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.

Kraken: enabling joint trajectory prediction by utilizing Mode Transformer and Greedy Mode Processing

Accurate and reliable motion prediction is essential for safe urban autonomy. The most prominent motion prediction approaches are based on modeling the distribution of possible future trajectories of each actor in autonomous system's vicinity. These "independent" marginal predictions might be accurate enough to properly describe casual driving situations where the prediction target is not likely to interact with other actors. They are, however, inadequate for modeling interactive situations where the actors' future trajectories are likely to intersect. To mitigate this issue we propose Kraken -- a real-time trajectory prediction model capable of approximating pairwise interactions between the actors as well as producing accurate marginal predictions. Kraken relies on a simple Greedy Mode Processing technique allowing it to convert a factorized prediction for a pair of agents into a physically-plausible joint prediction. It also utilizes the Mode Transformer module to increase the diversity of predicted trajectories and make the joint prediction more informative. We evaluate Kraken on Waymo Motion Prediction challenge where it held the first place in the Interaction leaderboard and the second place in the Motion leaderboard in October 2021.