Insights from NeurIPS 2025: Modeling Distributions and Venturing Beyond Them

NeurIPS 2025 Overview

From December 2–7, 2025, the Conference on Neural Information Processing Systems (NeurIPS) was held in San Diego, USA. The conference received 21,575 valid submissions and accepted 5,290 papers, yielding an overall acceptance rate of 24.52%.

NIRC Contributions

The Network Intelligent Research Center (NIRC) had four papers accepted, with three presented as posters by Liu Xingyu:

Unified 2D-3D Discrete Priors for Noise‑Robust and Calibration‑Free Multiview 3D Human Pose Estimation – The UniCodebook paper addresses the lack of robustness to input noise in transformer‑based multiview 3D pose methods. It introduces a unified 2D‑3D discrete codebook (UniCodebook) and a discrete‑continuous space attention (DCSA) mechanism, achieving higher accuracy while significantly improving noise resistance without requiring camera calibration. Experiments on three mainstream datasets show performance surpassing existing methods.

Do LVLMs Truly Understand Video Anomalies? Revealing Hallucination via Co‑Occurrence Patterns – The authors find that large‑vision‑language models (LVLMs) are often misled by visual‑text co‑occurrence patterns learned during training, producing hallucinated anomalies in normal scenes. This work systematically diagnoses the issue and proposes VAD‑DPO, a preference‑based semantic contrast optimization that markedly enhances scene understanding and anomaly detection reliability.

Generalizable Hand‑Object Modeling from Monocular RGB Images via 3D Gaussians – To tackle the high cost of dense 3D annotations or pre‑scanned object models, the paper presents HOGS, a hand‑object interaction reconstruction framework based on 3D Gaussian Splatting (3DGS). HOGS uses adaptively sensed 3D Gaussian primitives to achieve high‑quality hand‑object modeling from single‑view RGB images, maintaining strong generalization to unseen environments and complex motions.

Poster Discussions

In the poster area, a large‑scale pre‑training method for 3D scene reconstruction based on feed‑forward generation sparked discussion about the trade‑off between inference efficiency and structural fidelity, the generalization of domain‑out scenes, and the feasibility of unified 2D‑3D representation learning.

Industry Exhibition Highlights

Several leading AI companies showcased their latest visual technologies. Tesla demonstrated an end‑to‑end visual perception solution for autonomous driving and embodied intelligence, emphasizing large‑scale data‑driven industrial vision. Meta highlighted advances in general‑purpose visual perception, including SAM3 and its 3D extension SAM3D, illustrating the accelerating convergence of academic research and industry practice in foundational vision models and 3D perception.