What
Microsoft-Decision-1 is a specialized decision-scoring model designed for discrete workflow tasks like routing, classification, prioritization, and agentic control. Rather than acting as a traditional chat LLM that generates open-ended text, it operates as a specialized control-plane engine that outputs structured probability scores across a fixed set of choices in a single inference step.
How It Works Under the Hood
Underlying Microsoft-Decision-1 is an architecture specifically tuned for high-throughput single-pass classification:
- Base Architecture & Post-Training: It is built by post-training Qwen3.5-9B using a custom optimization process focused entirely on decision scoring rather than auto-regressive text generation.
- Single-Pass Logit Extraction: Standard generative LLMs generate answers token-by-token over multi-second output phases. Microsoft-Decision-1 evaluates the context and a fixed set of candidate choices (e.g., Yes/No, multiple choice, or rating rubrics) simultaneously, extracting the output logits directly from the final classification layer in a single forward pass.
- Calibrated Confidence Outputs: The model does not just output a single selection; it provides a calibrated probability distribution (from $0.0$ to $1.0$) across candidate options. If it outputs $0.90$ confidence, the action is statistically accurate 9 out of 10 times, enabling software logic to trigger fallback flows or human intervention dynamically if confidence drops below a given threshold.
- Perturbation Invariance: The model uses robustness fine-tuning so that harmless noise—such as paraphrasing prompts, changing keys, or reordering options—does not flip decisions. Across eight distinct perturbation types, its decision flip rate is 1.3% overall, and 0% when options are paraphrased or shuffled.
