Latent Reasoning with Supervised Thinking States
Ido Amos, Avi Caciularu, Mor Geva, Amir Globerson, Jonathan Herzig, Lior Shani, Idan Szpektor · 2026
Read Latent Reasoning with Supervised Thinking States by Ido Amos, Avi Caciularu, Mor Geva, Amir Globerson, Jonathan Herzig, Lior Shani, Idan Szpektor (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-08332.txt.
Description
Chain-of-thought reasoning helps language models solve hard problems, but writing out long rationales costs time at inference. Ido Amos and six co-authors propose Thinking States, a method that reasons while the input is still being read, instead of waiting until the end. Their 2026 arXiv paper interleaves short bursts of thought tokens with the incoming text and feeds those thoughts back as embeddings.
Because the thoughts stay in token form, they can be taught from ordinary language with a parallelizable training trick. In tests the method beats other latent-reasoning approaches, nearly matches full chain-of-thought on math, equals it on two-hop questions with lower latency, and even beats it on state tracking and longer sequences. Readers get a leaner path to reasoning that does not force a choice between speed and depth.
Catalog record
| CREATOR | Ido Amos, Avi Caciularu, Mor Geva, Amir Globerson, Jonathan Herzig, Lior Shani, Idan Szpektor |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 49 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-08332