VLN Interpretability

What do VLM-Based Vision-Language Navigation Models Rely on: Interpreting and Steering Policy Behavior

Débora Oliveira Makowski*1, Samiran Gode*1, Abhijeet Nayak*1, Marco Hutter2, Cordelia Schmid3, Lukas Rosenberger Schmid1, Wolfram Burgard1
1 University of Technology Nuremberg · 2 ETH Zürich · 3 Inria, École Normale Supérieure, CNRS, PSL Research University

* Equal contribution. Listing order is random.

Spotlight CoRL ’26 workshop The Science of Physical AI Safety · SPAIS 2026 ↗
Paper arXiv Explainer video

Abstract

Understanding what drives a navigation decision

Modern Vision-Language Navigation (VLN) models rely mostly on pre-trained large Vision-Language Models (VLMs) to predict navigation actions. While this fusion of language instructions and visual observations allows multimodal reasoning, it obscures how information is routed across modalities or what mechanisms drive navigation decisions. Thus, it remains unclear whether VLN models ground their predictions in relevant semantic cues or can track task progress. In this work, we study the interpretability and steerability of VLN models. We use intervention-based metrics that measure how visual observations, instructions, and visual memory causally influence navigation decisions. Our results show that these navigation policies are sensitive to all input modalities and do not depend on a single one. We further show that these agents encode navigation progress and retain semantic structure from their VLM backbones, enabling concept-level steering through internal activations. Finally, we extract activation vectors for abstract behaviors to transfer them zero-shot to out-of-distribution real-world scenarios, improving performance without additional fine-tuning.

The question

What drives a VLN policy's decisions?

Success alone does not reveal whether a policy grounded the instruction, used a shortcut, or recovered through another modality. Under distribution shift, a failed behavior may be absent or encoded but unused.

  1. What changes the action?Which image regions, instruction spans, and memory observations change the path?
  2. Does sensitivity imply fragility?Does an open-loop action change predict closed-loop failure?
  3. Can behavior be recovered?Can retained internal representations restore a missing behavior?
Two navigation observations with visual and instruction sensitivity maps
Figure 1. Attribution heatmaps generated with ISS for CorrectNav. Red indicates higher sensitivity and blue lower sensitivity. In the instruction, darker tokens indicate higher sensitivity. Active subtask verbs (e.g., stop, through) and semantic cues (e.g., kitchen, doors, vestibule) are more sensitive. Even distant objects such as kitchen in trajectory 1 activate their corresponding textual concepts.

Main findings

Four findings from analysis to control

01

Modality use is architecture-specific.

All four policies use images, instructions, and history, but their dependence and progress-tracking patterns differ.

02

Sensitivity is not robustness.

An intervention can change one action without causing the largest drop in closed-loop success.

03

Navigation semantics survive fine-tuning.

Navigation-associated value vectors remain, and clamping their neurons shifts motion behavior.

04

Suppressed reasoning remains recoverable.

A visual shift reduces reasoning from 12% to 3%; a simulation-derived direction restores it and transfers to the robot.

Explainer video

Method overview

Test what changes the action

We change one input component, rerun the unchanged policy, and measure the trajectory difference. This tests causal influence more directly than attention alone.

01Start from a decision

Use the current image, instruction, and visual history.

02Perturb one input

Blur visual content or modify an instruction span.

03Rerun the policy

Keep the model weights and remaining context fixed.

04Measure the change

Compare trajectories by position and orientation.

Why use three attribution methods?

MethodQuestionUse in this study
ISSHow much does the path change after an input intervention?Behavioral attribution for images, instructions, and memory.
Integrated GradientsHow does fixed-output likelihood change from a corrupted to original input?Gradient-based visual reference.
TAMWhich input tokens support a generated output token?Token-level localization, especially for concrete concepts.

Two forms of internal steering

Attribution probes the inputs; steering tests whether a retained representation can change behavior without updating weights.

Semantic steering

Clamp concept-associated FFN neurons

Project value vectors through the language head, select neurons associated with navigation concepts, and clamp positive and negative groups.

selected concept neurons → activation β

Behavior steering

Add a contrastive hidden-state direction

Subtract the mean action state from the mean reasoning state, then add that direction at the target layer. α sets its strength.

h̃ = h + αd

Evidence from the paper

What each result establishes

Visual attribution → progress tracking → retained semantics → recovery under distribution shift.

01

Visual evidence: intervention maps are sharper

We evaluate ISS and IG on 200 decisions with human-marked evidence regions.

ISS and Integrated Gradients heatmaps for five navigation subtasks
Figure 2a. ISS (top) and Integrated Gradients (bottom). Warmer areas have higher attribution.
Token Activation Maps for screen, stairs, door, and move forward
Figure 2b. TAM localizes concrete concepts better than the abstract token move forward.
Reading the plots. ISS is generally sharper than IG. TAM localizes concrete output concepts but not abstract motion tokens. Attribution outside a human mask may still represent free space or geometry.
02

Memory and progress differ across policies

We corrupt earlier observations or instruction spans and measure action change over trajectory progress.

ISS heatmaps for visual memory and text instruction corruption across four VLN policies
Figure 3. ISS heatmap of visual-memory (top) and textual memory (bottom) intervention across trajectory progress. Each cell reports the mean ISS between the unperturbed action and the action predicted after blurring a given fraction of the images currently stored in memory or substituting part of the navigation instruction with its antonyms. All baselines are compared on the same ground truth trajectories. ISS scores increase from blue to red.

No single pattern generalizes.

CorrectNav and DualVLN increasingly use distributed visual history. AwareVLN and NaVILA depend more on recent observations.

Language-progress patterns also vary, reflecting differences in memory, temporal encoding, reasoning, and action horizons.

03

Navigation semantics remain after fine-tuning

We find navigation-associated FFN value vectors, then intervene on their neurons.

Navigation-token-associated value vectors across Qwen2.5-VL and DualVLN layers
Figure 4. Value vectors whose top vocabulary projections include each navigation token; this is not activation frequency.

What remains?

Words such as forward, left, right, and stop persist after DualVLN fine-tuning. Symbolic arrows become more visible in later layers.

Why intervene?

Projection is correlational. Clamping the selected neurons shifts motion-token frequencies, showing causal influence.

Steering contrastForwardRightLeft
forward positive / turn negative+3.2%-0.8%-19.3%
left positive / right negative+0.8%-3.0%+7.8%
Takeaway. Before retraining a failed policy, test whether the relevant concept remains internally accessible.
04

Visual shift suppresses but does not erase reasoning

We compare matched simulator states under the training camera and a robot-like fisheye camera.

Nominal pinhole observation, simulated fisheye shift, and real robot camera observation
Figure 5. Training view, simulated visual shift, and real robot camera.

What changes?

Reasoning falls from 12 ± 2% to 3 ± 1%. Adding a simulation-derived action-to-reasoning activation direction restores it as α increases.

A random direction does not produce this effect, linking recovery to the learned direction rather than perturbation alone.

Conditionα=.5α=1α=2α=4
Nominal camera12 ± 2%
Visual shift3 ± 1%
Random direction3.0%2.7%2.6%1.6%
Mean reasoning direction20%55%90%100%

Real-world deployment

Zero-shot recovery on the robot

The reasoning direction is computed in simulation and applied without robot-specific steering data or weight updates.

The baseline runs first; an operator enables steering after an uncorrected wrong choice or loop. This is a sequential recovery protocol, not two independent policy runs.

Across 43 trials, baseline completes 16/43 (37.2%). Steering recovers 16/27 failures (59.3%), raising cumulative completion to 32/43 (74.4%).
Add: static/videos/exit-the-room.mp4
The instruction is "to exit the room", the robot goes the opposite way at which reasoning is triggered using activation steering and the model corrects itself
Add: static/videos/turn-right-and-stop-plants.mp4
The instruction is "to turn right and stop at the plants", the robot starts turning in place at which point reasoning is triggered using activation steering and the model goes the correct way.
Scope. Operator-triggered recovery is not compared with an unsteered continuation using the same extra action budget. Automatic triggering remains future work.

For the community

What can be reused

Evaluation

Separate sensitivity from robustness.

Pair fixed-state interventions with closed-loop evaluation.

Interpretability

Match the method to the question.

ISS, IG, and TAM expose different input–output relationships.

Adaptation

Test recoverability before retraining.

A shifted policy may retain a useful internal capability.

Citation

BibTeX

@misc{makowski2026vlmbasedvisionlanguagenavigationmodels,
  title         = {What do VLM-Based Vision-Language Navigation Models Rely on:
                   Interpreting and Steering Policy Behavior},
  author        = {Débora Oliveira Makowski and Samiran Gode and Abhijeet Nayak
                   and Marco Hutter and Cordelia Schmid
                   and Lukas Rosenberger Schmid and Wolfram Burgard},
  year          = {2026},
  eprint        = {2609.24576},
  archivePrefix = {arXiv},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2609.24576}
}