Intern-Decision scores structured choices in one pass
In plain words
Many interactive tasks ask an agent to pick from a short list, such as choosing a tool or an action. Intern-Decision is a family of models that makes these choices from text and, optionally, images.
It scores the available answers directly in the requested structured format, instead of writing a full response and having another system interpret it.
Across seven test suites, its 4B model had a higher average decision accuracy than Jev under the reported evaluation protocols. The suites did not all favor the same system, and the comparison does not show what caused the difference.
The research
When an agent needs to choose a tool or action from a fixed set, generating a full text response can add work before the choice reaches the system that uses it. Can a model return structured choices and their probabilities directly, even when a request contains several fields?
Intern-Decision scores predefined answers for multiple fields in one model pass, returning structured choices and probabilities without free-form answer generation.
Intern-Decision scores legal answers at placeholders in a structured response, handling several requested fields in one pass. The authors compare its models with other systems on seven suites, using decision accuracy and an equally weighted mean of suite scores.
The 4B model’s mean is 90.02%, versus 88.74% for Jev. This supports a higher average for the evaluated systems under these protocols, although suite-level exceptions remain.
The accuracy comparison covers the evaluated systems and test bundles; it does not isolate the effect of joint field scoring or model size. The public Jevbench items informed development, so they are diagnostic rather than independent evidence of generalization.
Analysis
Deep
- Original paper authors
- Tong Zhu, Zhe Xu, Zefeng He, Yunkai Chai, Xuyang Hu
- Submitted to arXiv
- License
- CC BY-SA 4.0