Robots Atlas>ROBOTS ATLAS
Training

Reasoning-to-Readout Self-Distillation

2026ResearchPublished: 17 November 2026Updated: 17 November 2026Published
Key innovation
A self-distillation method that teaches an immediate first-token decision "readout" to match the distribution the model reaches only after full reasoning — without gold labels and without a stronger teacher model.
Category
Training
Abstraction level
Technika
Operation level
Post-trainingInference
Use cases
Fast decision models returning a decision in a single forward passTraining well-calibrated decisions without gold labelsCompressing post-reasoning decisions into the first tokenFine-tuning decision models with full-weight or LoRA updates

How it works

For a given query the model first unrolls its reasoning and arrives at a probability distribution over the allowed options — this is the teaching signal (target). The same model is then trained so that its first-token readout — restricted to the allowed option labels — reproduces that post-reasoning distribution immediately. Because the target comes from the same model after reasoning, and not from an external teacher or gold labels, this is reasoning-to-readout self-distillation. At inference only the First-Token Decision Readout is used: a single forward pass returns the decision and its probability. The same readout handles all three JEVal formats (choice, judgment, score). Full-weight and LoRA variants are released.

Problem solved

Reasoning-based decision models (chain-of-thought) are slow and costly because they must generate a long token chain before deciding, and classic distillation needs gold labels or a stronger teacher. This technique removes both constraints: it obtains "post-reasoning" decision quality in a single forward pass and trains without gold labels and without an external teacher.

Components

First-Token Decision ReadoutDecision readout at inference

An inference mechanism that restricts the next-token distribution to the allowed option labels and returns the decision with its probability in a single forward pass.

Post-reasoning target distributionTeaching signal during training

The probability distribution over options that the same model reaches after unrolling reasoning; it serves as the teaching target in self-distillation (the model as its own teacher).

Evolution

Original paper · 2026 · arXiv preprint (2026) · Feiyu Duan
General Decision Models: Benchmarking and Insights Beyond Jev
Feiyu Duan, Jiayu Lin, Jia Wang, Jun Xiang, Jialiang Wu, Xinnong Zhang, Hanqi Yan, Siyuan Wang, Zhongyu Wei
2026
Method introduced in the SII-Decision work
Inflection point

The SII team presents Reasoning-to-Readout Self-Distillation and the First-Token Decision Readout, training SII-Decision-4B/27B without gold labels and without a teacher.