Reasoning on the Manifold: Bidirectional Consistency for Self-Verification in Diffusion Language Models

Jiaoyang Ruan, Xin Gao, Yinda Chen, Hengyu Zeng, Liang Du, Guanghao Li, Jie Fu, Jian Pu · 2026

Read Reasoning on the Manifold: Bidirectional Consistency for Self-Verification in Diffusion Language Models by Jiaoyang Ruan, Xin Gao, Yinda Chen, Hengyu Zeng, Liang Du, Guanghao Li, Jie Fu, Jian Pu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-16565.txt.

Description

Ruan and colleagues propose thinking on the manifold for self-checks in diffusion language models. Valid output paths are stable anchors on the high-density manifold, while invalid paths drift off-manifold. Their two-way manifold match metric, BMC, gauges sequence stability through forward hiding and backward rebuild, without training or labels. BMC checks solution truth without ground truth, focuses compute on complex tasks by refusal retries, and guides tuning as a dense shape reward. Inner shape stability thus serves as a robust sign of rightness for spread models.

Catalog record

CREATORJiaoyang Ruan, Xin Gao, Yinda Chen, Hengyu Zeng, Liang Du, Guanghao Li, Jie Fu, Jian Pu
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE109010
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.16565

Browse the full library · About the Tower of Babel