Open Models Almanac
Methodsmodel architecturemultimodal architecture

specific method · filed under model architecture

Per-layer key-value conditioning

Per-layer key-value conditioning is an interface that lets a separately trained action expert receive context from a vision-language model at every layer rather than from a single compressed representation.1 Each action-expert block cross-attends to the corresponding VLM layer's keys and values, giving the expert access to the same attention state the VLM uses while keeping the backbone modular and unchanged.2

How this definition was written

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

  1. Instead, MolmoAct2 conditions the expert on the VLM at every layer — MolmoAct2 Action Reasoning Models for Real-World Deployment
  2. Each action-expert block cross-attends to the corresponding VLM layer’s keys and values, after lightweight learned projections map the VLM attention state into the expert’s cross-attention width — MolmoAct2 Action Reasoning Models for Real-World Deployment

Also called per-layer KV conditioning, per-layer KV (key-value) conditioning, standard per-layer KV conditioning.

sources
2
model
1
lab adopt it
1
strongest
core

Documented in

Allen Institute for AI

Evidence

5 quotes from 2 sources, one card per model, each opening with its strongest.

How sources treat it
core 5

MolmoAct 2

Allen Institute for AIadopted2 sources

connects the autoregressive VLM to a flow-matching continuous action expert through per-layer KV (key-value) conditioning.

coreunclearin MolmoAct2Allen Institute for Artificial Intelligence (AI2)
4 more from this model

Each action-expert block cross-attends to the corresponding VLM layer’s keys and values

coremodel architecturein MolmoAct2

we post-train a continuous flow-matching expert using per-layer KV conditioning.

coreunclearin MolmoAct2Allen Institute for Artificial Intelligence (AI2)

The action expert receives VLM context through per-layer KV conditioning.

coreunclearin MolmoAct2 action expertAllen Institute for Artificial Intelligence (AI2)

The standard per-layer KV conditioning is strongest on average

coreunclearin MolmoAct2Allen Institute for Artificial Intelligence (AI2)
↑