Liquid AI's d1-ultra ranks #1 on Artificial Analysis Decision Index

Source published 1m agoTerminal published just now

Summary

Liquid AI has launched d1-ultra, which has taken the top spot on the Artificial Analysis Decision Index v1.0 with a score of 65, outperforming competitors like Jev 1.14 and Microsoft-Decision-1. The new decision model, which follows the earlier release of d1, boasts a significantly improved non-hallucination rate of 85%, indicating its ability to avoid answering when uncertain. However, this comes with a trade-off, as d1-ultra's accuracy has decreased to 32%, down from 41% for d1. This launch is part of Liquid AI's strategy to meet the growing demand for effective decision models that provide structured probabilities, as evaluated through benchmarking frameworks by Artificial Analysis, which assess AI capabilities such as decision accuracy and task efficiency.

Analysis

d1-ultra: d1-ultra is a decision model developed by Liquid AI as a follow-up to its d1 model released earlier in the same month. It is designed for scenarios where models must select from predefined options while accurately signaling uncertainty to reduce incorrect answers. The model is positioned within the emerging category of decision-focused AI systems evaluated on dedicated industry indices. Liquid AI: Liquid AI is an AI company specializing in decision models engineered to deliver probability-weighted choices from a fixed set of options in a single inference pass. The company focuses on building models optimized for calibration, speed, and controlled abstention in decision-oriented tasks. Its latest release, d1-ultra, directly extends this work by improving performance on benchmarks measuring knowledge boundaries. Benchmarking: Artificial Analysis publishes composite indices to evaluate AI models on specialized capabilities such as decision accuracy, hallucination resistance, and task efficiency. Decision Models: Decision models generate structured probability outputs over a fixed set of options in one call, differing from language models that produce open-ended text responses.

Summary

Analysis

Categories