Humanity's Last Exam

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 3,000 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Open-Vocabulary Segmentation with Unpaired Mask-Text Supervision

Contemporary cutting-edge open-vocabulary segmentation approaches commonly rely on image-mask-text triplets, yet this restricted annotation is labour-intensive and encounters scalability hurdles in complex real-world scenarios. Although some methods are proposed to reduce the annotation cost with only text supervision, the incompleteness of supervision severely limits the versatility and performance. In this paper, we liberate the strict correspondence between masks and texts by using independent image-mask and image-text pairs, which can be easily collected respectively. With this unpaired mask-text supervision, we propose a new weakly-supervised open-vocabulary segmentation framework (Uni-OVSeg) that leverages confident pairs of mask predictions and entities in text descriptions. Using the independent image-mask and image-text pairs, we predict a set of binary masks and associate them with entities by resorting to the CLIP embedding space. However, the inherent noise in the correspondence between masks and entities poses a significant challenge when obtaining reliable pairs. In light of this, we advocate using the large vision-language model (LVLM) to refine text descriptions and devise a multi-scale ensemble to stablise the matching between masks and entities. Compared to text-only weakly-supervised methods, our Uni-OVSeg achieves substantial improvements of 15.5% mIoU on the ADE20K datasets, and even surpasses fully-supervised methods on the challenging PASCAL Context-459 dataset.

An Efficient Transformer for Simultaneous Learning of BEV and Lane Representations in 3D Lane Detection

Accurately detecting lane lines in 3D space is crucial for autonomous driving. Existing methods usually first transform image-view features into bird-eye-view (BEV) by aid of inverse perspective mapping (IPM), and then detect lane lines based on the BEV features. However, IPM ignores the changes in road height, leading to inaccurate view transformations. Additionally, the two separate stages of the process can cause cumulative errors and increased complexity. To address these limitations, we propose an efficient transformer for 3D lane detection. Different from the vanilla transformer, our model contains a decomposed cross-attention mechanism to simultaneously learn lane and BEV representations. The mechanism decomposes the cross-attention between image-view and BEV features into the one between image-view and lane features, and the one between lane and BEV features, both of which are supervised with ground-truth lane lines. Our method obtains 2D and 3D lane predictions by applying the lane features to the image-view and BEV features, respectively. This allows for a more accurate view transformation than IPM-based methods, as the view transformation is learned from data with a supervised cross-attention. Additionally, the cross-attention between lane and BEV features enables them to adjust to each other, resulting in more accurate lane detection than the two separate stages. Finally, the decomposed cross-attention is more efficient than the original one. Experimental results on OpenLane and ONCE-3DLanes demonstrate the state-of-the-art performance of our method.