Self-Supervised Confidence Training Cuts AI Reasoning Tokens Without Sacrificing Accuracy, Capital One Team Finds
New approach teaches reasoning models to estimate their own confidence. It reduces token use without hurting accuracy across math, science, and coding tasks, using only internal signals.
- Publication
- arXiv
- Stage
- Preprint
- What we read
- Summary of the paper
- Authors
- Parsa Hosseini, Akasha Tigalappanavara, Sumit Nawathe, Chenrui Fan, Sourya Basu, Genta Indra Winata, Anirban Das, Soheil Feizi, Nima Chitsazan
- Universities and research institutions
- Not yet supplied in verified metadata; the Brief does not guess.
What they did and found
The study trained models to predict confidence at intermediate steps using only the model’s own probabilities. Inference used standard generation with no early stopping. Results show up to 25% fewer tokens at the same accuracy across several model families and tasks.
Why it matters
Could lower compute costs and energy use in real applications while keeping answers reliable in schools, public services, and businesses.
ConfSFT fine-tunes reasoning models to predict confidence at intermediate states using only token probabilities, without aiming to shorten reasoning. Accuracy stays stable while generated tokens drop.
Practically, this hints at lower compute needs for frequent reasoning tasks, with energy costs and latency potentially reduced in routine uses like math help or coding aids.
Original sources · 1
- Learning to Stop without Learning to Stop: Self-Supervised Confidence Training Improves Reasoning Efficiency ↗arXiv · 2026-09-25
Check the original paper for its authors, methods, version and access terms.
What remains uncertain
Results come from a subset of models and benchmarks; broader domain validity and real-world deployment effects remain untested.
What is your take?
Ask a question, add useful context or share a different perspective. Keep the conversation respectful and grounded.