PAIR Links Vision and Action in VLAs: OpenVLA-OFT Success Rate Rises From 51.4% to 65%

The PAIR authors propose a way to make explicit a step that most VLA models with continuous actions skip: the transition from a representation of the scene and instruction to a representation from which movements can be generated. On the LIBERO-Plus benchmark, VLA-Adapter's success rate rose from 59.1% to 64.2%, and on CALVIN the average completed chain length rose from 4.42 to 4.53. On seven real-world tasks, OpenVLA-OFT gained 13.6 percentage points, going from 51.4% to 65.0%.

The problem the paper addresses will be familiar to anyone following robotics built on large models. A VLA receives an image and a text command and must output continuous actions: a manipulator trajectory, a gripper position. The vision-language backbone is good at describing what is on the table and what the human is asking for. But features that work well for description are not necessarily suited to control. In many architectures this bridge stays implicit, and it is learned only through the final action-prediction error.

How PAIR works

The authors introduce a shared intermediate representation of perception and action. During training, a Masked Action Autoencoder encodes expert action chunks into Action Latent Tokens, aligned across the planning horizon. In parallel, a Bridge Module extracts task-relevant features from the current vision-language representations. PAIR then aligns these features with the action latent tokens. The result is Bridge Tokens: they retain task information while also reflecting the structure of expert movements.

Next, the Bridge Tokens are projected into the action-token space and blended into the initial Action Tokens. The Action Expert, which refines the actions, no longer starts from scratch but from a draft close to the target trajectory. At inference, the autoencoder is dropped: Bridge Tokens are built only from the current observation and instruction. No extra annotation or expert actions are required at run time.

• Base models in the experiments: OpenVLA-OFT and VLA-Adapter.
• Benchmarks: LIBERO, LIBERO-Plus, CALVIN ABC-D.
• LIBERO-Plus, VLA-Adapter: 59.1% → 64.2% (+5.1 pp).
• CALVIN, VLA-Adapter: average chain length 4.42 → 4.53.
• Seven real-world tasks, OpenVLA-OFT: 51.4% → 65.0% (+13.6 pp).

What the experiments showed

The gains in simulation are modest. On CALVIN, the average chain length grows by 0.11 steps, about 2.5% relative to the baseline. On LIBERO-Plus the gain is larger: 5.1 points, roughly 8.6% relative. Real-world tasks show the strongest effect, about 26% relative to the starting 51.4%. But the sample is small: seven tasks and a single model in this test, so it is too early to speak confidently about transfer to other platforms and conditions.

The authors also ran a separate analysis of the representations. According to them, Bridge Tokens retain task information and make information about continuous actions available even before the refinement stage in the Action Expert. This indirectly supports the idea: an intermediate layer between perception and action is useful as an explicit interface, not as a by-product of training. The paper is published in Hugging Face Daily Papers, and there are no independent reproductions yet.

What practitioners should do with this

The approach looks like an add-on rather than a new architecture built from scratch. Testing on two different base VLAs shows that PAIR can be built into existing models with an Action Expert. Since the autoencoder is used only during training, inference cost, judging by the description, barely increases: the only addition is generating Bridge Tokens from the current observation.

Teams that fine-tune OpenVLA-OFT or VLA-Adapter on their own tasks have reason to test the idea on their own data. It makes sense to start with real-world scenarios, where the gap between simulation and hardware is felt most. The main question for follow-up work: will the gain hold on longer horizons, when the manipulator changes, and on tasks outside the training distribution?