Скачать или смотреть ORION: Compact Reasoning for Large LMs

In this AI Research Roundup episode, Alex discusses the paper: 'ORION: Teaching Language Models to Reason Efficiently in the Language of Thought(2511.22891v1)' This work introduces ORION, a framework that trains large reasoning models to think in a compact, symbolic "Mentalese" instead of long, verbose chains of reasoning. By encoding abstract reasoning as ultra-compressed, structured tokens, the models can solve complex math and planning tasks with far fewer steps. The authors also propose Shorter Length Preference Optimization (SLPO), a reinforcement learning method that rewards concise but correct reasoning while still allowing longer chains when needed. Evaluated on challenging math benchmarks like AIME, Minerva-Math, and OlympiadBench, ORION models trace out a strong performance-efficiency Pareto frontier, preserving accuracy with significantly reduced computation. Paper URL: https://arxiv.org/pdf/2511.22891 #AI #MachineLearning #DeepLearning #LanguageModels #ReasoningModels #MathReasoning #ReinforcementLearning