InternLM Releases Intern-Decision 4B and 0.8B Multimodal Structured Decision Models

InternLM has released Intern-Decision 4B and 0.8B, multimodal structured decision models fine-tuned from Qwen3.5-4B. These models are designed to accept a shared state, a schema of named questions, and optional images, returning an answer distribution for every question in a single model forward pass.

Inference Mechanism

The models utilize a specific structured candidate scoring API rather than standard free-form text generation:

• Input Processing: The system preserves question and option order, mapping options to single-token symbols (A–Z, a–z, 0–9).
• Prompt Rendering: It renders the original system prompt, state, decision schema, and a complete assistant JSON skeleton with <decision> placeholders.
• Logit Extraction: During a causal Hugging Face forward pass, logits are read at the position immediately before each placeholder.
• Scoring: A softmax is applied only over the allowed candidate-symbol logits for each field, followed by the checkpoint's probability calibration.
• Output: Symbols are mapped back to original option values to return typed JSON answers.

This approach ensures that no gold answers are inserted into the prompt, and the inference compiler relies solely on the provided state, questions, and optional images.