🏢 Tsinghua University

SparseFlex: Achieves high-res, arbitrary-topology 3D shape modeling via sparse isosurface representation and sectional voxel training. Revolutionizing 3D generative AI!

ReaRAG enhances factuality in large reasoning models (LRMs) by integrating knowledge-guided reasoning with iterative retrieval augmented generation.

BIZGEN: Article-level Visual Text Rendering for Infographics Generation

Video-T1 enhances video generation through test-time scaling, improving quality and consistency by viewing generation as a search for optimal video trajectories.

LEMMA: LLMs learn math via mistake analysis and correction, boosting performance without external critics.

XAttention: Antidiagonal scoring unlocks block-sparse attention, slashing compute costs in long-context Transformers without sacrificing accuracy.

DeepMesh: RL-guided auto-regressive creation of artist-quality 3D meshes, enhanced by tokenization & DPO for human-aligned aesthetics.

DiffMoE: Dynamically selects tokens for scalable diffusion transformers, unlocking new efficiency levels in image generation.

DAPO: Open-sources a LLM reinforcement learning system that achieves SOTA AIME scores, fostering reproducible research at scale.

DeepPerception enhances MLLMs with cognitive visual perception, achieving superior grounding through knowledge integration & reasoning.

UniGoal: A novel framework for universal zero-shot goal-oriented navigation, outperforming task-specific methods with a unified approach.

KUDA unifies dynamics learning and visual prompting with keypoints for open-vocabulary robot manipulation.

GTR: Prevents thought collapse in RL-based VLM agents by process guidance, enhancing performance in complex visual reasoning tasks.

ProReflow: Improves diffusion model efficiency via progressive training and direction-focused velocity alignment.

SAGE: Precise RAG via semantic segmentation, adaptive chunking, and LLM feedback, boosting QA accuracy & cost-efficiency.

PQI: Accurately identify sensitive weights in post-quantization to enhance LLM compression & performance!

ArtGS: Achieves state-of-the-art, efficient interactable replicas of complex articulated objects via Gaussian Splatting.

FR-Spec: Frequency-Ranked Speculative Sampling accelerates LLMs by optimizing vocabulary space compression, achieving 1.12x speedup over EAGLE-2.

JL1-CD: New all-inclusive dataset & multi-teacher knowledge distillation framework for robust remote sensing change detection, achieving state-of-the-art results!

CRAW4LLM: Efficiently crawls web pages for LLM pretraining by prioritizing influence scores, boosting data quality & cutting crawling waste.