Build leaner robot policies by tracing the layers that drive action

A practical causal-intervention workflow for locating where vision becomes action, redesigning the policy around that interface, and validating the leaner model in simulation and on physical hardware.

August 2026

In short

A vision-language-action policy combines a vision-language model (VLM) that interprets the scene with an action expert that turns that representation into motion. Most architectures let the action expert read every VLM layer. Causal interventions showed that MolmoBot's action-relevant handoff is concentrated at layer 24.

Restricting cross-attention to that layer removed 97% of the action expert's cross-attention parameters and 93.5% of its static operations. The redesigned policy reached 60.8% held-out simulation success, compared with 59.5% for the released all-layer baseline. Applying the same localization procedure to π0.5 and MolmoAct2 revealed different causal bands: the workflow transfers, but the layer number does not.

These are static operation counts, not measured latency. The simulation comparison uses separately trained policies rather than a seed-matched control, and the physical evaluation is a 50-trial pilot. The evidence supports a leaner interface and a reusable localization method — not a universal layer number or a production-speed claim.

Step one

Locate the handoff from vision to action

Start by changing one internal representation at a time while holding the image, robot state, and sampling noise fixed. If the predicted action changes, that representation participates causally in the handoff from perception to control.

The handoff can be tested directly, on a policy nobody retrains. Interchange patching swaps one internal VLM representation for the one the same model produced under a different instruction — run the model on “reach for the tomato”, keep one layer's internal notes, and paste them into the run for “reach for the plate”. Ablation deletes a candidate read instead of replacing it, and steering adds a learned direction to it. Each is run with the image, the robot's joint state and the flow noise held fixed: these policies draw a fresh random sample on every rollout, so pinning that draw means the swap is the only thing that differs. If the predicted reach moves, that internal read is causally involved in the action.

The practical result is not "always use layer 24." It is "locate the causal interface before pruning it."

Depth axis

Figure 1 · Causal depth profiles across three VLA designs. Each model's intervention effect is drawn as a fraction of its own peak, because the recorded metrics differ: this is a qualitative localization view, not an effect-size comparison across models. Switch the horizontal axis between fraction of model depth and absolute layer; hover a point for the recorded effect. π0.5's spike at layer 0 reflects prompt-template structure rather than scene content, and its content-dependent band sits at layers 13–16.

MolmoBot's read peaks at layer 24

In a two-object reach, swapping only layer 24's context recovered 78.6% of the tomato-versus-plate reach gap. The peak stayed at layer 24 across three fresh flow-noise seeds, whose recovered fractions were 0.7711, 0.7658, and 0.7924.

The effect is sharply localized even though instruction information keeps accumulating deeper in the network. That separation is the lesson of this step: held-out decoding of the spatial reach target — training a small readout to recover the reach coordinates from a layer, on scenes it was not fitted on — peaks at r = 0.9286 at layer 22, two layers before the causal effect on the action peaks at 24. A representation can be easy to decode without being the one the action expert uses, which is why a decoding sweep alone would have pointed at the wrong layer.

Figure 2 · The target is computed before it is read. Held-out decoding of the spatial reach target (slate) rises through the middle layers and peaks at r = 0.9286 at layer 22. The causal effect on the action (ember) peaks two layers later, at 24. Both series are drawn as a fraction of their own peak so the ordering is visible on one axis; the shaded band marks the layers over which the target becomes decodable.

Figure 3 · Recorded MolmoBot trajectories under intervention. The baseline paths separate toward the instructed tomato and plate. Interchanging the causal band, or steering layer 24, redirects the plate-instructed reach toward the tomato. These are recorded action trajectories, not schematic arrows.

Step two

Turn the causal map into a smaller interface

Next, treat the causal map as an architecture proposal. Restrict the action expert to the identified interface, train the redesigned policy, and test it on held-out scenes — not just parameter count or training loss.

The causal map suggested a concrete redesign: keep the action expert's cross-attention — the wiring through which it reads the VLM — only at layer 24. The first question was not how many parameters disappeared. It was whether the policy still worked on scenes it had never seen.

Answering it meant training policies, not re-analyzing a released one. Each candidate layer was built two ways. A retrofit takes an existing trained policy and continues training it for 50K steps into the new interface. Full two-stage pretraining instead trains the policy through both of its usual stages from the start, which is the far more expensive route. Layers 24, 12 and 1 were each scored on roughly 1,000 held-out rollouts in the MolmoSpaces simulator — complete simulated attempts at the task, on scenes withheld from training, counted as successes and failures rather than as a loss curve.

Interface

All-layer action expert

Every action-expert block projects and attends to a long VLM context.

Cross-attention parameters

Action-expert operations

VLM + expert, with truncation

Full policy, with truncation

The VLM-plus-expert and full-policy reductions additionally assume truncating the VLM after layer 24. All four are static operation counts, not measured latency.

Figure 4 · What changes when MolmoBot reads only layer 24. The interface restriction itself drops 97% of cross-attention parameters and 93.5% of the action expert's static operations. The 40.2% VLM-plus-expert and 34.0% full-policy reductions additionally assume truncating the VLM after layer 24. None of them is a latency measurement.

Figure 5 · Held-out simulation success separates layer choice from training route. Points are success rates with 95% binomial intervals over roughly 1,000 MolmoSpaces rollouts per model; the row label carries the successes and rollouts behind each point. Filled markers are fully pretrained models, hollow markers 50K-step retrofits of an existing policy.

Layer 24 matches the released baseline in held-out simulation

The all-layer baseline reached 59.5% success (609/1023). Layer-24-only reached 60.8% when pretrained (620/1020) and 59.9% after a 50K-step retrofit (592/988). At this depth the two routes land in the same place, so an existing policy can be adapted without repeating the full pretraining stage.

Layer 12 exposes a task-success–training-cost tradeoff

Deeper in, the routes come apart, and that is what makes running both worthwhile. A 50K retrofit at layer 12 reached 47.5%, but full two-stage pretraining raised it to 57.0% (568/996; 95% CI 53.9–60.1), a 9.5-point gain. At layer 12 the training route, not only the layer choice, decided the outcome: the shallow cut needed the more expensive route to approach layer 24, while permitting more VLM truncation. Layer 1's 21.7% retrofit anchors the failure end of the frontier.

Step three

Repeat the localization sweep on each architecture

The procedure transfers; the layer number does not. Repeat the same localization sweep for each architecture to find where its action expert reads and which information it uses.

The narrow interface recurs, but its location and its carrier do not. π0.5 couples its action expert through a shared prefix key/value cache — the expert reads a cached summary of the prompt rather than attending into the VLM directly — and its main content-dependent band sits at layers 13–16, peaking at layer 14, while a separate layer-0 effect reflects template structure rather than the scene. MolmoAct2 uses dedicated cross-attention, and its state-span read concentrates at blocks 17–20, peaking at block 20.

The content differs too: not just where each model reads, but what it is reading. In the MolmoAct2 analysis, masking the discrete state span — the tokens encoding the arm's own configuration — shifted the action by 0.5406, while masking instruction positions shifted it by only 0.0049. π0.5's band was driven mostly by image and state. A reusable compression recipe therefore needs a localization sweep per architecture, not a universal layer number — the depth profiles in step one are plotted as a fraction of each model's own peak for exactly that reason, and are a qualitative localization view rather than an effect-size comparison across models.

Step four

Validate the redesign on physical hardware

Finally, move the redesigned interface out of simulation. Compare the unrestricted and pruned policies on physical hardware, then report task success and timing separately.

The transfer test ran on a Franka FR3 arm at 15 Hz on a fruit-in-basket task. The Vanilla π0.5 policy was unrestricted; the Pruned policy let its action expert read only VLM layers 14–16. The session ran 50 trials, 25 per arm, in four alternating policy blocks.

Both checkpoints were trained for 100K steps on the filtered DROID p98 subset. Success followed the operator label, except that more than 2000 control steps forced a failure; that rule applied to two over-cap episodes, and episode_0028 was scored a success per the operator after a console misclick.

Both arms succeeded on 22 of 25 trials (88%; Wilson 95% CI 70.0–95.8%; Fisher exact p = 1.000). Those are the observed counts. The pilot had no pre-registered non-inferiority margin — a threshold fixed in advance for how much worse the pruned policy could be and still count as not worse — so it does not establish equivalence or non-inferiority. Without one, a tie is only a tie.

Figure 6 · Every Franka trial, by duration. One dot per trial on its arm's row, with the median drawn as a bar: 53.6 s for Vanilla against 45.4 s for Pruned. The 6 failures are marked with crosses at the time the trial stopped; those times are censored, not completion times, and they are excluded from the medians. Row labels carry each arm's successes and Wilson interval.

Figure 7 · All 50 trials in session order. The strip preserves the four chronological policy blocks and every outcome; hover a trial for its episode, duration, and result. Blocks were interleaved rather than randomized or paired, which is why the timing comparison below is reported as suggestive.

Timing is suggestive, not established

Across all 50 trials, the median successful rollout took 53.6 seconds for Vanilla and 45.4 seconds for Pruned, a 8.2-second difference. The bootstrap interval crossed zero (−2.65 to 19.55 seconds) and the Mann–Whitney test gave p = 0.052808. This is suggestive timing evidence, not an established speed difference.

Why this differs from ordinary pruning

Ordinary pruning removes parameters using heuristics such as weight magnitude, redundancy, or measured sensitivity. This workflow begins with a causal question: which internal interface does the action expert actually use? The resulting map proposes a smaller architecture, which is then tested through retraining, held-out simulation, and physical hardware.

OpenVLA-OFT improves VLA fine-tuning and action generation. EfficientVLA compresses and accelerates diffusion-based VLAs without training. VLA-Cache reuses visual-token computation across control steps. Those methods optimize efficiency directly.

Mechanistic work by Häon and colleagues uses internal VLA representations for activation steering. Activation patching, causal tracing, and causal abstraction supply the broader intervention-based foundations. The evidence here combines those strands: causal localization predicts an interface redesign, held-out simulation tests the exact MolmoBot architecture, and a physical-robot pilot tests the broader layer-band principle in π0.5.

Among the cited work, that conjunction is the distinguishing contribution. It does not imply that no unpublished or contemporaneous work exists, nor that layer 24 transfers to another model.

What it took to validate the redesign

The localization sweep ran across three VLA architectures, each with its own set of interventions. The interface designs it suggested were then trained rather than argued for: policies at layers 24, 12 and 1, built by two different routes, a 50K-step retrofit of an existing policy and full two-stage pretraining. Each of those models was scored on roughly 1,000 held-out MolmoSpaces rollouts. The two π0.5 checkpoints behind the hardware test were each trained for 100K steps on the filtered DROID p98 subset, and the session itself ran 50 trials on a Franka FR3 at 15 Hz in four alternating blocks. Localization, retraining, held-out evaluation and a physical hardware test ran as one connected loop, and each stage was what set up the next.

What remains open

MolmoBot validates a single-layer design in simulation; π0.5 validates a three-layer band on a physical robot. The two tests are complementary, not identical.
The physical study used 25 trials per arm, block interleaving rather than randomization or pairing, and no pre-registered non-inferiority margin.
Completion time is computed over successful trials only. A definitive analysis should treat failures and timeouts as censored observations.
Unpinned flow-matching noise adds trial-level variance, so a timing claim needs a larger randomized or paired study.
Static operation reductions do not establish a wall-clock speedup.

Primary sources

Black, K. et al. π₀: A Vision-Language-Action Flow Model for General Robot Control. Robotics: Science and Systems XXI (2025). Primary record · arXiv
Black, K. et al. π₀.₅: a Vision-Language-Action Model with Open-World Generalization. Proceedings of Machine Learning Research 305 (CoRL 2025). Primary record · arXiv
Deshpande, A. et al. MolmoB0T: Large-Scale Simulation Enables Zero-Shot Manipulation. arXiv:2603.16861 (2026). Primary record
Kim, M. J., Finn, C. & Liang, P. Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. Robotics: Science and Systems XXI (2025). Primary record · arXiv
Yang, Y. et al. EfficientVLA: Training-Free Acceleration and Compression for Vision-Language-Action Models. Advances in Neural Information Processing Systems 38 (2025). Primary record · arXiv
Xu, S. et al. VLA-Cache: Efficient Vision-Language-Action Manipulation via Adaptive Token Caching. Advances in Neural Information Processing Systems 38 (2025). Primary record · arXiv
Häon, B., Stocking, K., Chuang, I. & Tomlin, C. Mechanistic Interpretability for Steering Vision-Language-Action Models. Proceedings of Machine Learning Research 305 (CoRL 2025). Primary record · arXiv
Zhang, F. & Nanda, N. Towards Best Practices of Activation Patching in Language Models: Metrics and Methods. ICLR (2024). Primary record · arXiv
Meng, K. et al. Locating and Editing Factual Associations in GPT. Advances in Neural Information Processing Systems 35 (2022). Primary record · arXiv
Geiger, A. et al. Causal Abstraction: A Theoretical Foundation for Mechanistic Interpretability. Journal of Machine Learning Research 26 (2025). Primary record · arXiv