The authors propose a two‑stage method.
Training only the root token enables efficient adaptive reasoning
Root-token Policy Optimization (RPO) trains only the initial gating token of large reasoning models, cutting adaptive‑thinking training cost to about 2% of prior methods.
Academic
Taehyeon Kim · Hyunsoo Lee · Youngsoo Jang · Moontae Lee
LG AI Research · UNIST · University of Illinois Chicago
Research Digest··3 min read
The authors introduce Root‑token Policy Optimization (RPO), a framework that allows large reasoning models (LRMs) to decide adaptively whether to engage in explicit reasoning by optimizing only the first token (the root token) that gates the reasoning process.
Why this paper
From UNIST and 2 others
In one line
Root-token policy optimization teaches reasoning models when to think while using about 2% of the training compute required by prior adaptive-reasoning methods.
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.
§