV and explore its potential for multi-spectral and panchromatic fusion, bridging a key gap in remote sensing image fusion.

We propose a novel DP-driven multimodal learning framework that automatically optimizes the balance between prominent intra-modal representation learning and cross-modal alignment.

We introduce ACM, a framework that enhances model merging by incorporating layer-specific merging coefficients based on activation mutual information.

a unified multimodal model purely based on discrete flow matching, achieving comparable performance with AR-based MLLMs

We meta-learn a transformer based in-context learning fMRI visual cortex encoder which can adapt to new human subjects without any fine-tuning

MedAgentBoard finds multi-agent useful for some medical tasks, but single LLMs or conventional methods are often better, especially for VQA/EHR. Task-specific AI choice is key.

We propose a progressive consistency distillation framework that enhances the efficiency of MLLMs by significantly reducing computational cost while preserving strong performance.

We distill commonsense spatial priors from large language models into a flow-based generative model to help agents imagine unseen parts of indoor scenes for Object Goal Navigation task.

A framework for generating modular building layouts from textual descriptions driven by code

Improving 3D representation learning for multimodal large language models.

The paper proposes an LLM-guided knowledge graph traversal method that enhances retrieval efficiency and accuracy for complex queries while reducing costs.

We introduce SEAL, a semantic-aware hierarchical learning framework that leverages real-world coarse-to-fine taxonomies to provide stronger supervision for GCD.

we established optimal risk bound of $1/(\gamma^2 n)$ for GD with deep ReLU networks

We propose TreeSynth, a tree-guided subspace-based data synthesis approach, achieving superior data diversity, model performance, robust scalability, and data balance efficacy.

We propoed the first application of RLVR to directly enhance LLMs' proficiency in optimization modeling: SI-RL, achieves state-of-the-art on diverse public benchmarks

Instruction-based Video Editing Dataset