aiqbench analysis
This research introduces DART, a novel training framework designed to enhance the stability and robustness of recurrent reasoning models by replacing single-point supervision with local target distributions. By aligning model outputs through an adversarial objective, the method effectively mitigates issues where valid solutions occupy a tiny region of the solution space as task difficulty increases. The approach demonstrates improved performance on complex structured problems like chess and maze solving without relying solely on standard softening techniques.
DART: Distributional Adversarial Recurrent Training for Algorithm Learning
arXiv:2609.05988v1 Announce Type: new Abstract: Recurrent reasoning models (RRMs) can solve structured problems, achieving easy-to-hard generalization through iterative computation in hidden space. These models are typically trained with instance-level supervision, which becomes increasingly problematic as task difficulty grows: valid solutions occupy a tiny region of the solution space, while invalid solutions proliferate rapidly. We propose Distributional Adversarial Recurrent Training (DART), a training framework that replaces single-point supervision with a local target distribution around the ground-truth solution and aligns model outputs with this distribution through an adversarial objective. DART provides a richer learning signal and encourages more stable iterative trajectories toward valid solutions. When evaluated on Maze, Chess, and masked Sudoku with multiple RRMs, including Deep Thinking Systems and Tiny Recursive Models, DART improves solution quality, stability, and robustness under the evaluated distribution shifts. Comparisons with label smoothing, Gaussian softened targets, and progressive training show that DART is not explained by target softening alone and is complementary to training schemes that stabilize long-horizon recurrence. These results identify DART as a promising approach for improving robustness across the evaluated recurrent reasoning models.
Abstract:Recurrent reasoning models (RRMs) can solve structured problems, achieving easy-to-hard generalization through iterative computation in hidden space. These models are typically trained with instance-level supervision, which becomes increasingly problematic as task difficulty grows: valid solutions occupy a tiny region of the solution space, while invalid solutions proliferate rapidly. We propose Distributional Adversarial Recurrent Training (DART), a training framework that replaces single-point supervision with a local target distribution around the ground-truth solution and aligns model outputs with this distribution through an adversarial objective. DART provides a richer learning signal and encourages more stable iterative trajectories toward valid solutions. When evaluated on Maze, Chess, and masked Sudoku with multiple RRMs, including Deep Thinking Systems and Tiny Recursive Models, DART improves solution quality, stability, and robustness under the evaluated distribution shifts. Comparisons with label smoothing, Gaussian softened targets, and progressive training show that DART is not explained by target softening alone and is complementary to training schemes that stabilize long-horizon recurrence. These results identify DART as a promising approach for improving robustness across the evaluated recurrent reasoning models.
Machine Learning (cs.LG); Artificial Intelligence (cs.AI)
arXiv:2609.05988 [cs.LG]
(or arXiv:2609.05988v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2609.05988 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 272 words
Related coverage
Latent reasoning architectures would likely undermine CoT, our strongest oversight tool
AI Alignment Forum · Sep 23, 2026
A Mike's-Eye View of ARC's Research
AI Alignment Forum · Jun 9, 2026
On the Escaping Efficiency of Distributed Adversarial Training Algorithms
arXiv · Oct 2, 2026
The Alignment Problem: Computer Science Meets Analytic Philosophy
AI Alignment Forum · Oct 18, 2026