Open Models Almanac
Methodspost-trainingsupervised fine-tuning

single mention · post-training

Embodied specialization stage

Embodied specialization is the first stage of a specialize-then-rehearse recipe: it fine-tunes an existing multimodal checkpoint on a new embodied corpus.1,2 Text-only data is mixed in to preserve language competence, and the stage rapidly improves pointing, video embodied QA, and multi-image reasoning.3,4

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. Starting from the Molmo2–4B mid-training checkpoint — MolmoAct2 Action Reasoning Models for Real-World Deployment
  2. we fine-tune for K steps on the Molmo2-ER corpus — MolmoAct2 Action Reasoning Models for Real-World Deployment
  3. augmented with Tulu-3 (Lambert et al., 2025) text-only data to preserve language competence — MolmoAct2 Action Reasoning Models for Real-World Deployment
  4. pointing accuracy, video embodied QA, and multi-image reasoning all improve sharply — MolmoAct2 Action Reasoning Models for Real-World Deployment

Also called Embodied specialization.

Filed under post-training › supervised fine-tuning