Do Vision-Language-Action Models Mean What They Say?

01 Faithfulness

What is faithfulness?

Embodied Chain-of-Thought (CoT) promises to make Vision-Language-Action (VLA) models more capable and interpretable. But does this verbalized reasoning actually reflect the policy's internal decision process or is it a plausible story told after the fact?

We separate two properties of reasoning that are usually conflated. Functional reasoning improves task performance; faithful reasoning reflects the process by which the policy selects actions. Prior work optimizes for the former, and when addressed, faithfulness is typically reduced to reasoning-action alignment — a necessary but insufficient criterion that admits traces whose intermediate steps are ungrounded, mutually inconsistent, or disconnected.

A two-by-two matrix of functionality versus faithfulness for a fallen-tree driving scenario: a chain-of-thought can be functional without being faithful to the process that produces the action. — *Functionality and faithfulness are distinct axes of embodied reasoning.* **Functionality** (vertical) measures whether CoT improves task performance; **faithfulness** (horizontal) measures whether it reflects the process that actually produces the action. Hence, a reasoning trace may be functional, improving task performance, without being faithful to the model's underlying computations that produce an action.

Why do we study faithfulness?

If CoT is to expand a robot's domain of competency e.g., by exploring alternative solution strategies, then embodied reasoning must be load-bearing for action prediction, rather than a mechanism to rationalize bias and precomputed decisions. Hence, inspired by human-decision-making strategies and prior work, we hypothesize that faithful reasoning is important in the long-tail of robot experience, where data and expert supervision is limited.

Our Contributions

Through a human study on a State-of-The-Art (SoTA) driving model, we find that reasoning quality and trajectory improvement are only loosely coupled. Motivated by this observed gap, we formalize the notion of faithfulness for reasoning-based robot policies — characterizing precisely what it means for a generated trace to reflect a policy's underlying decision-making process — and operationalize it as a tractable behavioral consistency objective, deriving a scalable surrogate from pairwise semantic consistency constraints.

We instantiate this objective in Pinocchio, a learned critic that identifies inconsistencies between observations, intermediate reasoning, and actions to provide a dense reward signal for post-training. Our planner trained with Pinocchio improves faithfulness in embodied reasoning by +4% and +18% over SoTA alignment and trajectory-error baselines, respectively, with competitive task performance — and, interestingly, shows 1.6× the responsiveness compared to Alpamayo 1.5, a SoTA driving policy, on rare, counterfactual scenarios.

02 Method

Pinocchio: a learned critic of embodied faithfulness.

Frontier VLAs use an embodied CoT to produce actions, and while there is variantion in reasoning structure between models, a general theme is to draw inspiration from the foundational See-Think-Act paradigm in robotics. We build a reasoning VLA that factorizes CoT along this structure i.e., Observation → Scene → Move Justification → Meta-Action → Action, wherein we probe for signatures of unfaithful and inconsistent reasoning.

We formalize faithfulness as pairwise semantic consistency along the reasoning graph; then approximate each check with a fine-tuned critic aligned with human judgment that emits a binary <CONSISTENT> / <INCONSISTENT> verdict per edge. Its log-probabilities become a dense reward during alignment post-training.

A necessary condition, made trainable

Mechanistic faithfulness — proving the reasoning trace causes the action — is an interventional property, which is difficult to optimize directly in practice. We instead require that a faithful policy's generation act like a Markov chain, where each stage in the reasoning chain is semantically traceable to the one before it. This yields five checkable pairwise consistency constraints. In our paper, we prove consistency is necessary for mechanistic faithfulness (Prop. 1): a policy that fails these checks cannot be faithful, even though passing them cannot fully certify the underlying computation. Consistency is the tractable surrogate that lets us optimize for this necessary condition at scale.

Dense, decomposed reward

We turn the necessary condition for faithfulness into a soft, relaxed objective rather than a hard constraint. In practice, Pinocchio scores each of the five reasoning edges independently, so the policy learns exactly where it demonstrated unfaithful reasoning. Importantly, we validate our critic's training labels with a user study among four human annotators and judge the assessment of a frontier Vision-Language Model (VLM) against the human pool, allowing for scalable annotations aligned with human judgement. Then, we post-train a VLM planner against Pinocchio and a standard error metric (ADE) on a subset of expert driving demonstrations.

03 In the wild

Faithfulness pays off in the long tail.

Faithful reasoning means that the policy's justifications determine the resultant action. Hence, we leverage image inpainting tools (e.g., Nano Banana Pro) to construct a counterfactual benchmark, wherein we inpaint safety-critical hazards in otherwise benign imagery, to quantify the responsiveness of a policy's stated justifications and actions to out-of-distribution scenarios. We evaluate our planner trained with Pinocchio against a suite of baslines e.g., Alpamayo 1.5, and observe that our planner is the most responsive architecture, articulating waypoints that appropriately recognize the hazard in +7% of scenarios in comparison to Alpamayo.

Model	Reasoning ↑	Waypoints ↑	Overall ↑
Ours	18.2	31.8	7.6
LLM-Judge	9.1	21.2	4.5
ADE	16.7	24.2	4.5
Alpamayo-1.5-10B	16.9	15.4	4.6

Hazard response (%) on the adversarial long-tail benchmark. Reasoning and Waypoints measure independent response to the inpainted hazard; Overall measures their causal alignment.

On the Overall metric, which quantifies the percentage of cases in which the policy revises both its reasoning and its action appropriately to the controlled counterfactual intervention, our planner achieves a 1.6× responsiveness score compared to Alpamayo, the strongest baseline. Nevertheless, the low absolute scores across all methods indicate that robust, causally faithful reasoning in rare driving scenarios remains a persistent challenge for reasoning VLAs. Below we include select qualitative examples of each model's response to a synthetically generated hazard with the vehicle's waypoints represented through a Bird's Eye View and speed profile.

Injected cyclist

ADE verbalizes slowing but speeds up; Alpamayo plans to yield but never decelerates. Ours updates its justification and brakes.

Construction worker

Both baselines name the worker; only ours actively decelerates relative to its unaugmented prediction — a coherent physical reaction.

Stopped delivery truck

Others hallucinate a left nudge for a truck on the right. Ours is the only model whose reasoning and speed profile agree.

Qualitative counterfactuals from the adversarial benchmark. Click a card to enlarge the full comparison: original vs. augmented scene, predicted trajectory, speed profile, and reasoning trace for ADE (top row), ours (center), and Alpamayo 1.5 (bottom).

04 Deeper dive

Additional results.

Here you'll find a preliminary probe into the relationship between Reinforcement Learning (RL) trajectory gains and reasoning quality, and the performance of our planner on held-out nominal driving scenarios. We encourage you to read the paper for more discussion!

Trajectory gains are inconclusively connected to reasoning quality.

Present alignment strategies supervise embodied reasoning indirectly through action error, which raises a central question: Do trajectory-level gains from RL post-training translate into improved reasoning, or does the policy simply learn to route around the reasoning trace? We probe this question on the Alpamayo family, a frontier autonomous driving model, comparing the Supervised Fine-Tuned (SFT) and post-RL checkpoints. An annotator was shown a response from both models and asked 1) whether the responses were substantively different and 2) whether each response was grounded in the scene. We also record the Average Displacement Error (ADE) of each model.

Our preliminary finding indicates a weak and inconclusive coupling between reasoning quality and trajectory improvement under RL post-training. In the case of lateral error, RL frequently improves trajectory prediction even when the accompanying reasoning remains unchanged or degrades. A similar pattern emerges for the longitudinal case: RL achieves comparable improvements regardless of whether the reasoning is grounded, provided the responses are judged qualitatively similar. Our results indicate that trajectory gains from RL post-training are not always reflected in the generated reasoning, suggesting that the trace is inconclusively coupled to the policy's decision-making process.

Per-stratum win rate of the RL policy vs. the SFT baseline (ADE). Rows split by whether the RL response differs from SFT ("CoT changed") and which response aligns with the scene (S = SFT, R = RL; ✓ aligns, ✗ misaligns). Markers show deviation from a 50/50 split (right/blue: RL lower ADE; left/amber: baseline lower ADE); dot size ∝ n.

Highest overall faithfulness, competitive driving.

We evaluate the efficacy and faithfulness of our planner on nominal held-out German (DE) driving data (~20,000), disjoint from the US training set. We benchmark our planner against four Group Relative Policy Optimization reward variants initialized from the same SFT checkpoint to isolate the contribution of our faithfulnes objective. ADE is a standard trajectory error metric, VLM-Judge rewards causal alignment between generated and expert reasoning with a pre-trained VLM judge, while ADE-Reason and ADE-Swap are faithfulness baselines we introduce. Full implementation details are provided in the manuscript. Here, we evaluate both open-loop trajectory performance and reasoning faithfulness, with faithfulness measured by a frontier VLM validated against human judgement.

Model	ADE ↓	Overall ↑	E1	E2	E3	E4	E5
SFT	6.117	27.7	48.0	68.5	95.1	60.1	43.2
ADE	4.169	43.4	51.7	72.7	95.5	73.6	59.1
VLM-Judge	4.282	57.5	76.9	89.9	97.4	78.3	64.8
ADE-Reason	5.546	25.0	45.7	67.5	95.2	51.6	38.0
ADE-Swap	4.196	43.9	55.6	75.5	95.6	73.4	58.3
Pinocchio (ours)	4.324	61.4	69.2	86.4	97.3	82.5	71.1

Consistency in %. Best in column in bold. ADE averaged over three trials; consistency judged by Gemini. Evaluation on ~20,000 withheld DE scenarios.

This evaluation demonstrates a tension between open-loop prediction accuracy and reasoning faithfulness: ADE and ADE-Swap post the strongest trajectories, yet are also the most prone to justifications that are behaviorally inconsistent with the resulting action, echoing our earlier observation that trajectory gains don't necessarily track with reasoning quality. VLM-Judge performs best on the visually-grounded edges (E1, E2), suggesting limitations in our learned critic propagate directly into the downstream reward signal. Even so, our planner is strongest on overall faithfulness by substantially raising consistency on E4 and E5, i.e., whether the proposed action is supported by the model's reasoning, at only a modest ~5% cost in ADE, showing that faithfulness can be improved while largely preserving driving performance.

Do Vision-Language-Action Models Mean What They Say?
On the Role of Faithfulness in Embodied Reasoning