Open Models Almanac
Methodstraining objectiveauxiliary loss

specific method · filed under training objective

Knowledge insulation

Detaches VLM-provided conditioning from the flow loss so its gradients update the action expert rather than back-propagating through the VLM.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See auxiliary loss for the methods filed beside it.

Also called stop-gradient to the KV-cache.

sources
3
models
3
labs adopt it
3
strongest
used

Documented in

Allen Institute for AI
Physical Intelligence

Evidence

5 quotes from 3 sources, one card per model, each opening with its strongest.

How sources treat it
evaluated 1used 4

MolmoAct 2

Allen Institute for AIadopted1 source

The expert conditions on the VLM keys and values, but these tensors are detached before they enter the expert, so updates the action expert and its adapter projections without back-propagating through the VLM.

usedoptimizationin MolmoAct2
2 more from this model

Adding knowledge insulation is also close but slightly worse than the final recipe.

evaluatedunclearin MolmoAct2Allen Institute for Artificial Intelligence (AI2)

In post-training, this conditioning path is detached from the VLM for the flow loss.

usedunclearin MolmoAct2Allen Institute for Artificial Intelligence (AI2)

π0.5

Physical Intelligenceadopted1 source

an upgraded version of π₀ with better open-world generalization trained with knowledge insulation

usedtraining objectivein π₀.₅Physical Intelligence
↑