method · first seen 23 Oct 2026
reasoning token reduction
Also: thinking budget reduction
Techniques that shorten a reasoning model's intermediate thinking tokens while keeping answer accuracy roughly unchanged.
Mentions per week
PapersCodeDiscussion
Earliest sources
- Gemma-4-26B-A4B with 39-69% less thinking at same accuracyDiscussion · 23 Oct 2026 · armcat · 127 points
Latest sources
- Gemma-4-26B-A4B with 39-69% less thinking at same accuracyDiscussion · 23 Oct 2026 · armcat · 127 points