Changing a robot’s mind

Pick It Up: Does π0.5 Have an Editable Instruction Following Circuit?

Code & data ↗Watch the robots ↓Supporting figures ↓
In this report

What problem am I trying to solve?

I investigated whether there is an instruction following pathway for the instructions given to Vision Language Action (VLA) models and if it can be edited to influence real world behavior. This model biology question was inspired by Othello (Nanda et al.), especially using a model native coordinate system to solve for a hidden world state (latent space or variable). While the original paper intervened in the residual stream of a custom Transformer to predict the next move on boards, we edited the VLA’s image token positions to redirect a robot’s simulated behavior.

I hypothesized that fine-grained LLM style steering can successfully be applied to VLAs to change high level robot actions.

Why is this important?

Fine-grained control of models is increasingly important to solve alignment problems that reinforcement learning has limitations on. Models can behave in a black box and exhibit evolutionary greedy algorithm-like behavior when objective hacking (Hugging Face incidents). This is especially salient as we scale physical AI in the real world with VLAs and world models, some of which operate in latent space (JEPA, DINO).

These steering methods have seen early results that are promising with CAFT in LLMs and COAST in VLAs (Casademunt et al, Miao et al). Contributing to the applied interpretability toolkit for models that can take actions with real consequences is critical. This knowledge can help diagnose unsafe decisions and know which internal signals can control the models’ actions.

Most of all, LLMs often support cleaner token level forensics. I was interested in a less developed problem (VLAs) that had an ambiguous problem space and systematically use clean, mechanistic interpretability techniques to make a contribution to that space.

Finding 1: A broad image-state edit changed the robot’s action

I found that instructions did control actions (18/20 of the episodes succeeded). However, we had to replace all 512 Transformer positions in layers 12–17 to successfully intervene on robot behavior. This suggests the instruction effect is spread broadly across the model’s layers instead of in a few select positions. We found no evidence that fine-grained ablations at the token position level can successfully steer π0.5 in simulated robotics tasks, which responded only to broad edits. One possibility is that semantic representations of high-level instructions ("pick this object up") are spread across multiple token positions and layers, and may change during action generation.

As shown in the first graph above, we did find a pathway when giving the model the wrong instruction in π0.5 (and patching in the image position activations from the correct instruction scenes). However, this only had a significant effect after ablating much of the broader image token layers. This test assessed whether the internal visual signal could override the wrong instructions. This is interesting especially since vision is thought of as a “second class citizen” in VLAs by some robotics researchers (Jim Fan 2026).

Conflicting promptFailed
“Pick up the tomato sauce…” Tomato sauce is touched first; the cream-cheese task fails. No intervention.
Correct promptComplete
“Pick up the cream cheese…” Cream cheese is touched first and the task completes, without an intervention.
Late-state repairComplete
The prompt still says tomato sauce. Replacing image K/V at layers 12–17 redirects the robot to cream cheese and completes the task.
Early-layer controlFailed
The same replacement at layers 0–5 does not redirect this run: tomato sauce is touched first and the task fails.

High-resolution reruns of the illustrated case: matching task outcomes, with up to one simulator step of timing variation.

Code found here
Small patches barely move the action
Image-position K/V · prefix layers 12–17
Object-centeredCount-matched random
512 / 512
Replacing 128 or 256 selected image positions produced median action progress of 0.043 and 0.082; replacing all 512 produced 0.773. Curves show medians over 24 scene-direction rows: eight directed prompt-pair cells, each with three initial states. At 512, both selections are the same intervention. The animation reveals measured doses; it does not show elapsed training time.

Finding 2: Readable language was not controlling behavior

Conversely, editing text token states barely affects actions. It only has a 1% effect. In early layers, the model may keep the copy of the instructions in the text token states. However, these instructions may no longer directly control a robot’s actions after the model processes a scene.

Readable information and causal influence separate
Held-out decoding and paired forward-pass interventions

Instruction classification

Action progress after replacement

Text positionsImage positions
Left: held-out 10-way instruction-classification accuracy (chance = 0.10). Right: median action progress over six directed prompt-pair medians. The intervention study contains 300 paired forward-pass cases and 18,000 records, with shared prompts, initial states, and timepoints; these are not 300 independent prompts. This is an immediate-action experiment, not a rollout test.

Model biology for physical AI

Dissecting the model biology of models that operate with real world consequences is essential. It has practical safety implications, particularly since fields such as robotics are traditionally more closed-source when it comes to benchmarks, data, and model weights.

Late image-cache replacement has the strongest effect
Equal six-layer bands · all 512 image positions
Action progress for early, middle, and late six-layer image K/V replacement bands
Each point is one directed prompt-pair median; the line joins medians across six pairs. Layers 12–17 produce the strongest immediate-action effect. These are interventions on whole six-layer bands, not independent measurements at each layer.

Representation geometry

However, VLAs and world models have a tricky problem. They have longer causal chains than LLMs. Visual concept, language, and geometry can be entangled. And especially for JEPA family architecture, the model may learn its own private coordinate system in latent state space. This can mean mapping pathways or mechanisms that can be more easily found in LLMs, e.g. those associated with safety or obedience, may involve more complex, representational geometry questions.

One projection, two ways to label it
800 recorded initial states · two principal components
Color by

Select a label to highlight its points.

PCA of globally centered, concatenated mean image-position residuals entering layers 12–17 (12,288 dimensions). PC1 explains 52.4% of variance and PC2 explains 5.6%; the other 42.0% is not shown. This is a separate exploratory dataset, not the 80-episode repair experiment. Coordinates are fixed when colors change. Neither visible clusters nor separation prove a causal mechanism.

Future directions

The experiments used a limited set of simulated pick-and-place tasks, so they do not establish how broadly this intervention pattern generalizes. Perhaps the highest ROI intervention would have been to choose a stimuli set that specifically involved this stop/go mechanism to directly address the obedience mechanism that we were looking for.

Supporting figures

The dose curve varies across prompt pairs
Eight directed prompt-pair median dose curves for selected and random positions
Thin lines show eight directed prompt-pair medians over three initial states each; thick lines show the pooled descriptive medians used in the main dose curve. Random selection sometimes moves the action at a smaller dose, but not consistently across cases.
Task completion across all 20 matched cases
Twenty matched task and state cases under clean, conflict, late repair, and early repair conditions
One square per task/state case in each condition. Filled squares indicate simulator task success. The conditions share the same 20 starting cases: clean 19/20, conflict 0/20, late repair 18/20, early control 0/20.
Repair does not reproduce the clean trajectory exactly
Paired simulator episode lengths for clean and repair conditions across ten cases
Lines connect clean and repaired episode lengths for 10 additional task/state pairs. A cross marks the repair failure; its duration is the time limit, not time to successful completion. Eight repaired episodes lasted longer, with a median difference of +15 steps across all pairs. This is an exploratory screen.
Follow the hand through the recorded rollout
Top-down projection of measured end-effector positions
ConflictCorrectLate repairEarly control
280
Actual X–Y hand positions from task 1, initial state 20, the registered case illustrated in the videos. The third spatial dimension is omitted; this is a physical workspace projection, not latent space. Each trace stops at the end of its episode.
Links & references

Code and further reading

Related work already demonstrates VLA steering. This report focuses on the text/image cache comparison, the breadth of the tested intervention, and its behavioral validation in this checkpoint.