Robots Atlas>ROBOTS ATLAS
SII-Decision-4B

SII-Decision-4B

4B · Family: SII-Decision
Open-weight 4B decision model (SII-Decision family) producing structured JEVal decisions in a single forward pass via First-Token Decision Readout.
🔬 Research🔬 Research only⚖ Open sourceLLMSpecialized AI📁 SII-Decision
Parameters
4B
parameters
Access:DownloadDeployment:💻 Local

Overview

SII-Decision-4B is an open-weight 4-billion-parameter decision model from the SII-Decision family. It is a fine-tune of Qwen3.5-4B, built for fast, structured decisions in the JEVal format: single-option choice, binary judgment and score assignment.

The model uses the First-Token Decision Readout method, which constrains the next-token probability distribution of an open-weight LLM to the allowed option labels only and returns a decision in a single forward pass, without unrolling full reasoning.

Training relies on Reasoning-to-Readout Self-Distillation: the readout learns to match the distribution the model arrives at after reasoning, without gold labels and without a stronger teacher model. Full-weight and LoRA variants are available.

On the JEVal benchmark, the 4B variant reaches 71.92% accuracy, 0.0843 ECE and 0.383 raw Brier score. The model was developed by the Shanghai Innovation Institute (SII) together with Fudan University, Tongji University, King's College London and the Chinese University of Hong Kong. Weights and code are released under the Apache-2.0 license.

Classification
LLMSpecialized AI
Family: SII-Decision
Access & deployment
Download
Local
Weights: Open source
Key parameters
🧩 Parameters: 4B
📥 Input: text

Technical specification

Parameters
4B
parameters
License
Apache-2.0
Modalities
⬇ Input
text
⬆ Output
structured_datatext

Capabilities and applications

Native model capabilities
Decision making
The ability to choose the best option from a defined set of possibilities based on input state, usually with a probability estimate for each option.
Category: reasoning
Classification
Assigning an observation to one of predefined classes (binary or multi-class). Output: class label and optionally probabilities.
Category: other
Structured output
Producing data in structured formats such as JSON.
Category: structured_generation
Instruction following
Precisely following instructions contained in the prompt: response format, length, style, constraints (e.g. 'reply in six words'). GPT-5.1 significantly improved this capability compared to GPT-5.
Category: language
Reasoning
The model's ability to reason logically and solve complex problems.
Category: reasoning

Benchmark results

3 benchmarks
JEVal
accuracy · First-Token Decision Readout; SII-Decision-4B variant
71.92%
📄 paper
JEVal
ECE · Expected Calibration Error; 4B variant
0.0843
📄 paper
JEVal
Brier score (raw) · Raw Brier score; 4B variant
0.383
📄 paper

Technical architecture

Core Architecture
Model Form