🏢 Tsinghua University

LLMs’ training quality is exponentially improving, enabling models with half the parameters to match state-of-the-art performance every 3 months, thus reducing inference costs.

MIDI: a novel multi-instance diffusion model generates compositional 3D scenes from single images by simultaneously creating multiple 3D instances with accurate spatial relationships and high generali…

2DGS-Room: Seed-guided 2D Gaussian splatting with geometric constraints achieves state-of-the-art high-fidelity indoor scene reconstruction.

X-Prompt: a novel autoregressive vision-language model achieves universal in-context image generation by efficiently compressing contextual information and using a unified training framework for super…

Unified 3D latent representation (SLAT) enables versatile high-quality 3D asset generation, significantly outperforming existing methods.

Train high-performing Process Reward Models (PRMs) cheaply using only outcome-level labels, eliminating the need for costly step-by-step annotations!

AlphaTablets: A novel 3D plane representation enabling accurate, consistent, and flexible 3D planar reconstruction from monocular videos, achieving state-of-the-art results.

Researchers created TemplateGSM, a massive dataset of 7M+ grade-school math problems and solutions, using GPT-4 to generate templates, significantly advancing LLM training for mathematical reasoning.

TAPTRv3 achieves state-of-the-art long-video point tracking by cleverly using spatial and temporal context to enhance feature querying, surpassing previous methods and demonstrating strong performance…

HiAR-ICL, a novel automated reasoning paradigm using Monte Carlo Tree Search, surpasses state-of-the-art accuracy in complex mathematical reasoning by shifting focus from specific examples to abstract…

Boosting Large Language Model (LLM) reasoning without massive datasets: A novel training method encourages ‘patient’ reasoning, improving accuracy by up to 6.7% on benchmark tasks.

SageAttention2 achieves 4-bit accurate attention, boosting inference speed by 2x compared to FlashAttention2, while maintaining end-to-end accuracy across diverse models.

AnimateAnything: A unified approach enabling precise & consistent video manipulation via a novel optical flow representation and frequency stabilization.

Boosting multimodal reasoning in LLMs, researchers developed Mixed Preference Optimization (MPO) and a large-scale dataset (MMPR), significantly improving reasoning accuracy and achieving performance …

LLaMA-Mesh: Unifying 3D mesh generation with LLMs by directly representing meshes as text, enabling efficient text-to-3D conversion within a single model.

Sharingan extracts user action sequences from desktop recordings using novel VLM-based methods, achieving 70-80% accuracy and enabling RPA.

JanusFlow harmonizes autoregression and rectified flow for unified multimodal understanding and generation, achieving state-of-the-art results on standard benchmarks.

DimensionX generates photorealistic 3D and 4D scenes from a single image via controllable video diffusion, enabling precise manipulation of spatial structure and temporal dynamics.

WEBRL: A self-evolving online curriculum reinforcement learning framework empowers open LLMs to excel as high-performing web agents, surpassing proprietary models.

Researchers discovered predictable scaling laws for activation sparsity in LLMs, showing how data, architecture, and model size influence sparsity, paving the way for more efficient and interpretable …