Modern causal language models, followed by rapid developments in discrete diffusion models, can now produce a wide variety of interesting and useful content. However, these families of models are predominantly trained to output tokens with a fixed (left-to-right) or random order, which may deviate from the logical order in which tokens are generated originally. In this paper, we observe that current causal and diffusion models encounter difficulties in problems that require adaptive token generation orders to solve tractably, which we characterize with the $\mathcal{V}$-information framework. Motivated by this, we propose Reinforced Context Order Recovery (ReCOR), a reinforcement-learning-based framework to extract adaptive, data-dependent token generation orders from text data without annotations. Self-supervised by token prediction statistics, ReCOR estimates the hardness of predicting every unfilled token and adaptively selects the next token during both training and inference. Experiments on challenging reasoning and planning datasets demonstrate the superior performance of ReCOR compared with baselines, sometimes outperforming oracle models supervised with the ground-truth order.

This year, Co.Design asked a handful of design firms to take on the moral dilemma of self-driving car decision-making: What does a car do when it has to choose between saving its passenger and saving a pedestrian? Their solutions included smart roads, flying airbags, and air traffic control-style systems. But the San Francisco-based design firm Box Clever is focusing on safety first. The studio imagines creating a new layer of public infrastructure in the form of security drones that can warn self-driving cars of things they can't see. Combined with the use of smart materials in cars themselves that can better absorb the impact of crashes, the studio's vision doesn't just apply to a world of autonomy–it could help make our streets safer for pedestrians right now.