By carefully coordinating off-the-shelf models with inference only, we show the DSP framework can achieve surprisingly good results in theorem proving, comparable to the frontier models with RL-based large-scale training.

This paper presents the "Latent Harmony" framework, which achieves synergistic unified Ultra-High Definition (UHD) image restoration through latent space regularization and controllable refinement.

We present Auto-Connect, a method that establishe a new standard for automatic character rigging and skinning, offering a more reliable and effective solution for animation pipelines

We boost masked graph autoencoders for molecules by introducing DyCC, a dynamic, chemistry-aware training framework with adaptive masking (GIBMS) and soft label reconstruction (SLG).

We propose UtilGen, a novel utility-centric data augmentation framework which adaptively optimizes data generation based on downstream task feedback to produce high-utility training data.

We propose CARML, a novel retrieval-augmented editing framework that integrates conflict-aware dynamic retrieval with multi-level collaborative guidance for reliable lifelong multimodal editing.

We introduce Self-Consistency Sampling (SCS), a framework that improves the faithfulness of reasoning trajectory.

We develop a unified large recommender model with intrinsic reasoning capabilities.

We ddvance vision-language action models via introducing comprehensive knowledge prediction.

We show that the pairwise angular structure of pre-trained weights serves as a domain-invariant semantic anchor for continual test-time adaptation.

Identify the purest quantum state among $K$ unknown $n$-qubit quantum states using total $N$ quantum state copies.

This work is the first to address the challenge of large-scale (over 100km^2) urban cross-view LiDAR-based place recognition with high-resolution remote sensing imagery.

We introduce PHYSICS, a dataset containing 16,568 high-quality physics problems spanning subjects and difficulty levels, and an improved evaluation framework combining rule- and model-based judgments to advance LLMs' physical reasoning.

A domain-robust mixture-of-experts for MILP solution prediction across domains

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.

HiFlow is a training-free and model-agnostic framework to unlock the resolution potential of pre-trained flow models.

We propose a novel Dynamic Separator Configuration (DynSep) method that models separator configuration in different rounds as a reinforcement learning task, making decisions based on an incremental triplet graph updated by iteratively added cuts.

preference fading discrete diffusion tailored for recommendation via modeling preference ratios

We propose CLAdapter to refine and adapt the rich representations learned from large-scale data for transfer across diverse, data-limited scientific downstream domains.

We propose a novel progressive 3D editing framework to achieve consistent and precise 3D editing

We propose ModuLM, a flexible framework for LLM-based molecular relational learning, supporting multimodal inputs and dynamic model construction.

Contrastive decoding fails to mitigate object hallucinations in MLLMs—apparent improvements stem from misleading factors rather than genuine effectiveness.

We develop a new framework that combines autoregressive modeling and diffusion for video data.

Our work introduces a GNN-based method for learning SPAI preconditioners, enabling efficient and robust training across various datasets and outperforming traditional and learning-based approaches in accelerating Conjugate Gradient solvers on GPUs.

GoRA improves LoRA by adaptively assigning ranks and initializing weights using gradient info, achieving better performance while staying efficient. It outperforms LoRA and full fine-tuning on GLUE (+5.88) and GSM8k (+5.13) benchmarks.

The paper presents AnyMDP, a framework for procedurally generating diverse tasks to enhance In-Context Reinforcement Learning (ICRL) scalability, and explores the trade-off between generalization and adaptation efficiency.

We propose a projection-based scoring function in KV cache eviction for LLM acceleration.

Advancing Machine-Generated Text Detection from an Easy to Hard Supervision Perspective

This paper introduces BPR+, a novel loss function for graph self-supervised learning that extends Bayesian Personalized Ranking (BPR) by incorporating even-hop paths to better capture global connectivity and topological structure.

To resolve long video understanding task, we propose Deep Video Discovery agent to iterative reasoning and gather information from video content via an agentic search and tool use strategy..

By jointly modeling patient conditions and substructure-aware representations, Our model enhances both accuracy and interpretability, ultimately enabling safer, more personalized drug recommendations.

This paper introduce CoRT, a post-training framework for teaching large reasoning LLMs to leverage CI effectively and efficiently.

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 introduce the first learning-augmented streaming algorithms for Correlation Clustering on both complete and general graphs, improving the best-known space-approximation tradeoffs.

Enhancing Linguistic Steganography Payload in Practice with Large Language Models

We propose a global pruning framework that efficiently learns unstructured sparsity for LLMs.

A practical and super-fast QAT method for general vision applications; source for recognition and detection will be released.

In this work, we address unsupervised domain adaptation for LiDAR segmentation by introducing BeyondMix, a framework that overcomes critical limitations in existing mix-based methods.

We propose DuetGraph, a dual-pathway global-local fusion model with coarse-to-fine optimization that mitigates over-smoothing in KG reasoning, achieving SOTA performance, with up to an 8.7% improvement in quality and a 1.8$\times$ acceleration.

We develop the first HDR 3D measurement framework for event-based structured light systems, enabling accurate 3D reconstruction even under extreme conditions.

We propose an adversarial policy learning framework for humanoid robots, aiming for precise motion tracking of the upper body and robust locomotion of the lower body.

We introduce a novel reasoning paradigm -- pixel-space reasoning. We identified the learning trap when cultivating this ability and proposed a curiosity-driven RL approach to address it.

A Planning Representation and Paradigm Investigation of Vision-Language-Action Models

We propose a novel perspective to understand AI-generated image detection through the lens of OOD detection and introduce a generalizable training-free generated image detection method.

improve both generation quality and efficiency of autoregressive image generation by leveraging entropy of token-wise probability distributions

Common prior knowledge can distort edge weights in DAG structure learning, creating characteristic graph patterns. Our framework detects these patterns as signals to adaptively adjust learning process, improving prior integration and DAG accuracy.

REArtGS

This study introduces VADB, the largest video aesthetic database with 10,490 videos, and VADB-Net, a novel framework that outperforms existing models in video aesthetic assessment.

We introduce Mesh-RFT, a novel framework leveraging localized preference optimization with topology-aware scoring to achieve state-of-the-art high-fidelity 3D mesh generation.

Optimize KV cache eviction by adaptively allocating budgets across different attention heads for efficient LLM inference

A VLM method for vehicles and robots.

An Effective Levelling Paradigm for Unlabeled Scenarios

Loong-X enables hands-free image editing using multimodal neural signals, achieving performance comparable to text-driven methods by combining BCIs with the proposed diffusion-based generative methods.

We propose XLFM-Former, a physics-guided and self-supervised Transformer framework for efficient volumetric reconstruction in light field microscopy, achieving state-of-the-art performance with minimal supervision.

We introduce rStar-Coder to train advanced code reasoning LLMs, with our 14B model achieving comparable performance to QWQ-32B.

This paper presents HumanoidGen, an automated task creation and demonstration collection framework that leverages atomic dexterous skills and LLM reasoning to generate relational constraints.

We enhance long-term time series forecasting by proposing MEW, a new metric, and two modules (IBF and HTM) that help Transformers better use historical data.

A vector database watermarking scheme with fewer query errors is proposed.