Papers

Multimodal Reasoning Classifiers Approach the Frontier at Lower Cost and Suggest Two Axes of Intelligence in Language Models

A reasoning model thinks over a text or an image, then scores every option in one pass. Reasoning raises Gemma 4 by 0.14 to 0.16 where the answer must be inferred, and setting what reasoning adds against what scale adds separates two axes of capability.

Any Language Model Is a Classifier: Calibrated One-Pass Decisions over Arbitrary Label Sets from Pretrained LLMs

Scoring every option in one forward pass, then training that distribution on 18 tasks, turns a pretrained model into a calibrated classifier for label sets it has never seen. A 77-way decision from a 1.7B model takes 54 ms.