Welcome to the Best of NeurIPS meetup, your virtual pass to some of the groundbreaking research, insights, and innovations that defined this year’s conference. Live streaming from the authors to you.

Operating rooms (ORs) demand precise coordination among surgeons, nurses, and equipment in a fast-paced, occlusion-heavy environment, necessitating advanced perception models to enhance safety and efficiency. Existing datasets either provide partial egocentric views or sparse exocentric multi-view context, but do not explore the comprehensive combination of both. We introduce EgoExOR, the first OR dataset and accompanying benchmark to fuse first-person and third-person perspectives. Spanning 94 minutes (84,553 frames at 15 FPS) of two emulated spine procedures, Ultrasound-Guided Needle Insertion and Minimally Invasive Spine Surgery, EgoExOR integrates egocentric data (RGB, gaze, hand tracking, audio) from wearable glasses, exocentric RGB and depth from RGB-D cameras, and ultrasound imagery. Its detailed scene graph annotations, covering 36 entities and 22 relations (568,235 triplets), enable robust modeling of clinical interactions, supporting tasks like action recognition and human-centric perception. We evaluate the surgical scene graph generation performance of two adapted state-of-the-art models and offer a new baseline that explicitly leverages EgoExOR's multimodal and multi-perspective signals. This new dataset and benchmark set a new foundation for OR perception, offering a rich, multimodal resource for next-generation clinical perception.

Few-shot segmentation requires recognizing novel object categories from only a few annotated examples, demanding both accurate mask generation and strong visual correspondence. While Segment Anything 2 (SAM2) provides powerful prompt-based segmentation and built-in feature matching, its representations are entangled with tracking-specific cues that limit higher-level semantic generalization. We show that SAM2 nonetheless encodes rich latent semantic structure despite its class-agnostic training. To leverage this, we introduce SANSA, a lightweight framework that makes this structure explicit and adapts SAM2 for few-shot segmentation with minimal modifications. SANSA achieves state-of-the-art generalization performance, outperforms generalist in-context methods, supports flexible prompting, and remains significantly faster and smaller than prior approaches.

We present Nested Learning (NL), a new learning paradigm for continual learning that views machine learning models and their training process as a set of nested and/or parallel optimization problems, each of which with its own context flow, frequency of update, and learning algorithm. Based on NL, we design a new architecture, called Hope, that is capable of continual learning and also modifying itself, if it is needed.

VLMs sound convincing—but are their explanations actually true? This talk introduces Explanation-Driven Counterfactual Testing (EDCT), a simple and model-agnostic method that evaluates whether VLM explanations align with the evidence models truly use. By perturbing the very features a model claims to rely on, EDCT exposes mismatches between stated reasoning and real decision pathways. I will show surprising failure cases across state-of-the-art VLMs and highlight how EDCT can guide more trustworthy explanation methods.