InfoDensity: AUC-Based Intermediate Reasoning Rewards Reduce Reasoning Token Verbosity Without Accuracy Loss

arXiv·medium signal

ArXiv paper 2603.17310 introduces InfoDensity, a training reward based on AUC of information gain across reasoning steps rather than final-answer correctness alone. This directly targets the 'reasoning theater' problem where extended chain-of-thought adds tokens without proportional accuracy gains. The approach is compatible with existing RLHF pipelines and shows consistent verbosity reduction while maintaining benchmark scores.