Interactive agents often need to choose a tool, classify a screen or set a score, not write an essay. A Shanghai AI Lab-led team says its new Intern-Decision models can score several such fields together from text and optional images in one forward pass, returning structured choices with candidate probabilities. Its 4B model averaged 90.02% across seven test suites, the authors report, with important limits on what that comparison shows.
A Model Interface for Fixed Choices
Intern-Decision comes in 0.8B, 2B and 4B sizes. The researchers fine-tuned Qwen3.5 language backbones while keeping the vision tower and projector frozen, so an agent can provide text, optional images and multiple typed questions in one request.
The model receives a fixed JSON template with one blank for each decision. In one model run, it scores the allowed answers for each blank; a formatting step puts the highest-scoring option into the requested field. This lets an agent use the choices directly instead of asking the model to write free-form prose first. The structured format is predictable, though valid JSON does not prove an answer is correct.
For an agent that must make several small decisions from the same screen or state, scoring the fields together could simplify the boundary between a model and its controller. The team’s release links code and model checkpoints, but does not include its training data.
The Headline Score Has a Narrow Meaning
The paper reports a 90.02% unweighted average accuracy for Intern-Decision-4B across seven evaluation suites, compared with 88.74% for Jev v1.13.0, a structured decision API used as a baseline in the evaluation. The bundles contain 10,751 rows and 12,351 decisions, but the headline is an average of seven suite-level scores: each suite counts equally, regardless of size. It is not pooled accuracy over all decisions.
The comparison also needs context. The 4B model scored 82 of 111 items, or 73.87%, on the public Jevbench-Hard set. The authors say public Jevbench informed development, so the benchmark is a diagnostic rather than independent evidence of generalization. The paper does not isolate how much of the result comes from the scoring method or the model size.
Latency Applies to a Short Request
On one RTX 4090, the 4B model averaged 44.16 milliseconds on a fixed 289-token request with three fields, using BF16 and the Hugging Face backend. The 0.8B and 2B models averaged about 34 milliseconds. The measurement covers one compact request, not the full loop of reading a changing image, choosing an action and waiting for an environment to respond.
The paper’s Jev timing comparison lacks full serving and transport conditions, and some timing details such as warm-up and repetition count are not specified. The reported numbers therefore do not establish a controlled hardware speedup over the baseline.
Low-Confidence Cases Go to a Larger Model
In a separate experiment on 111 Jevbench-Hard items, the team routed 42 low-confidence cases to the larger Atria-Dawn-Preview model and kept 69 with Intern-Decision-4B. The paper describes comparing a scalar confidence with a cutoff, but does not specify the confidence measure or the cutoff used for these results. The hybrid got 97 items right, compared with 82 for the 4B model alone. Its reported 140.16-second total is an estimate; the 300.03 seconds for sending every item to Atria is measured.
The result sketches one way a compact decision model could handle routine choices while escalating uncertain ones. It shows one routing setup and does not establish which confidence threshold works best.
Illustrative image: A Complex Graphics Card.jpg by Nick Stathas (EvolvedWiki), Wikimedia Commons, CC BY-SA 4.0, unmodified. The pictured AMD Radeon R9 270X is not the RTX 4090 used in the paper’s timing test.
