Pith. sign in

REVIEW 4 major objections 5 minor 70 references

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

T0 review · 4 major / 5 minor · reviewed 2026-07-11 · grok-4.5

Pith's one-line read Faithful reasoning traces—not just functional ones—make embodied VLAs more robust to rare hazards.

desk verdict Solid training recipe for multi-edge faithfulness in driving VLAs; the 1.6x OOD number is real but Gemini-tied, so treat it as directional. read the letter →

arxiv 2607.04681 v1 pith:DWJ46F7G submitted 2026-07-06 cs.RO cs.AI

classification cs.ROcs.AI
keywords embodiedreasoningvision-language-actionmodelsfaithfulnesschain-of-thoughtautonomousdrivingreinforcementlearningPinocchiocriticcounterfactualevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Vision-language-action models often produce chain-of-thought explanations that look helpful but may not actually drive the actions they accompany. This paper separates functional reasoning (does the text help performance?) from faithful reasoning (does the text track the real decision process?). Using autonomous driving as the testbed, a human study shows that trajectory gains from reinforcement learning do not reliably improve reasoning quality. The authors then train a critic, Pinocchio, that scores whether each step of the reasoning is grounded in the scene and coherent with the next step and the final waypoints, and they use those scores as dense rewards. The resulting planner raises overall faithfulness while keeping competitive trajectory error, and on synthetic long-tail hazards it responds with causally aligned reasoning and motion 1.6 times more often than a strong baseline. The claim is that making reasoning load-bearing, rather than a post-hoc rationalization, is a practical route to more trustworthy robot policies.

What carries the argument

Pinocchio: a learned VLM critic that scores pairwise semantic consistency along the edges of a first-order Markov reasoning graph (observation→scene, scene→justification, justification→meta-action, meta-action→waypoints, justification→waypoints). Its log-probabilities supply a dense, decomposable faithfulness reward inside GRPO post-training.

What would settle it

A controlled intervention that overwrites an intermediate reasoning step (for example, changing the meta-action from lane-keeping to turn-right) and shows that the predicted waypoints still ignore the overwrite and follow the original observation, proving the trace is not load-bearing.

Watch

Extended reading notes

Core claim

State-of-the-art alignment that only requires the final action to match the text is necessary but insufficient for faithfulness; intermediate steps can still be ungrounded or disconnected. Optimizing a learned critic of observation grounding and stepwise coherence during RL post-training improves overall faithfulness by 4% over a VLM-judge alignment baseline and 18% over trajectory-error-only RL on held-out driving data, while remaining competitive on average displacement error, and yields 1.6× better causal hazard response on a synthetic counterfactual set.

Load-bearing premise

That the product of five pairwise consistency checks is a tight enough stand-in for true causal mediation, so that rewarding those checks actually forces the action to be generated through the reasoning rather than merely rationalized afterward.

Editorial extensions

If this is right

  • Post-training rewards can and should supervise intermediate reasoning edges, not only final trajectory error or action–text agreement.
  • Policies trained this way remain competitive on ordinary benchmarks while becoming more responsive when rare hazards appear.
  • Faithfulness metrics that only check final action consistency will systematically under-detect ungrounded or disconnected CoT.
  • Synthetic counterfactual scene edits become a practical probe of whether stated justifications actually control behavior.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same pairwise-consistency critic could be applied at inference time as a runtime filter that aborts or re-samples inconsistent rollouts.
  • If the necessity gap is large, closed-loop simulators that force reactive recovery will expose residual unfaithfulness that open-loop ADE misses.
  • Edge-level rewards may transfer to other VLA domains (manipulation, navigation) wherever CoT is structured into perception–justification–action stages.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper distinguishes functional from faithful embodied Chain-of-Thought in VLAs and argues that action-level alignment is necessary but insufficient for faithfulness. Using autonomous driving as a testbed, it reports a human study on Alpamayo showing weak coupling between RL trajectory gains and reasoning quality (Section 4, Figure 2). It formalizes mechanistic faithfulness via a first-order Markov DAG (Eqs. 2–3) and a behavioral surrogate as the product of five pairwise semantic-consistency relations (Eq. 4, set C). A VLM critic, Pinocchio, is trained on Gemini-labeled edges and used as a dense GRPO reward (Eq. 7). On held-out DE/US data the method improves overall consistency by ~4% over a VLM-Judge baseline and ~18% over ADE-only RL while remaining competitive on ADE (Tables 3–4). On a 66-scene synthetic counterfactual hazard set it reports 1.6× overall causal hazard response relative to Alpamayo-1.5-10B (Table 5).

Significance. If the results hold under independent evaluation, the work would be a meaningful contribution to embodied reasoning: it elevates faithfulness from a post-hoc interpretability property to an explicit training objective, supplies a structured multi-edge consistency decomposition that is more granular than action–text alignment, and shows that optimizing that objective can improve long-tail responsiveness without large ADE collapse. Strengths include the human pilot of Gemini as judge (Table 1, κ within inter-annotator range on four of five edges), balanced-accuracy validation of Pinocchio (Table 2), controlled GRPO baselines spanning functionality and faithfulness (Section 6.2), and an explicit necessary-but-not-sufficient statement of the behavioral surrogate (Section 5.2, Limitations). The formalization and the dense multi-edge reward design are useful even if the OOD multiplier is revised.

major comments (4)
  1. [Section 6.3, Table 5; also Tables 3–4] Tables 3–5 and the OOD protocol (Section 6.3, Appendix 8.6.1) use Gemini 3.1 Pro both to label D_faith for critic training (Section 6.1) and to score final consistency and hazard response. Although the human pilot (Table 1) shows Gemini–majority agreement within human–human range on four of five edges, the 1.6× Overall claim in Table 5 is therefore partly a same-judge-family agreement metric. An independent human or held-out non-Gemini judge on at least the OOD set (and ideally a subsample of Tables 3–4) is needed before the headline long-tail claim can be treated as established.
  2. [Section 5.1–5.2, Eqs. (3)–(4), (7); Limitations] Proposition 1 and Section 5.2 correctly state that F(τ)=1 (Eq. 4) is necessary but not sufficient for mechanistic faithfulness (Eq. 3): a policy can emit fully consistent traces while computing a via a pathway that bypasses z. The GRPO objective (Eq. 7) maximizes critic log-probabilities, not interventional mediation. The paper’s central claim that faithful traces “contribute to more robust, generalizable” behavior therefore rests on a behavioral surrogate whose gap to Eq. 3 is untested. At minimum, report a simple intervention test (e.g., overwrite z_k and measure change in a, as sketched after Eq. 3) on a subset of rollouts for the final policy versus ADE and VLM-Judge; without this, the mechanistic language in the abstract and conclusion overreaches the evidence.
  3. [Section 6.3, Table 5] The OOD benchmark has N=66 synthetic inpainted scenes, no ground-truth trajectories, and absolute Overall rates of only 7.6% (Ours) vs 4.5–4.6% (baselines) (Table 5). The 1.6× relative gain is therefore sensitive to small absolute counts and to Gemini’s hazard-response rubric. Report confidence intervals or bootstrap over scenes, and either enlarge the set or add a non-synthetic long-tail subset (e.g., real rare events from the cited WOD-E2E taxonomy) so the robustness claim does not rest solely on AI-generated imagery that none of the models saw in training.
  4. [Section 6.2.2, Tables 3–4, Eq. (7)] On DE data, Ours ADE is 4.324 vs ADE baseline 4.169 (~3.7% worse) and vs ADE-Swap 4.196; on US, 3.858 vs 3.734–3.745 (Table 3–4). The abstract’s “maintaining competitive downstream task performance” is fair, but the paper should quantify whether faithfulness gains survive under a constrained multi-objective setting (e.g., Pareto front or λ-sweep on ADE vs r_faith) rather than a single fixed equal-weight composite. Without that, it remains unclear whether the method improves the functionality–faithfulness frontier or merely trades a small amount of ADE for Gemini-aligned consistency.
minor comments (5)
  1. [Section 4, Figure 2] Figure 2’s win-rate visualization is hard to parse (overlapping markers, “pp from 50%”); a tabular breakdown of n and win rates per stratum would make the Section 4 claim easier to verify.
  2. [Section 6.1, Appendix 8.3.2] Edge naming is inconsistent across the main text (E1–E5) and Appendix 8.3.2 (image→justification vs image→scene); align labels so readers can map Table 1–4 columns to C without ambiguity.
  3. [Abstract; Tables 3–4] The abstract states “improves faithfulness by 4% and 18%”; specify that these are absolute percentage-point gains on overall consistency (e.g., 61.4 vs 57.5 and 43.4 on DE) so they are not read as relative improvements.
  4. [Limitations] Limitations correctly flag open-loop evaluation; a short closed-loop note or deferred experiment plan would strengthen the outlook without requiring full simulator results in this revision.
  5. [Throughout] Typographical inconsistencies appear (e.g., “SoT A”, “W ork”, “T races”, “V alidating”); a pass for spacing and OCR-like artifacts would improve polish.

Circularity Check

1 steps flagged · score 3.0 of 10

Moderate evaluation loop from Gemini labeling the critic training set and also scoring the reported faithfulness/OOD metrics; the core formalization and RL objective are not definitionally circular.

  1. other [Sections 6.1.1–6.1.3 (critic training) + 6.2.2 / Tables 3–4 (evaluation) + 6.3 / Table 5 (OOD)]
    "we query Gemini 3.1 Pro ... to judge the consistency of the five edges in Ggen ... We therefore use Gemini to label all of Dfaith ... faithfulness measured using the Gemini-based protocol from Section 6.1.1 ... These judgments are produced by Gemini using a hazard-response evaluation prompt"

    Gemini supplies the edge-level CONSISTENT/INCONSISTENT labels that train Pinocchio (the dense reward). The same Gemini family then produces the overall consistency percentages and the OOD hazard-response rates that constitute the paper’s strongest empirical claims. Because the reward model is fitted to Gemini’s judgments, subsequent gains on Gemini-scored metrics can partly reflect agreement with the judge rather than independent causal mediation (Eq. 3). The loop is imperfect (Pinocchio is a different frozen model; human pilot exists) but still contaminates the headline numbers.

full rationale

The paper is an empirical robotics/ML methods paper, not a first-principles derivation. Proposition 1 correctly states only necessity of the product of pairwise consistencies F(τ) for mechanistic faithfulness, and the authors repeatedly flag that the surrogate is necessary-but-not-sufficient (Section 5.2, Limitations). The RL objective (Eq. 7) optimizes a frozen learned critic (Pinocchio, a fine-tuned Qwen3-VL-4B) rather than Gemini itself; the policy is therefore not optimizing the evaluation metric by construction. Human pilot agreement (Table 1) further shows Gemini is within inter-annotator range on 4/5 edges. The residual circularity is the shared Gemini family for (i) constructing Dfaith labels that train the critic and (ii) the Gemini-based protocol that produces the headline consistency numbers in Tables 3–4 and the 1.6× OOD hazard-response numbers in Table 5. This is a real but moderate contamination risk of the kind common in LLM-as-judge pipelines; it does not collapse the claimed gains to a tautology. No self-definitional equations, no uniqueness theorems imported from the authors, no ansatz smuggled via self-citation, and no renaming of a known result. Score 3 reflects the shared-judge loop without over-claiming definitional circularity.

Assumptions & free parameters 2 free parameters · 3 assumptions · 2 invented entities

The central claim rests on a first-order Markov generative assumption for the reasoning trace, on the necessity (but not sufficiency) of pairwise semantic consistency for mechanistic faithfulness, and on the reliability of a Gemini-derived critic as a dense reward. No continuous free parameters are fitted to the final evaluation metrics; the λ weights are treated as fixed hyper-parameters. The main invented entity is the Pinocchio critic itself.

free parameters (2)
  • λ_u,v reward weights
    Fixed hyper-parameters that trade off ADE against each consistency edge; tuned on a held-out validation set but not dual-ascended (Section 5.2).
  • GRPO group size / learning rate / epochs
    Standard RL hyper-parameters (batch 8, 16 rollouts, 4 epochs, 1e-6) that affect absolute numbers but are shared across all baselines.
assumptions (3)
  • ad hoc to paper A faithful reasoning policy factorizes as a first-order Markov chain along the DAG G_gen (Eq. 2).
    Adopted to make every edge locally interrogable; not derived from first principles of the underlying VLA.
  • domain assumption Trace consistency F(τ)=1 is a necessary condition for mechanistic faithfulness (Proposition 1).
    Standard causal-mediation reasoning; the paper correctly notes it is not sufficient.
  • domain assumption Gemini 3.1 Pro judgments of semantic consistency are sufficiently reliable to serve as large-scale labels and evaluation.
    Supported by a 100-sample human pilot (Table 1) but remains an external modeling choice.
invented entities (2)
  • Pinocchio critic c_ϕ
    purpose: Dense, multi-edge reward that scores observation grounding and stepwise coherence for GRPO.
    A fine-tuned 4B VLM trained on Gemini-labeled edges; no independent existence outside this training pipeline.
  • Five-edge consistency set C
    purpose: Operationalizes behavioral faithfulness as product of pairwise semantic relations.
    Chosen by the authors to close rationalization loopholes; not uniquely determined by theory.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Do Vision-Language-Action Models Mean What They Say? On the Role of Faithfulness in Embodied Reasoning." pith.science (2026). https://pith.science/paper/DWJ46F7G

@misc{pith2026260704681,
  author       = {Pith},
  title        = {Pith review of: Do Vision-Language-Action Models Mean What They Say? On the Role of Faithfulness in Embodied Reasoning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DWJ46F7G}},
  note         = {Machine review of arXiv:2607.04681}
}
read the original abstract

Embodied Chain-of-Thought has emerged as a promising mechanism to enhance robot decision-making and interpretability in black-box Vision-Language Action (VLA) models. However, whether this verbalized Chain-of-Thought truthfully reflects the policy's underlying decision process remains poorly understood. We distinguish between functional reasoning, in which reasoning improves task performance, and faithful reasoning, in which reasoning truly reflects the policy's internal decision process. We argue that SoTA alignment strategies offer a necessary but insufficient notion of faithfulness, admitting reasoning whose intermediate steps can mask the causal links in action prediction through confounding factors (e.g., reasoning that is ungrounded in the environment and internally disconnected or inconsistent), restricting policy generalization. We study this gap through a human evaluation of a SoTA reasoning model for autonomous driving, revealing an inconsistent coupling between reasoning quality and downstream trajectory improvement. We then operationalize a behavioral surrogate for embodied faithfulness through a learned critic, Pinocchio, scoring observation grounding and stepwise coherence, and use this critic as a dense reward signal in post-training an embodied policy with reinforcement learning. Across withheld driving benchmarks, our post-trained planner improves faithfulness by 4% and 18% over SoTA alignment and trajectory error post-training baselines, respectively, while maintaining competitive downstream task performance. Finally, on a synthetic out-of-distribution test set, post-training for faithfulness improves policy responsiveness to rare counterfactual scenarios by 1.6x that of a SoTA policy, suggesting that faithful reasoning traces contribute to more robust, generalizable, and interpretable embodied intelligence. Project page: https://mjf-su.github.io/pinocchio/

Figures

Figures reproduced from arXiv: 2607.04681 by the authors.

Figure 1
Figure 1. Functionality and faithfulness are distinct axes of embodied reasoning. Functionality (vertical) measures whether the CoT improves task performance; faithfulness (horizontal) measures whether it reflects the process that actually produces the action. In this work, we revisit the role of reasoning in physical intelligence through the lens of faithfulness and introduce Pinocchio, a learned critic that operationalizes … view at source ↗
Figure 2
Figure 2. Per-stratum win rate of the RL policy vs. SFT baseline (ADE). Rows are split by whether the response from the RL policy is demonstrably different than SFT baseline (i.e., “CoT Changed”) and by which model response aligns with the driving scene (S=SFT, R=RL; ✓=aligns, ✗=misaligns). Markers show deviation from a 50/50 split (right/blue: RL lower ADE; left/orange: baseline lower ADE); size of each dot is ∝n. 4 [PITH_F… view at source ↗
Figure 3
Figure 3. Model responses to a synthetically injected cyclist. The ADE baseline (top) and Alpamayo-1.5 (bottom) [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The three example images provided to Gemini for a chain-of-thought annotation request. The camera frame [PITH_FULL_IMAGE:figures/full_fig_p014_4.png]
Figure 5
Figure 5. Figure 5: Model responses to scene augmentation. A cyclist is introduced on the right side of the intersection. [PITH_FULL_IMAGE:figures/full_fig_p029_5.png]
Figure 6
Figure 6. Figure 6: Model responses to scene augmentation. A construction worker is added to an intersection scene. Alpamayo [PITH_FULL_IMAGE:figures/full_fig_p029_6.png]
Figure 7
Figure 7. Figure 7: Model responses to scene augmentation. An original driving scene is augmented by inserting a person in a [PITH_FULL_IMAGE:figures/full_fig_p030_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

70 extracted references · 1 canonical work pages

  1. [1]

    Adcock et al

    A. Adcock et al. The Llama 4 herd: Architecture, training, evaluation, and deployment notes. arXiv preprint, arXiv:2601.11659, 2026. URLhttps://api.semanticscholar.org/CorpusID: 284910371

  2. [2]

    Arcuschin, J

    I. Arcuschin, J. Janiak, R. Krzyzanowski, S. Rajamanoharan, N. Nanda, and A. Conmy. Chain-of- thought reasoning in the wild is not always faithful.arXiv preprint, arXiv:2503.08679, 2025

  3. [3]

    Bai et al

    S. Bai et al. Qwen3-VL technical report.arXiv preprint, arXiv:2511.21631, 2025. URLhttps: //arxiv.org/abs/2511.21631

  4. [4]

    Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Robert Equi, Chelsea Finn, Niccolo Fusai, Manuel Y . Galliker, Dibya Ghosh, Lachy Groom, Karol Hausman, brian ichter, Szymon Jakubczak, Tim Jones, Liyiming Ke, Devin LeBlanc, Sergey Levine, Adrian Li-Bell, Mohith Mothukuri, Suraj Nair, Karl Pertsch, Allen Z. Ren,...

  5. [5]

    Bommasani et al

    R. Bommasani et al. On the opportunities and risks of foundation models.arXiv preprint, 2021. URLhttps://crfm.stanford.edu/assets/report.pdf

  6. [6]

    Training strategies for efficient embodied reasoning.arXiv preprint arXiv:2505.08243, 2025

    William Chen, Suneel Belkhale, Suvir Mirchandani, Oier Mees, Danny Driess, Karl Pertsch, and Sergey Levine. Training strategies for efficient embodied reasoning.arXiv preprint arXiv:2505.08243, 2025

  7. [7]

    Droid: A large-scale in-the-wild robot manipulation dataset.ArXiv, abs/2403.12945, 2024

    Alexander Khazatsky et al. Droid: A large-scale in-the-wild robot manipulation dataset.ArXiv, abs/2403.12945, 2024. URLhttps://api.semanticscholar.org/CorpusID:268531351

  8. [8]

    π0.5: a vision-language-action model with open-world generalization.ArXiv, abs/2504.16054, 2025

    Kevin Black et al. π0.5: a vision-language-action model with open-world generalization.ArXiv, abs/2504.16054, 2025. URLhttps://api.semanticscholar.org/CorpusID:277993634

Show all 70 references
  1. [9]

    Measuring faithfulness in chain-of-thought reasoning.ArXiv, abs/2307.13702,

    Tamera Lanham et al. Measuring faithfulness in chain-of-thought reasoning.ArXiv, abs/2307.13702,

  2. [10]

    URLhttps://api.semanticscholar.org/CorpusID:259953372

  3. [11]

    Causal mediation analysis: what is it and how can it be used to inform practice and policy?F amily Practice, 42 (4):cmaf043, 2025

    P Fernainy, C Godard-Sebillotte, A Lacasse, G Layani, C Longo, J Kaczorowski, MA Rodriguez, ME Poitras, M Breton, MT Lussier, Y Couturier, C Hudon, and N Sourial. Causal mediation analysis: what is it and how can it be used to inform practice and policy?F amily Practice, 42 (4...

  4. [12]

    Self-Supervised Bootstrapping of Action-Predictive Embodied Reasoning

    Milan Ganai, Katie Luo, Jonas Frey, Clark Barrett, and Marco Pavone. Self-Supervised Bootstrapping of Action-Predictive Embodied Reasoning. InProceedings of Robotics: Science and Systems, Sydney, Australia, July 2026

  5. [13]

    Steervla: Steering vision-language-action models in long-tail driving scenarios.ArXiv, 2026

    Tian Gao, Celine Tan, Catherine Glossop, Timothy Gao, Jiankai Sun, Kyle Stachowicz, Shirley Wu, Oier Mees, Dorsa Sadigh, Sergey Levine, and Chelsea Finn. Steervla: Steering vision-language-action models in long-tail driving scenarios.ArXiv, 2026. URLhttps://arxiv.org/abs/2602.08440

  6. [14]

    Gemini 3 pro model card

    Google DeepMind. Gemini 3 pro model card. https://storage.googleapis.com/ deepmind-media/Model-Cards/Gemini-3-Pro-Model-Card.pdf, December 2025. Model card

  7. [15]

    D. Guo, D. Y ang, H. Zhang, J. Song, P . Wang, Q. Zhu, R. Xu, R. Zhang, S. Ma, X. Bi, X. Zhang, X. Y u, Y . Wu, Z. F. Wu, et al. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning.Nature, 645(8081):633–638, 2025. doi: 10.1038/s41586-025-09422-z

  8. [16]

    J. Hu, Y . Zhang, Q. Han, D. Jiang, X. Zhang, and H.-Y . Shum. Open-reasoner-zero: An open source approach to scaling up reinforcement learning on the base model. InConf. on Neural Information Processing Systems, 2025. URLhttps://openreview.net/forum?id=NFM8F5cV0V. 11

  9. [17]

    Open-reasoner-zero: An open source approach to scaling up reinforcement learning on the base model.ArXiv, abs/2503.24290, 2025

    Jingcheng Hu, Yinmin Zhang, Qi Han, Daxin Jiang, Xiangyu Zhang, and Heung yeung Shum. Open-reasoner-zero: An open source approach to scaling up reinforcement learning on the base model.ArXiv, abs/2503.24290, 2025. URL https://api.semanticscholar.org/CorpusID: 277468189

  10. [18]

    Hwang, R

    J.-J. Hwang, R. Xu, H. Lin, W.-C. Hung, J. Ji, K. Choi, D. Huang, T. He, P . Covington, B. Sapp, Y . Zhou, J. Guo, D. Anguelov, and M. Tan. EMMA: End-to-end multimodal model for autonomous driving.arXiv preprint, arXiv:2410.23262, 2024

  11. [19]

    Alphadrive: Unleash- ing the power of vlms in autonomous driving via reinforcement learning and reasoning.ArXiv, abs/2503.07608, 2025

    Bo Jiang, Shaoyu Chen, Qian Zhang, Wenyu Liu, and Xinggang Wang. Alphadrive: Unleash- ing the power of vlms in autonomous driving via reinforcement learning and reasoning.ArXiv, abs/2503.07608, 2025. URLhttps://api.semanticscholar.org/CorpusID:276928398

  12. [20]

    Farrar, Straus and Giroux, New Y ork, 2011

    Daniel Kahneman.Thinking, F ast and Slow. Farrar, Straus and Giroux, New Y ork, 2011

  13. [21]

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P . Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P . Liang, and C. Finn. OpenVLA: An open-source vision-language-action model.arXiv preprint...

  14. [22]

    Drive-r1: Bridging reasoning and planning in vlms for autonomous driving with reinforcement learning.ArXiv, abs/2506.18234, 2025

    Y ue Li, Meng Tian, Dechang Zhu, Jiangtong Zhu, Zhenyu Lin, Zhiwei Xiong, and Xinhai Zhao. Drive-r1: Bridging reasoning and planning in vlms for autonomous driving with reinforcement learning.ArXiv, abs/2506.18234, 2025. URLhttps://api.semanticscholar.org/CorpusID: 279999326

  15. [23]

    Adathinkdrive: Adaptive thinking via reinforcement learning for autonomous driving.ArXiv, abs/2509.13769, 2025

    Y uechen Luo, Fang Li, Shaoqing Xu, Zhiyi Lai, Lei Y ang, Qimao Chen, Ziang Luo, Zixun Xie, Shengyin Jiang, Jiaxin Liu, Long Chen, Bing Wang, and Zhi-Xin Y ang. Adathinkdrive: Adaptive thinking via reinforcement learning for autonomous driving.ArXiv, abs/2509.13769, 2025. URL ...

  16. [24]

    Walk the talk? measuring the faithfulness of large language model explanations

    Katie Matton, Robert Ness, John Guttag, and Emre Kiciman. Walk the talk? measuring the faithfulness of large language model explanations. InThe Thirteenth International Conference on Learning Representations, 2025. URLhttps://openreview.net/forum?id=4ub9gpx9xw

  17. [25]

    Introducing nano banana pro.https://blog.google/ innovation-and-ai/products/nano-banana-pro/, 2025

    Google Deepmind Naina Raisinghani. Introducing nano banana pro.https://blog.google/ innovation-and-ai/products/nano-banana-pro/, 2025. Accessed: 2026-06-24

  18. [26]

    Bjorck, F

    NVIDIA, J. Bjorck, F. Castaneda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y . Fang, D. Fox, F. Hu, et al. GR00T N1: An open foundation model for generalist humanoid robots.arXiv preprint, arXiv:2503.14734, 2025

  19. [27]

    Wang, et al

    NVIDIA, Y . Wang, et al. Alpamayo-R1: Bridging reasoning and action prediction for generalizable autonomous driving in the long tail.arXiv preprint, arXiv:2511.00088, 2025

  20. [28]

    O’Neill, A

    Open X-Embodiment Collaboration, A. O’Neill, A. Rehman, A. Gupta, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, et al. Open X-Embodiment: Robotic learning datasets and RT-X models. InProc. IEEE Conf. on Robotics and Automation, pages 6892–6903, 2024

  21. [29]

    Making reasoning matter: Measuring and improving faithfulness of chain-of-thought reasoning

    Debjit Paul, Robert West, Antoine Bosselut, and Boi Faltings. Making reasoning matter: Measuring and improving faithfulness of chain-of-thought reasoning. InConference on Empirical Methods in Natural Language Processing, 2024. URLhttps://api.semanticscholar.org/CorpusID: 267770195

  22. [30]

    Z. Peng, W. Ding, Y . Y ou, Y . Chen, W. Luo, T. Tian, Y . Cao, A. Sharma, D. Xu, B. Ivanovic, B. Li, B. Zhou, Y . Wang, and M. Pavone. Counterfactual VLA: Self-reflective vision-language-action model with adaptive reasoning.arXiv preprint, arXiv:2512.24426, 2025

  23. [31]

    Radford et al

    A. Radford et al. Learning transferable visual models from natural language supervision. In Int. Conf. on Machine Learning, 2021. URLhttps://api.semanticscholar.org/CorpusID: 231591445

  24. [32]

    Simlingo: Vision-only closed-loop autonomous driving with language-action alignment

    Katrin Renz, Long Chen, Elahe Arani, and Oleg Sinavski. Simlingo: Vision-only closed-loop autonomous driving with language-action alignment. InConference on Computer V ision and P attern Recognition (CVPR), 2025. 12

  25. [33]

    Pei et al. Sun. Scalability in perception for autonomous driving: Waymo open dataset. InProceedings of the IEEE/CVF Conference on Computer V ision and P attern Recognition (CVPR), June 2020

  26. [34]

    FRIT: Using causal importance to improve chain-of-thought faithfulness

    Anand Swaroop, Akshat Nallani, Saksham Uboweja, Adiliia Uzdenova, Michael Nguyen, Kevin Zhu, Sunishchal Dev, Ashwinee Panda, V asu Sharma, and Maheep Chaudhary. FRIT: Using causal importance to improve chain-of-thought faithfulness. InFirst W orkshop on F oundations of Reasoni...

  27. [35]

    P . Tang, S. Xie, B. Sun, B. Huang, K. Luo, H. Y ang, W. Jin, and J. Wang. Mind to hand: Purposeful robotic control via embodied reasoning.arXiv preprint, arXiv:2512.08580, 2025

  28. [36]

    Gemini robotics: Bringing ai into the physical world.ArXiv, abs/2503.20020,

    Gemini Robotics Team. Gemini robotics: Bringing ai into the physical world.ArXiv, abs/2503.20020,

  29. [37]

    URLhttps://api.semanticscholar.org/CorpusID:277322650

  30. [38]

    Turpin, J

    M. Turpin, J. Michael, E. Perez, and S. R. Bowman. Language models don’t always say what they think: unfaithful explanations in chain-of-thought prompting. InConf. on Neural Information Processing Systems, New Orleans, LA, USA, 2023

  31. [39]

    Wei et al

    J. Wei et al. Chain-of-thought prompting elicits reasoning in large language models.arXiv preprint, arXiv:2201.11903, 2022

  32. [40]

    Do what you say: Steering vision-language-action models via runtime reasoning-action alignment verification.2026 IEEE International Conference on Robotics & Automation (ICRA), 2026

    Yilin Wu, Anqi Li, Tucker Hermans, Fabio Ramos, Andrea Bajcsy, and Claudia P˜AˇSrez-D’Arpino. Do what you say: Steering vision-language-action models via runtime reasoning-action alignment verification.2026 IEEE International Conference on Robotics & Automation (ICRA), 2026

  33. [41]

    Xu et al

    R. Xu et al. WOD-E2E: Waymo Open Dataset for end-to-end driving in challenging long-tail scenarios.arXiv preprint, arXiv:2510.26125, 2025

  34. [42]

    RLPR: Extrapolating RL VR to general domains without verifiers, 2026

    Tianyu Y u, Bo Ji, Shouli Wang, Shu Y ao, Zefan Wang, RuanLiqing, Kaidong Zhang, Ganqu Cui, Ning Ding, Y uan Y ao, Zhiyuan Liu, Maosong Sun, and Tat-Seng Chua. RLPR: Extrapolating RL VR to general domains without verifiers, 2026. URLhttps://openreview.net/forum?id= T03kNBYq81

  35. [43]

    Zaman and S

    K. Zaman and S. Srivastava. A causal lens for evaluating faithfulness metrics. InProceedings 2025 Conf. on Empirical Methods in Natural Language Processing, pages 29425–29449, 2025

  36. [44]

    Zawalski, W

    M. Zawalski, W. Chen, K. Pertsch, O. Mees, C. Finn, and S. Levine. Robotic control via embodied chain-of-thought reasoning.arXiv preprint, arXiv:2407.08693, 2024

  37. [45]

    X. Zhou, X. Han, F. Y ang, Y . Ma, V . Tresp, and A. C. Knoll. OpenDriveVLA: Towards end-to-end autonomous driving with large vision language action model. InProc. AAAI Conference on Artificial Intelligence, volume 40, pages 13782–13790, 2026. doi: 10.1609/aaai.v40i16.38386

  38. [46]

    Autovla: A vision-language-action model for end-to-end autonomous driving with adaptive reasoning and reinforcement fine-tuning.arXiv preprint arXiv:2506.13757, 2025

    Zewei Zhou, Tianhui Cai, Y un Zhao, Seth Z.and Zhang, Zhiyu Huang, Bolei Zhou, and Jiaqi Ma. Autovla: A vision-language-action model for end-to-end autonomous driving with adaptive reasoning and reinforcement fine-tuning.arXiv preprint arXiv:2506.13757, 2025. 13 8 Appendix Thi...

  39. [47]

    Stop for static constraints Trigger: a stop line, red signal, school-zone rule, or rail crossing ahead requires the vehicle to reach and hold zero speed

  40. [48]

    Yield (agent right-of-way) Trigger: the vehicle slows or stops to concede priority to a dynamic agent --- and no static control point is the primary cause

  41. [49]

    Lead obstacle following Trigger: a lead vehicle is present in the ego lane and the trajectory reflects time-gap management to that specific vehicle

  42. [50]

    Gap-searching (for LC/merge/zipper) Trigger: speed is being adjusted to open a gap for an imminent lateral maneuver

  43. [51]

    Acceleration for passing/overtaking Trigger: speed is increasing to pass a slower lead, with an associated lateral plan already in progress

  44. [52]

    Speed adaptation (road events) Trigger: speed is adjusted for a road geometry feature --- curve, grade, roundabout, ramp, or speed bump

  45. [53]

    Set speed tracking Trigger: none of the above apply; vehicle maintains or converges to a target cruise speed on an unconstrained road

  46. [54]

    LATERAL TAXONOMY (assign the MOST SPECIFIC bin): 1a

    None --- use only when fully stopped for a non-classifiable reason. LATERAL TAXONOMY (assign the MOST SPECIFIC bin): 1a. Turn left 1b. Turn right 2a. Lane change left 2b. Lane change right

  47. [55]

    Out-of-lane nudge left 4b

    Merge / Split (facility change) 4a. Out-of-lane nudge left 4b. Out-of-lane nudge right 5a. In-lane nudge left 5b. In-lane nudge right

  48. [56]

    Pull-over / curb approach

  49. [57]

    Lateral maneuver abort

  50. [58]

    Lane keeping & centering (default)

  51. [59]

    The vehicle is approximately 1.5--2 m wide

    None Compare the red polyline against visible lane markings to distinguish lane changes and nudges from lane keeping. The vehicle is approximately 1.5--2 m wide. Use the BEV plot as a fallback when the polyline is short or absent; on a curving road the BEV path sweeps sideways...

  52. [60]

    SURGICAL | change the minimum number of concepts necessary

  53. [61]

    Do not introduce obvious non-sequiturs or unrelated content

    PLAUSIBLE | the modified field must read as a natural, fluent piece of driving description when read in isolation. Do not introduce obvious non-sequiturs or unrelated content

  54. [62]

    red light

    SEMANTIC | the inconsistency must require reasoning about the meaning of the content to detect. Avoid keyword-level contradictions that a simple text matching rule would catch (e.g. do not write "red light" when the action is PROCEED if the original said "green light" | instea...

  55. [63]

    Do not introduce additional inconsistencies beyond the one specified

    SINGLE | corrupt exactly one edge as instructed. Do not introduce additional inconsistencies beyond the one specified. Return a JSON object with the fields specified in the task description. Do not include any preamble, explanation, or markdown fencing. LLM-based Corruption: U...

  56. [64]

    Is geometrically inconsistent with the trajectory described by the waypoints in Step 1

  57. [65]

    Is implausible or wrong for the scene context identified in Step 2

  58. [66]

    Differs from the source action on at least one axis

  59. [67]

    Both single-axis and two-axis flips are valid

    Uses the exact label strings from the vocabulary below --- no variants, no capitalisation changes. Both single-axis and two-axis flips are valid. Consider near-miss counterfactuals as well as clearly-different ones; don’t always default to the most dramatic flip. Rules: - Modi...

  60. [68]

    If any of{image to mj, mj to action, action to waypoints, mj to waypoints}is INCONSISTENT→overall = INCONSISTENT

  61. [69]

    Else if scene grounding is HALLUCINATED→overall = INCONSISTENT

  62. [70]

    " #! * * ( % # % ) ! %& # % # %!#' $

    Else→overall = CONSISTENT. Do NOT re-judge the overall holistically. Apply rules 1--3 mechanically. Confidence HIGH --- evidence is clear and unambiguous. MEDIUM --- plausible but some uncertainty. LOW --- genuine difficulty; a careful human might also be uncertain. Output for...

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.