The authors studied 21 model-task configurations covering three training recipes (including COCONUT and per-step supervision), three reasoning tasks (logical, arithmetic, and others), and two backbone architectures.
Latent chain-of-thought often lacks genuine reasoning in logical tasks
The authors test 21 model-task configurations and find that on logical tasks, latent chains frequently encode pre-existing answers and are causally unnecessary.
Independent
Jingjie Qiu · Sheldon Yu · Xunyi Jiang · Jiajin Liu · Zihan Huang · Zhendong Chu · +3 more
Research Digest··2 min read
Thread:Reasoning Trace Correction
The authors investigate whether latent chain-of-thought models, which use continuous hidden states instead of explicit reasoning steps, actually perform reasoning.
Why this paper
Independent
In one line
Latent chain-of-thought often does not reason: readable answers can precede the chain, and deleting it leaves predictions nearly unchanged.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ·No stated limitations found
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§