PaperlayerView exact revision on arXiv

Source-linked AI summary

Measuring the Reasoning-Action Coupling in Driving Vision-Language-Action Models through Counterfactuals

Gabriel Manalu, Saad Wazir, Thao Do, Seongah Kim, Hyeontaek Hwang, Tolcha Yalew Kidane, Daeyoung Kim

arXiv:2610.30953v1cs.AI

TL;DR

Driving VLAs present generated reasoning as an explanation of their actions, but separate text and trajectory evaluations do not show whether reasoning tracks visual causes or informs the resulting trajectory. This paper applies reasoning and scene-level counterfactual interventions to four models across benchmarks. It finds weak reasoning–action coupling, with reasoning updates not always matched by corresponding trajectory changes.

  • Problem

    Separate evaluations of generated reasoning and trajectories do not establish whether the trajectory depends on reasoning content or whether reasoning tracks visual evidence, leaving a key test of explainability unanswered.

    Source: §1 · p. 1, §3.1 · p. 3

  • Method

    The paper intervenes on generated reasoning while holding the scene fixed and removes reasoning-identified objects from camera frames to test reasoning and trajectory responses.

    Source: §3.1 · p. 3, §4.2 · p. 5

  • Results

    Across four driving VLAs, reasoning–action coupling is weak: meaning changes often produce limited action response, and revised reasoning after object removal does not always match the updated trajectory.

    Source: §7 · p. 9

  • Takeaways & Limitations

    Plausible, visually responsive reasoning alone is not a faithful explanation of a driving VLA’s action, so evaluation should also test whether stated reasoning accounts for its action.

    Source: §7 · p. 9

    The directional analysis is bounded by ORION’s small number of scenes with committed lateral decisions.

    Source: §6.3 · p. 8

Abstract

from arXiv · show

Driving vision-language-action (VLA) models bring an appealing form of explainability to a safety-critical domain, since a system that can state the reasons for its maneuvers offers the transparency that deployment and public trust in autonomous driving demand. These models read the world into language before they act on it, describing the scene and reasoning over it in text that ranges from chain-of-thought to multi-stage question answering on the way to a trajectory. That framing quietly assumes two separable properties, namely that the trajectory follows the generated reasoning and that the generated reasoning tracks the causes that actually produce the trajectory. To thoroughly investigate these properties, we apply causal, counterfactual interventions to four driving VLAs across different benchmarks. First, we perturb the models' own reasoning and show that altering its meaning produces small trajectory changes relative to blanking the same models' camera inputs, a pattern that persists across benchmarks and model architectures. We then construct scene-level counterfactuals through photorealistic removal of key objects named in the models' reasoning and analyze whether the generated reasoning and trajectories track these changes. Overall, our findings indicate that the reasoning these models present to account for their resulting driving behavior is weakly coupled to that behavior itself, pointing to limits on the explainability these systems are intended to provide.

1. Introduction

Separate text and trajectory scores do not establish whether a driving VLA’s reasoning governs its action or tracks visual causes. The paper tests these links with reasoning- and image-side counterfactuals, reporting weak coupling across four VLAs and two benchmarks.

1. Introduction

  • Separate evaluations of generated text and predicted trajectories do not establish that trajectories depend on reasoning content or that reasoning is grounded in cited visual evidence.The text may be plausible while the trajectory follows shortcut priors, even when both outputs score well separately.Source: §1 · p. 2
  • The paper asks whether a VLA’s trajectory depends on its reasoning’s semantic content and whether its reasoning and trajectory respond when the observed world changes.Source: §1 · p. 2
  • Reasoning-side interventions edit generated text while holding the scene fixed; image-side interventions remove an action-relevant object and regenerate both reasoning and trajectory.Together, the comparisons test whether reasoning steers action and whether edited visual evidence changes the model’s outputs.Source: §1 · p. 2
  • Across four driving VLAs, reasoning–action coupling is weak, and intervention responses reveal different model-specific explainability shortcomings.The evaluation spans two benchmarks.Source: §1 · p. 2
  • Meaning-preserving paraphrases control for wording variation, while camera blanking supplies a model-specific reference for visual sensitivity.Source: §1 · p. 2

2. Related Work

Prior driving-VLA work uses generated language for scene description, question answering, deliberation, or planning, but evaluations have not directly tested whether free-text reasoning explains the resulting trajectory. This paper adapts intervention-based faithfulness tests to trajectories and distinguishes them from evaluations based on discrete answers.

2. Related Work

  • Driving VLAs generate scene-conditioned language ranging from descriptions and question–answer outputs to deliberation and planning, sometimes jointly with action predictions.The paper uses “reasoning” for scene-conditioned language generated during inference before action prediction.Source: §2 · p. 2
  • Prior driving evaluations do not directly test whether the semantic content of generated free-text reasoning explains the resulting trajectory.Ablations of reasoning training do not test whether a fixed model’s trajectory causally depends on what its reasoning says for the same scene.Source: §2 · p. 2
  • Behavioral faithfulness tests intervene on reasoning traces or input evidence and assess resulting outputs, whereas many prior tests rely on discrete-answer changes.This work adapts intervention-based evaluation to semantic reasoning-to-action effects and responses to removal of reasoning-selected visual objects.Source: §2 · p. 2
  • Content-changing perturbations MISTAKE, NEGATE, and SWAP are compared against meaning-preserving PARAPHRASE.Source: Table 1 · p. 3

3. Measuring the Reasoning–Action Coupling

The section tests whether driving trajectories depend on the meaning of generated reasoning and whether reasoning and actions respond when a visually identified object is removed. It contrasts direct interventions on a fixed model with prior training-time ablations.

3.1. Motivation

  • Prior training-time ablations did not test whether a fixed model’s trajectory depends on the meaning of its own generated reasoning.The paper intervenes directly on generated reasoning and complements this with object-removal counterfactuals.Source: §3.1 · p. 3

3.2. Intervening on the generated reasoning

  • Reasoning-side interventions edit the model’s generated text while keeping inference conditions otherwise identical, then compare trajectory displacement with the unedited-reasoning baseline.A meaning-preserving PARAPHRASE control is subtracted to account for wording sensitivity.Source: §3.2 · p. 3

3.3. Intervening on the images

  • Image-side interventions remove an object identified as action-relevant, rerun the model on the edited input, and compare its new reasoning and trajectory with factual outputs.Trajectory responses are reported relative to camera blanking as a model-specific reference for visual sensitivity.Source: §3.3 · p. 4

4. Constructing the Visual Counterfactuals

The visual counterfactuals remove model-identified objects from PhysicalAI-AV scenes, using reviewed masks and a selected removal method. Candidate methods are compared by human ratings of realism, removal, associated effects, and preservation.

4. Constructing the Visual Counterfactuals

  • Object-removal counterfactuals are built on PhysicalAI-AV because DriveLM annotations usually lack a single removable key object identified by the model’s reasoning.Selected objects are model-specific because VLAs use different temporal contexts and may name different objects in the same scene.Source: §4 · p. 4

4.1. Locating the key object

  • The model’s key-object phrase is parsed and used to prompt an open-vocabulary detector, followed by segmentation and temporal mask propagation across relevant camera views.Automatic masks are reviewed and corrected manually before use.Source: §4.1 · pp. 4–5

4.2. Selecting an object-removal method for counterfactual generation

  • ObjectClear received the highest human rating on every question, with an overall mean of 4.61, and was selected for final counterfactual inputs.The four questions assess realism, target removal and reconstruction, removal of associated visual effects, and preservation outside the edited region.Source: §4.2 · p. 5
  • Model–scene edits are excluded if the object or identifying effects remain visible, or if reconstruction artifacts are substantial.The final quality check verifies removal of the key object and visually salient associated effects.Source: §4.2 · p. 5

5. Experimental Setup

The study evaluates four driving VLAs with different reasoning-to-action interfaces on held-out validation data from two benchmarks. It uses paired counterfactual comparisons across three seeds and measures intervention-induced trajectory displacement.

5.1. Models

  • The four evaluated VLAs connect reasoning to action through different interfaces, including autoregressive action tokens, a conditioned diffusion expert, and planning-token conditioning.Alpamayo-R1 and Alpamayo-1.5 share an architecture but differ in post-training; three additional models were excluded because their interfaces did not support localized semantic edits.Source: §5.1 · pp. 5–6

5.2. Datasets

  • DriveLM validation includes 799 nuScenes-derived scenes, while PhysicalAI-AV validation includes 349 real-world driving events.ORION is evaluated on DriveLM; Alpamayo-R1 and AutoVLA are evaluated on PhysicalAI-AV, with AutoVLA inputs constructed following its nuScenes protocol.Source: §5.2 · p. 6

5.3. Protocol and metrics

  • Comparisons are paired within scene and seed, with results based on three seeds unless stated otherwise.Reasoning edits are compared with a meaning-preserving paraphrase through the same injection path; image-side factual and counterfactual frames use otherwise identical evaluation inputs and configurations.Source: §5.3 · p. 6
  • ADE and FDE measure intervention-induced trajectory change: average waypoint displacement over the horizon and displacement at the final waypoint, respectively.These metrics compare baseline and counterfactual trajectories from the same model, rather than measuring planning error against an expert trajectory.Source: §5.3 · p. 6

6. Findings

Across the evaluated models and benchmarks, changing whether or how reasoning is supplied can affect trajectories more than changing its driving content. Object removal can update a model’s reasoning without ensuring its trajectory follows the revised maneuver, and stated direction does not reliably predict the factual trajectory for both Alpamayo models.

6. Findings

  • ADE summarizes the findings because ADE and FDE give the same qualitative ordering; larger intervention responses mean more trajectory change, not necessarily a better, worse, or less safe trajectory.Percentages are relative to each model’s camera-blanking response.Source: §6 · p. 7

6.1. Models split between reasoning-presence dependence and content-sensitive action

  • ORION’s DELETE/FILLER effects average 15.65% vs. 3.26% ADE for content-changing edits; Alpamayo-R1 responds most to MISTAKE, while SWAP and NEGATE remain below 3.20% ADE.For ORION, the reasoning slot’s presence matters more than its counterfactual driving content.Source: §6.1 · p. 7
  • Alpamayo-1.5 shows consistent content sensitivity, with all three content-changing interventions reaching 14.48%–19.46% of its camera-side ADE response.Even its largest effect, MISTAKE, reaches only 19.46% of that response.Source: §6.1 · p. 7

6.2. Object removal updates reasoning and action, but their updates can disagree

  • After object removal, Alpamayo-1.5 drops the removed object from its reasoning, but its trajectory does not always implement the regenerated maneuver.Figure 3 shows a qualitative case where reasoning updates but the trajectory does not stay faithful to it.Source: Figure 3 · p. 7, §6 · p. 7, §6.2 · p. 8

6.3. A directional finding: stated direction in reasoning does not reliably predict trajectories

  • AutoVLA follows its stated direction on 187 of 202 DriveLM scenes and 89.2% of PhysicalAI-AV scenes, while Alpamayo-1.5 and R1 do so on 71 of 155 and 59 of 121 scenes.Alpamayo-1.5’s median left/right endpoint separation is 0.01 m, and R1’s is 0.27 m; Figure 4 groups factual trajectories by stated direction.Source: §6.3 · p. 8, Figure 4 · p. 8

6.4. The reasoning–action disconnect is not a benchmark artifact

  • Across DriveLM and PhysicalAI-AV, AutoVLA’s DELETE and FILLER effects remain largest, so reasoning availability or form matters more than most changes to its stated driving content.The absolute response magnitudes and ordering among content-changing edits differ between sets.Source: §6.4 · p. 9

7. Conclusion

Across four driving VLAs, reasoning–action coupling is model-dependent and often weaker than generated text suggests. Paired reasoning edits, object-removal counterfactuals, and comparisons of stated versus decoded directions reveal limits to treating plausible, visually responsive reasoning as a faithful account of action.

7. Conclusion

  • Across four driving VLAs, reasoning–action coupling is model-dependent and often weaker than the generated text suggests.ORION responds more to reasoning presence or surface form than to counterfactual driving content; Alpamayo models show stronger but different content sensitivities.Source: §7 · p. 9
  • ORION trajectory changes track reasoning presence or surface form more than counterfactual driving content, while Alpamayo models show stronger but different content sensitivities.Alpamayo models’ stated leftward and rightward decisions also do not reliably describe their factual trajectories’ lateral directions.Source: §7 · p. 9
  • After object removal, models may revise their reasoning, but the updated trajectory does not always implement the described maneuver.Source: §7 · p. 9
  • Plausible, visually responsive reasoning is not by itself a faithful explanation of a driving VLA’s action.The conclusion calls for evaluating whether stated reasoning causally and consistently accounts for accompanying actions, alongside reasoning quality and trajectory accuracy.Source: §7 · p. 9

Related papers

Loading 2610.30953v1…