Think Wider: Mitigating Latent Rank Collapse in Implicit Chain-of-Thought Reasoning
A spectral regularization method for improving latent-token diversity and mitigating rank collapse in implicit reasoning.
A spectral regularization method for improving latent-token diversity and mitigating rank collapse in implicit reasoning.
Website-prior task and trajectory co-synthesis for web agents.
A lightweight geometric regularization method for stabilizing implicit reasoning trajectories without adding inference-time overhead.
A study of premature upper-layer attention specialization during language model pretraining.