The authors compared a vanilla model, a chain-of-thought model, a pause-token model, and two latent-reasoning models on ProsQA-Ext, an extended synthetic task requiring multi-hop reasoning over graph relations.
Latent recurrence learns graph search beyond its training depth
On a controlled multi-hop task, latent-reasoning models discovered a recurrent reachability algorithm that generalized to longer paths better than token-based alternatives.
Academic
Huzi Cheng · Zhewei Zhang
University of Minnesota · Independent Researcher
Research Digest··2 min read
Cheng and Zhang trained five versions of the same GPTNeoX backbone from scratch, varying only how each model performed intermediate computation.
Why this paper
From University of Minnesota and Independent Researcher
In one line
Latent reasoning discovers a recurrent search algorithm that generalizes beyond training depth; vanilla, CoT, and pause-token models do not.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·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.
§