Reasoning accuracy in latent CoT depends on mutual information fidelity between latent trajectories and explicit steps, with generative reconstruction preserving capacity better than geometric compression.
Language models can learn implicit multi-hop reasoning, but only if they have lots of training data
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2026 1verdicts
UNVERDICTED 1representative citing papers
citing papers explorer
-
What Makes Effective Supervision in Latent Chain-of-Thought: An Information-Theoretic Analysis
Reasoning accuracy in latent CoT depends on mutual information fidelity between latent trajectories and explicit steps, with generative reconstruction preserving capacity better than geometric compression.