Think Wider: Mitigating Latent Rank Collapse in Implicit Chain-of-Thought Reasoning
Jan 1, 2026·
,·
1 min read
Yuwen Hao
Menglin Yang
Image credit: UnsplashAbstract
This work identifies latent rank collapse in implicit chain-of-thought reasoning, where multi-step latent tokens over-concentrate along shared dominant directions, reducing effective rank and increasing inter-step redundancy. It proposes WIDER, a lightweight spectral regularization strategy that penalizes projection onto dominant latent directions during training.
Type
Publication
Under Review
Status
Open access
Note
This paper is currently under review.
Think Wider studies how implicit reasoning trajectories can collapse into low-rank latent representations. The proposed WIDER objective estimates dominant singular directions and discourages latent tokens from concentrating along them, improving effective rank and reducing redundant steps without adding inference-time overhead.

Authors
MPhil Student in Artificial Intelligence
I am an MPhil student in Artificial Intelligence at the Hong Kong
University of Science and Technology (Guangzhou), working with Assistant
Professor Menglin Yang in DIGAI Lab.
My current research focuses on continuous
latent-space reasoning, implicit chain-of-thought reasoning, and language
model reasoning systems.