Intern-Decision paper details 0.8B–4B multimodal decision models
A new arXiv paper by Tong Zhu and co-authors presents Intern-Decision, a family of 0.8B, 2B and 4B multimodal decision models that return structured choices and candidate probabilities from text and optional images. The method places one decision placeholder per field in a complete assistant JSON skeleton and evaluates all fields in a single causal forward pass, with ground-truth symbols used only in supervision. The authors report that the 4B model reaches 90.02% unweighted average accuracy across seven evaluation suites versus 88.74% for Jev v1.13.0, and a mean latency of 44.16 ms for a fixed 289-token, three-field request on one RTX 4090; they also describe a confidence-based handoff study in which routing 42 of 111 Jevbench-Hard items to Atria-Dawn-Preview raised accuracy from 82/111 to 97/111.
The authors say these results support compact decision models as a practical component of larger agent systems, arguing that generating a free-form response for each choice adds latency and complicates the model-environment interface.
Computer Science > Computation and Language
arXiv:2610.22131 (cs)
[Submitted on 15 Oct 2026]
Title:Intern-Decision: Efficient Multimodal Decision Making with Autoregressive Masked Language Modeling
Authors:Tong Zhu, Zhe Xu, Zefeng He, Yunkai Chai, Xuyang Hu, Dehong Zheng, Yun Luo, Yafu Li, Qi Zhang, Yu Cheng
View a PDF of the paper titled Intern-Decision: Efficient Multimodal Decision Making with Autoregressive Masked Language Modeling, by Tong Zhu and 9 other authors