Intern-Decision: Efficient Multimodal Decision Making with Autoregressive Masked Language Modeling

Intern-Decision. Efficient Multimodal Decision Making with Autoregressive Masked Language Modeling. By Tong Zhu and nine others, from the Shanghai AI Lab, Shanghai Jiao Tong University, and two other institutions. Published on the fifteenth of October. Its code repository is worked out: about fifty source files, with a readme and a licence.

Interactive agents keep making choices from a finite menu: an action, a label, an ordinal outcome. The usual way to get such a choice out of a language model is to let it generate a free-form response and read the answer from that. Doing so for every choice adds latency, and it complicates the interface between the model and its environment, which ends up more elaborate than the decision it carries. The output space is small. The machinery used to produce it is not.

Intern-Decision is a family of multimodal decision models at three sizes: zero point eight billion, two billion, and four billion parameters. They take text, and optionally images, and return structured choices together with probabilities for the candidates, instead of prose. The training objective is called autoregressive masked language modeling. A complete assistant JSON skeleton is written out, and each field in it holds exactly one decision placeholder, the slot where that field's answer belongs.

Ground-truth symbols appear only in supervision. When the model is evaluated, every field is filled in a single causal forward pass, and the scoring is done by the backbone's own native language-model head rather than by an added component. A request with several fields therefore does not need several rounds of generation. One pass reads the whole skeleton and returns, for each placeholder, a choice and the probabilities of the candidates behind it.

Across seven evaluation suites, the four billion parameter model reaches an unweighted average accuracy of ninety point zero two percent, against eighty-eight point seven four percent for Jev v1.13.0. Speed was measured on a fixed request of two hundred eighty-nine tokens with three fields, on one RTX 4090. Mean latency is forty-four point one six milliseconds for the largest model, and roughly thirty-three to thirty-four milliseconds for the smaller ones.

The probabilities were examined too. On a separate pilot of ninety-six cases drawn from a known distribution, temperature scaling lowers expected multiclass Brier from zero point six two eight to zero point five five, and expected calibration error from zero point two one three to zero point zero eight nine. The authors also describe qualitative demonstrations in games, visual interfaces, and a browser.

The authors conclude that compact decision models are a practical component of larger agent systems, and a confidence-based handoff study shows what that can look like. On one hundred and eleven public Jevbench-Hard items, forty-two were routed to Atria-Dawn-Preview, and correct answers rose from eighty-two to ninety-seven: from seventy-three point eight seven percent to eighty-seven point three nine percent. A small model that reports how sure it is can answer quickly when confident and pass the rest along.