Pith. sign in

REVIEW 2 major objections 6 minor 111 references

Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs

T0 review · 2 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Showing a driving VLM the logged future trajectory during reasoning degrades causal faithfulness and roughly doubles hallucinations; deferring trajectory exposure until after a scene-derived decision restores faithful reasoning, improves…

desk verdict A solid empirical study of a real supervision failure in AD VLM CoT training; the anchoring-bias result is well-supported, but the rubric reward is a genuine loophole the paper doesn't fully close. read the letter →

arxiv 2608.01755 v2 pith:SMLD5RD7 submitted 2026-08-03 cs.AI

Zixuan Huang , Yang Zhou , Kaixuan Wang , Guli Zhang , Hongyan Xie , Yakun Zhu , Hao Geng , Xiaozhi Chen
show 2 more authors
Yikun Ban Deqing Wang
This is my paper
classification cs.AI
keywords trajectoryanchoringbiasdeferredexposureautonomousdrivingchain-of-thoughtreinforcementlearningwithverifiablerewardsvision-languagemodelmultiple-choicebenchmarkrubricreward
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

This paper argues that a standard practice in training driving vision-language models—showing the teacher the logged ground-truth future trajectory while it writes its chain-of-thought—teaches the model to rationalize a known answer rather than reason from the scene. The authors call this trajectory anchoring bias and show, in a human-scored study, that it lowers causal faithfulness and roughly doubles severe hallucinations. They propose a verifiable benchmark, AD-MCQ, that turns planning into selection among explicit candidate trajectories, and a training procedure, DEFT-RLVR, that hides the candidates until after the policy has committed to a scene-derived decision, then uses exact trajectory correctness plus rubric-based process rewards. On two VLM backbones, DEFT-RLVR improves AD reasoning accuracy and reasoning quality while preserving or slightly improving general visual capability, and the gains transfer to an out-of-domain driving dataset.

What carries the argument

The load-bearing object is the two-turn DEFT interaction built on the AD-MCQ task. AD-MCQ discretizes logged ego futures into a K=8192-prototype codebook via K-means and constructs six-option questions in which the oracle is the nearest prototype to the logged future and distractors are similarity-bounded hard negatives. DEFT's Turn 1 asks the policy to produce causal reasoning and a high-level decision from scene context alone; Turn 2 then reveals the candidate trajectories and asks for a binding match to the committed decision. Both turns are optimized jointly with GRPO on a product reward: exact-choice correctness gates a per-question rubric reward, where rubrics are generated offline by a vision-language model from the same frames and graded online by a text-only judge that never sees images, candidates, or the ground-truth trajectory.

What would settle it

Run the same controlled human study on 100 causally challenging scenes with a different strong teacher model, such as a non-Qwen VLM: if GT-conditioned chains of thought do not show significantly more severe hallucinations than candidate-blind ones, or if human raters do not rate the candidate-blind CoTs as more causally faithful, the trajectory anchoring bias claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that future-trajectory information, when available before reasoning, acts as an anchor that drives post hoc rationalization in driving VLMs: teachers produce action-consistent but causally unfaithful chains of thought and invent scene cues to justify the revealed outcome. The discovery is that exposing the trajectory after a decision has been committed—deferred exposure—converts it from a reasoning premise into a verification target, and that combining this ordering with an exact multiple-choice verifier and rubric-based process supervision yields better driving decisions and reasoning quality than either joint exposure or outcome-only rewards. The authors demonstrate this through a controlled human study, ablation comparisons against joint-exposure and online-rubric variants, and cross-domain evaluation on nuScenes.

Load-bearing premise

The argument assumes that a text-only judge, checking a chain of thought against an offline-generated per-question rubric, can reliably verify whether that reasoning is grounded in the scene; if the rubric criteria omit or misstate the visible evidence, or the judge can be gamed by style, the process reward can reinforce non-causal reasoning even when final choices are correct.

Editorial extensions

If this is right

  • If the anchoring bias is general, current CoT-annotation pipelines for AD that condition on logged futures likely inject hallucinated causal evidence into supervision; DEFT-RLVR offers a drop-in alternative that starts from base VLMs.
  • Deferred exposure plus exact trajectory selection gives a verifiable RL reward without forcing the VLM to generate continuous coordinates, avoiding the trajectory-error and capability-forgetting bottlenecks of direct trajectory-token prediction.
  • Rubric-gated process supervision, applied only to correct choices, improves reasoning fidelity scores (CFS) and decision consistency (HLD) while adding only about 0.5% training overhead over correctness-only reward.
  • The learned scene-to-decision reasoning transfers to an out-of-domain nuScenes set, suggesting the paradigm teaches transferable reasoning rather than dataset-specific cues.

Reading between the lines

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

  • The deferred-exposure principle may generalize beyond driving: any VLM task where the answer is known during rationale annotation, such as VQA with ground-truth answers or logged tool-use executions, could suffer analogous anchoring, and the two-turn verify-after-commit format could apply whenever the answer space is enumerable as candidates.
  • AD-MCQ's controlled distractor construction gives a difficulty dial that could serve as a testbed for studying whether RLVR shortcuts recur as candidates become more similar; the paper shows gains hold across twenty constructions, but the hardest regimes still compress accuracy.
  • A testable extension is to check whether the text-only rubric judge can be gamed by style—for example, verbose but ill-grounded chains of thought that mention rubric cues—by comparing its reward against an image-conditioned judge; the paper's own control variant suggests the text-only scheme behaves differently, but the relative strictness is not fully characterized.
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

2 major / 6 minor

Summary. The paper studies how chain-of-thought supervision for autonomous-driving VLMs should be generated. It first provides a human-scored comparison (Appendix B) showing that exposing the logged ground-truth future trajectory to the teacher leads to lower causal faithfulness and more hallucinations than a causal-planning prompt that hides the future. It then proposes AD-MCQ, a benchmark that turns trajectory planning into selection among decoded waypoint candidates from a K-means codebook, and DEFT-RLVR, a two-turn interaction in which the policy first commits to a scene-derived high-level decision, then sees the candidates and grounds its decision in one of them. Training uses GRPO with an exact-choice reward multiplied by a process reward computed by a text-only judge against an offline, per-question VLM-generated rubric. Experiments on two backbones report gains in AD-MCQ accuracy, causal-faithfulness score (CFS), and high-level-decision consistency (HLD), with preserved or improved general visual benchmarks, plus OOD transfer to nuScenes and robustness across candidate constructions.

Significance. The paper has real strengths: the anchoring-bias hypothesis is tested with a controlled paired human study; AD-MCQ provides an exactly verifiable interface for trajectory-level planning; the DEFT/JEFT ablation cleanly isolates candidate-exposure order; the OOD and candidate-construction robustness experiments are extensive; and the release of code, model, and dataset supports reproducibility. The human validation of the CFS and HLD judges (Appendix E) and the transparency about training details (Algorithm 3, runtime accounting) are also commendable. If the rubric-reward verification chain is secured, the work would be a useful contribution to verifiable reasoning for AD VLMs. The main unresolved risk is that the process reward R_RUB (Eq. 10-11) is not itself verified against the scene, so the claim that rubric supervision suppresses hallucination may be weaker than the experiments suggest.

major comments (2)
  1. [Section 4.3, Eqs. (8)-(11), Appendix F.2-F.3] The central claimed advantage of DEFT-RLVR over correctness-only RLVR is that R_RUB steers reasoning toward causal faithfulness and suppresses hallucination (Section 5.4). This claim rests on two unverified links: the offline rubric generator G never fabricates scene facts, and the text-only judge J can detect grounding failures from text alone. G is a VLM prompted with the same kind of frames that already produce hallucinated CoTs in Fig. 1, and the history-only constraint is a prompt instruction, not a verified property. J is explicitly instructed to treat every cue asserted by the CoT as a textual claim and check it only against the rubric text (Appendix F.3); if a criterion encodes a hallucinated scene fact, any CoT that repeats it receives full credit. Because R_MCQ (Eq. 7) gates but does not penalize fabricated evidence, DEFT-RLVR can in principle be optimized to parrot rubric criteria while remaining causally unfaithful. The paper should either (a) report a manual audit of the generated rubrics, (b) show that R_RUB scores correlate with human or image-conditioned CFS judgments during training, or (c) demonstrate robustness to deliberately injected false rubric criteria. Without one of these, the reported CFS gains over DEFT+RLVR(R_MCQ) and the RGEN variant may partly reflect rubric-criteria matching rather than genuine scene grounding.
  2. [Table 2, Table 3, Appendix H] All main results are reported as means, and although the text states that each evaluation is run eight times, no variance or confidence intervals are given. This matters for load-bearing comparisons: for example, on Qwen3-VL-8B, DEFT-RLVR ACC is 77.9 vs. 75.2 for DEFT+RLVR(R_MCQ R_GEN), and on Qwen3.5-4B the CFS values are 0.822 vs. 0.819; the general-visual averages differ by only a few points. The claim that DEFT-RLVR preserves or even enhances general visual capability requires knowing the spread of these numbers. Please report standard errors or confidence intervals, or per-run values, for the AD-specific metrics and the general-capability aggregates.
minor comments (6)
  1. [Appendix B.2] The human study reports aggregate percentages and scores but no inter-annotator agreement; please report Cohen's kappa (or an equivalent reliability measure) for the 1,600 dimension-level ratings and for the pairwise preference judgments.
  2. [Section 5.1, Table 2] The rows marked + DEFT for both base models show dashes for CFS and HLD; clarify whether these metrics were intentionally omitted for training-free DEFT or whether the model produced no parseable Turn-1 output.
  3. [Appendix D.3, Algorithm 3] The text says that when grading fails after retries the method falls back to the exact-correctness reward, and Algorithm 3 sets R_RUB to 1 in that case; please report how often this fallback occurs, since a high failure rate would change the effective reward distribution.
  4. [Appendix C, Table 5] The caption states that per-cell ADE standard deviation is at most 0.005, but it does not define what a cell is or over how many seeds it is computed; please specify the number of K-means initializations used for the reported aggregate values.
  5. [Figure 6] The training-dynamics discussion relies on qualitative statements such as by far the largest policy drift; please add quantitative summaries (for example, final policy drift or KL values) so the reader can verify the claimed differences.
  6. [Section 2 / Appendix B.1] Please state explicitly whether the 100 scenes used in the anchoring-bias study are disjoint from the AD-MCQ-500 Test split, and whether the two annotators had professional driving or autonomous-driving domain expertise.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: DEFT-RLVR's deferred-exposure design uses the logged future only as a post-decision verification target, and the rubric reward is not fitted to the reported evaluation metrics.

full rationale

The paper's derivation chain is self-contained against external benchmarks. The central empirical claim—trajectory anchoring bias—is established by a controlled human study (Appendix B.2) that varies only the presence of the GT trajectory block while holding the task, model, and decoding fixed; the reported differences in CFS, hallucination rate, and pairwise preference are measured, not built into the definitions of the compared settings. In the DEFT-RLVR training loop, the oracle MCQ answer enters only through the exact-choice gate R_MCQ (Eq. 7) after the policy has committed to a candidate-blind Turn-1 decision (Eq. 5); the candidates are revealed only in Turn 2 (Eq. 6). The rubric reward R_RUB (Eqs. 8-11) is generated offline from scene frames without the GT trajectory, and the text-only judge evaluates the normalized Turn-1 trace against those rubric criteria without access to frames, candidates, or the oracle. No parameter is fitted to CFS or HLD and then reported as a prediction: CFS and HLD are scored by a separate, larger judge (Qwen3.5-397B-A17B) whose agreement with humans is validated in Appendix E, and they are not optimized directly. The cited self-references ([21,22,23,81]) appear only as related-work context and are not load-bearing for the main claims; no uniqueness theorem or ansatz is imported from the authors' prior work. The skeptic concern that the offline rubric generator might hallucinate scene facts is a validity and robustness risk, not a circularity by construction, because the rubric is not derived from the target metric it is used to improve.

Assumptions & free parameters 6 free parameters · 7 assumptions · 0 invented entities

The methods do not hide fitted constants as predictions; the free parameters are benchmark and training design choices. The load-bearing assumptions are that demonstrated trajectories define correctness, that the codebook preserves planning distinctions, and that LLM judges and rubrics faithfully measure causal faithfulness. No physical entities are introduced.

free parameters (6)
  • Codebook size K = 8192
    Selected from reconstruction fidelity versus utilization trade-off (Appendix C.5); determines oracle quantization error and candidate separation.
  • Clustering corpus size N = 489,042
    Chosen because larger N closes the out-of-sample gap (Table 6); affects quantization quality underlying AD-MCQ.
  • Candidate count M = 6
    Number of options per AD-MCQ instance; ACC and difficulty depend on M.
  • Hard-negative similarity bounds (rho_min, rho_max) = 0.30, 0.85 train; 0.92 Dev/Test
    Defines the distractor pool in Eq. 3 and Eq. 12; bounds are hand-selected and control benchmark difficulty.
  • Scale-match tolerance = 0.15 relative endpoint displacement
    Used in Dev/Test distractor selection to force scale-matched negatives that are not trivially separable by endpoint distance.
  • Rubric criterion weights = positive integers 1-10 per criterion
    Weights in Eq. 10 are generated by the rubric model and determine the process reward; no sensitivity analysis is provided.
assumptions (7)
  • domain assumption The logged future trajectory is a valid oracle for driving decisions, and the quantized nearest prototype is the single correct AD-MCQ answer.
    Section 3 and Appendix D.1 define the oracle this way; the paper acknowledges it is not the unique safe trajectory, so benchmark correctness is relative to demonstrated behavior.
  • domain assumption The K-means trajectory codebook at K=8192 preserves planning-relevant distinctions in braking, speed, and lateral geometry.
    Appendix C argues reconstruction fidelity and utilization; if the codebook is too coarse, distractors may not represent meaningfully different plans.
  • domain assumption Human-rated causal faithfulness dimensions (grounding, no-hallucination, specificity, coherence) capture reasoning quality, and the automatic CFS judge agrees with humans adequately.
    Appendix E reports Spearman 0.78 and 92% HLD agreement on 200 pooled samples; the entire CFS metric rests on this transfer.
  • ad hoc to paper The text-only rubric grader can determine whether a CoT contains scene-grounded evidence solely from criteria text and CoT text, with no images or answers.
    Section 4.3 and Appendix F.3; this assumption is introduced by the proposed method and is not externally validated.
  • ad hoc to paper Offline-generated rubrics from Qwen3.6-35B-A3B are accurate, complete, and unbiased regarding the scene.
    Eq. 8 and Appendix F.2; if criteria hallucinate or miss cues, R_RUB can reward unfaithful traces.
  • domain assumption Waymo Open E2E plus an internal corpus provides enough coverage of hard causal driving scenes for generalization claims.
    Section 5.1 and Appendix D.1; the internal corpus is unreleased, and hard-causal Dev/Test are curated, so the target population may not be fully represented.
  • standard math Euclidean waypoint distance is an appropriate trajectory similarity metric for selecting distractors and evaluating correctness.
    Eq. 1-2 and Table 10; this is a reasonable metric but a design choice, not physically forced.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs." pith.science (2026). https://pith.science/paper/SMLD5RD7

@misc{pith2026260801755,
  author       = {Pith},
  title        = {Pith review of: Deferred Exposure of Future Trajectories for Verifiable Reasoning in Autonomous Driving VLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SMLD5RD7}},
  note         = {Machine review of arXiv:2608.01755}
}
read the original abstract

Recent Vision-Language-Action (VLA) models for autonomous driving (AD) increasingly utilize chain-of-thought (CoT) supervision to enhance the reasoning capabilities of their Vision-Language Model (VLM) components, yet existing annotation pipelines commonly expose the teacher model to the logged ground-truth (GT) future trajectory. We empirically show that this induces trajectory anchoring bias: teacher models rationalize the revealed outcome rather than infer a decision from scene evidence, producing less causally faithful CoTs and substantially more severe hallucinations, especially in causally challenging scenes. Removing the GT trajectory eliminates this shortcut, but open-ended trajectory generation entangles high-level decision-making with precise geometric synthesis and low-level dynamics. To make trajectory-level driving decisions verifiable without requiring open-ended trajectory synthesis, we introduce Autonomous-Driving Multiple-Choice Question (AD-MCQ), which casts planning as selection among explicit trajectory candidates. Taking this a step further, we propose Deferred Exposure of Future Trajectories for RLVR (DEFT-RLVR) to transform future trajectories from pre-decision anchors into post-decision verification targets. Experimental results show that DEFT-RLVR improves AD reasoning while preserving or even enhancing general visual capabilities. With VLM-only inference and controllable difficulty through candidate construction, AD-MCQ provides a flexible, scalable, and extensible foundation for future research on verifiable AD reasoning.

Figures

Figures reproduced from arXiv: 2608.01755 by the authors.

Figure 9
Figure 9. Complete causal-planning chat template. The teacher receives the scene history and ego state but no logged future. Per-scene inputs are shown as variables. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_9.png] view at source ↗
Figure 10
Figure 10. Complete GT-conditioned chat template. It is identical to [PITH_FULL_IMAGE:figures/full_fig_p022_10.png] view at source ↗
Figure 11
Figure 11. Turn-1 policy chat template. The policy observes the scene, ego state, and navigation command, but no candidate trajectory. The role-separated panel preserves the production message order, and variables in braces are instantiated per scene. Turn-2 Candidate-Trajectory Matching Prompt System Prompt Continuation of Turn 1; no new system message. User Message — Turn 2: Option Match PART 2 — OPTION MATCH (only now use t… view at source ↗
Figures from the paper (4 more)
Figure 12
Figure 12. Figure 12: Turn-2 policy chat template. The environment reveals six deterministically shuffled candidate tra￾jectories only after the policy has committed to its first-turn plan. The parser uses the last FINAL_CHOICE field. 36 [PITH_FULL_IMAGE:figures/full_fig_p036_12.png]
Figure 14
Figure 14. Figure 14: Online text-only grader chat template. The grader returns one binary decision per stored criterion; the weighted present rate supplies the process score for an outcome-correct rollout. F.4 Image-Conditioned Rubric Reward for the Controlled Variant For DEFT + RLVR (RMC…
Figure 15
Figure 15. Figure 15: Image-conditioned four-axis rubric-judge template for the controlled variant (axis weights: 0.30/0.30/0.10/0.30). F.5 Joint-Exposure Policy Prompt (JEFT) [PITH_FULL_IMAGE:figures/full_fig_p039_15.png]
Figure 16
Figure 16. Figure 16: JEFT joint-exposure policy template with all candidate trajectories visible before reasoning. 40 [PITH_FULL_IMAGE:figures/full_fig_p040_16.png]

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

111 extracted references · 33 canonical work pages

  1. [1]

    Chain-of-thought reasoning in the wild is not always faithful.arXiv preprint arXiv:2503.08679, 2025

    Iván Arcuschin, Jett Janiak, Robert Krzyzanowski, Senthooran Rajamanoharan, Neel Nanda, and Arthur Conmy. Chain-of-thought reasoning in the wild is not always faithful.arXiv preprint arXiv:2503.08679, 2025

    arXiv 2025

  2. [2]

    Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    Shuai Bai, Yuxuan Cai, Ruizhe Chen, Keqin Chen, Xionghui Chen, Zesen Cheng, Lianghao Deng, Wei Ding, Chang Gao, Chunjiang Ge, et al. Qwen3-vl technical report.arXiv preprint arXiv:2511.21631, 2025

    arXiv 2025

  3. [3]

    A closer look at bias and chain-of-thought faith- fulness of large (vision) language models

    Sriram Balasubramanian, Samyadeep Basu, and Soheil Feizi. A closer look at bias and chain-of-thought faith- fulness of large (vision) language models. In Christos Christodoulopoulos, Tanmoy Chakraborty, Carolyn Rose, and Violet Peng, editors,Findings of the Association for Computational Linguistics: EMNLP 2025, pages 13406– 13439, Suzhou, China, November ...

    work page doi:10.18653/v1/2025.findings-emnlp.723.https://aclanthology.org/2025 2025

  4. [4]

    nuscenes: A multimodal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020

    2020

  5. [5]

    Answermatchingoutperforms multiple choice for language model evaluation.arXiv preprint arXiv:2507.02856, 2025

    NikhilChandak, ShashwatGoel, AmeyaPrabhu, MoritzHardt, andJonasGeiping. Answermatchingoutperforms multiple choice for language model evaluation.arXiv preprint arXiv:2507.02856, 2025. doi: 10.48550/arXiv.2507. 02856

    arXiv doi:10.48550/arxiv.2507.02856 2025

  6. [6]

    Reasoning models don’t always say what they think

    Yanda Chen, Joe Benton, Ansh Radhakrishnan, Jonathan Uesato, Carson Denison, John Schulman, Arushi Somani, Peter Hase, Misha Wagner, Fabien Roger, et al. Reasoning models don’t always say what they think. arXiv preprint arXiv:2505.05410, 2025

    arXiv 2025

  7. [7]

    Transfuser: Imitation with transformer-based sensor fusion for autonomous driving.IEEE transactions on pattern analysis and machine intelligence, 45(11):12878–12895, 2022

    Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driving.IEEE transactions on pattern analysis and machine intelligence, 45(11):12878–12895, 2022

    2022

  8. [8]

    Embspatial-bench: Benchmarking spatial understanding for embodied tasks with large vision-language models

    Mengfei Du, Binhao Wu, Zejun Li, Xuan-Jing Huang, and Zhongyu Wei. Embspatial-bench: Benchmarking spatial understanding for embodied tasks with large vision-language models. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 346–355, 2024

    2024

Show all 111 references
  1. [9]

    Vlmevalkit: An open-source toolkit for evaluating large multi-modality models

    Haodong Duan, Xinyu Fang, Junming Yang, Xiangyu Zhao, Zerun Ma, Yuxuan Qiao, Mo Li, Tianhao Liang, Lin Zhu, Amit Agarwal, et al. Vlmevalkit: An open-source toolkit for evaluating large multi-modality models. arXiv preprint arXiv:2407.11691, 2024

    2024 arXiv

  2. [10]

    Libero-plus: In-depth robustness analysis of vision-language-action models.arXiv preprint arXiv:2510.13626, 2025

    Senyu Fei, Siyin Wang, Junhao Shi, Zihao Dai, Jikun Cai, Pengfang Qian, Li Ji, Xinzhe He, Shiduo Zhang, Zhaoye Fei, et al. Libero-plus: In-depth robustness analysis of vision-language-action models.arXiv preprint arXiv:2510.13626, 2025

    arXiv

  3. [11]

    Srft: A single-stage method with supervised and reinforcement fine-tuning for reasoning.arXiv preprint arXiv:2506.19767, 2025

    Yuqian Fu, Tinghong Chen, Jiajun Chai, Xihuai Wang, Songjun Tu, Guojun Yin, Wei Lin, Qichao Zhang, Yuanheng Zhu, and Dongbin Zhao. Srft: A single-stage method with supervised and reinforcement fine-tuning for reasoning.arXiv preprint arXiv:2506.19767, 2025

    2025 arXiv

  4. [12]

    Acceleratingstructuredchain-of-thoughtinautonomousvehicles.arXiv preprint arXiv:2602.02864, 2026

    Yi Gu, Yan Wang, Yuxiao Chen, Yurong You, Wenjie Luo, Yue Wang, Wenhao Ding, Boyi Li, Heng Yang, Boris Ivanovic, etal. Acceleratingstructuredchain-of-thoughtinautonomousvehicles.arXiv preprint arXiv:2602.02864, 2026

    2026

  5. [13]

    Anisha Gunjal, Anthony Wang, Elaine Lau, Vaskar Nath, Bing Liu, and Sean M. Hendryx. Rubrics as rewards: Reinforcement learning beyond verifiable domains.arXiv preprint arXiv:2507.17746, 2025

    2025 arXiv

  6. [14]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    2025 arXiv

  7. [15]

    Rethinking multiple-choice questions for rlvr: Unlocking potential via distractor design

    Xu Guo, Qiming Ge, Jian Tong, Kedi Chen, Jin Zhang, Xiaogui Yang, Xuan Gao, Haijun Lv, Zhihui Lu, Yicheng Zou, et al. Rethinking multiple-choice questions for rlvr: Unlocking potential via distractor design. InFindings of the Association for Computational Linguistics: ACL 2026...

    2026

  8. [16]

    Llm-rubric: A multidi- mensional, calibrated approach to automated evaluation of natural language texts

    Helia Hashemi, Jason Eisner, Corby Rosset, Benjamin Van Durme, and Chris Kedzie. Llm-rubric: A multidi- mensional, calibrated approach to automated evaluation of natural language texts. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (...

    2024

  9. [17]

    Planning-oriented autonomous driving

    Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17853–17862, 2023

    2023

  10. [18]

    Gameformer: Game-theoretic modeling and learning of transformer- based interactive prediction and planning for autonomous driving

    Zhiyu Huang, Haochen Liu, and Chen Lv. Gameformer: Game-theoretic modeling and learning of transformer- based interactive prediction and planning for autonomous driving. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 3903–3913, 2023

    2023

  11. [19]

    Differentiable integrated motion prediction and planning with learnable cost function for autonomous driving.IEEE transactions on neural networks and learning systems, 2023

    Zhiyu Huang, Haochen Liu, Jingda Wu, and Chen Lv. Differentiable integrated motion prediction and planning with learnable cost function for autonomous driving.IEEE transactions on neural networks and learning systems, 2023

    2023

  12. [20]

    Gen-drive: Enhancing diffusion generative driving policies with reward modeling and reinforcement learning fine-tuning.arXiv preprint arXiv:2410.05582, 2024

    Zhiyu Huang, Xinshuo Weng, Maximilian Igl, Yuxiao Chen, Yulong Cao, Boris Ivanovic, Marco Pavone, and Chen Lv. Gen-drive: Enhancing diffusion generative driving policies with reward modeling and reinforcement learning fine-tuning.arXiv preprint arXiv:2410.05582, 2024

    arXiv

  13. [21]

    Adaptive sample scheduling for direct preference optimization.arXiv preprint arXiv:2506.17252, 2025

    Zixuan Huang, Yikun Ban, Lean Fu, Xiaojie Li, Zhongxiang Dai, Jianxin Li, and Deqing Wang. Adaptive sample scheduling for direct preference optimization.arXiv preprint arXiv:2506.17252, 2025

    2025

  14. [22]

    Does your reasoning model implicitly know when to stop thinking?arXiv preprint arXiv:2602.08354, 2026

    Zixuan Huang, Xin Xia, Yuxi Ren, Jianbin Zheng, Xuanda Wang, Zhixia Zhang, Hongyan Xie, Songshi Liang, Zehao Chen, Xuefeng Xiao, et al. Does your reasoning model implicitly know when to stop thinking?arXiv preprint arXiv:2602.08354, 2026

    2026 arXiv

  15. [23]

    Real-time aligned reward model beyond semantics

    Zixuan Huang, Xin Xia, Yuxi Ren, Jianbin Zheng, Xuefeng Xiao, Hongyan Xie, Huaqiu Li, Songshi Liang, Zhongxiang Dai, Fuzhen Zhuang, Jianxin Li, Yikun Ban, and Deqing Wang. Real-time aligned reward model beyond semantics. 2026.https://api.semanticscholar.org/CorpusID:285240754

    2026

  16. [24]

    Emma: End-to-end multimodal model for autonomous driving.arXiv preprint arXiv:2410.23262, 2024

    Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, et al. Emma: End-to-end multimodal model for autonomous driving.arXiv preprint arXiv:2410.23262, 2024

    arXiv

  17. [25]

    Drivelmm-o1: A step-by-step reasoning dataset and large multimodal model for driving scenario understanding

    Ayesha Ishaq, Jean Lahoud, Ketan More, Omkar Thawakar, Ritesh Thawkar, Dinura Dissanayake, Noor Ahsan, Yuhao Li, Fahad Shahbaz Khan, Hisham Cholakkal, et al. Drivelmm-o1: A step-by-step reasoning dataset and large multimodal model for driving scenario understanding. In2025 IEE...

    2025

  18. [26]

    Vad: Vectorized scene representation for efficient autonomous driving

    Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. InProceed- ings of the IEEE/CVF International Conference on Computer Vision, p...

    2023

  19. [27]

    Senna: Bridging large vision-language models and end-to-end autonomous driving.arXiv preprint arXiv:2410.22313, 2024

    Bo Jiang, Shaoyu Chen, Bencheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving.arXiv preprint arXiv:2410.22313, 2024

    arXiv

  20. [28]

    Viewspatial-bench: Evaluating multi-perspective spatial localization in vision-language models.arXiv preprint arXiv:2505.21500, 2025

    Dingming Li, Hongxing Li, Zixuan Wang, Yuchen Yan, Hang Zhang, Siqi Chen, Guiyang Hou, Shengpei Jiang, Wenqi Zhang, Yongliang Shen, et al. Viewspatial-bench: Evaluating multi-perspective spatial localization in vision-language models.arXiv preprint arXiv:2505.21500, 2025

    2025

  21. [29]

    Discrete diffusion for reflective vision-language-action models in autonomous driving.arXiv preprint arXiv:2509.20109, 2025

    Pengxiang Li, Yinan Zheng, Yue Wang, Huimin Wang, Hang Zhao, Jingjing Liu, Xianyuan Zhan, Kun Zhan, and Xianpeng Lang. Discrete diffusion for reflective vision-language-action models in autonomous driving.arXiv preprint arXiv:2509.20109, 2025

    2025

  22. [30]

    Womd-reasoning: A large-scale dataset for interaction reasoning in driving, 2025.https://arxiv.org/abs/2407.04281

    YihengLi, CunxinFan, ChongjianGe, ZhihaoZhao, ChenranLi, ChenfengXu, HuaxiuYao, MasayoshiTomizuka, Bolei Zhou, Chen Tang, Mingyu Ding, and Wei Zhan. Womd-reasoning: A large-scale dataset for interaction reasoning in driving, 2025.https://arxiv.org/abs/2407.04281

    2025 arXiv

  23. [31]

    Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving.arXiv preprint arXiv:2411.15139, 2024

    Bencheng Liao, Shaoyu Chen, Haoran Yin, Bo Jiang, Cheng Wang, Sixu Yan, Xinbang Zhang, Xiangyu Li, Ying Zhang, Qian Zhang, et al. Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving.arXiv preprint arXiv:2411.15139, 2024

    2024 arXiv

  24. [32]

    Toolanchor: Anchoring counterfactual context to boost agentic tool-use capability.arXiv preprint arXiv:2607.14145, 2026

    Weiting Liu, Jieyi Bi, Wanqi Zhou, Jianfeng Feng, Yining Ma, Ai Han, and Wenlian Lu. Toolanchor: Anchoring counterfactual context to boost agentic tool-use capability.arXiv preprint arXiv:2607.14145, 2026

    2026 arXiv

  25. [33]

    Automated optimization modeling via a localizable error-driven perspective.arXiv preprint arXiv:2602.11164, 2026

    Weiting Liu, Han Wu, Yufei Kuang, Xiongwei Han, Tao Zhong, Jianfeng Feng, and Wenlian Lu. Automated optimization modeling via a localizable error-driven perspective.arXiv preprint arXiv:2602.11164, 2026. 13

    2026

  26. [34]

    Styletai- lor: Towards personalized fashion styling via hierarchical negative feedback, 2025.https://arxiv.org/abs/2508

    Hongbo Ma, Fei Shen, Hongbin Xu, Xiaoce Wang, Gang Xu, Jinkai Zheng, Liangqiong Qu, and Ming Li. Styletai- lor: Towards personalized fashion styling via hierarchical negative feedback, 2025.https://arxiv.org/abs/2508. 06555

    2025

  27. [35]

    Weijian Ma, Ruoxin Chen, Keyue Zhang, Shuang Wu, and Shouhong Ding. Instruct where the model fails: Generative data augmentation via guided self-contrastive fine-tuning.Proceedings of the AAAI Conference on Artificial Intelligence, 39(6):5991–5999, Apr. 2025. doi: 10.1609/aaai...

    2025 doi

  28. [36]

    Thinking with blueprints: Assisting vision-language models in spatial reasoning via structured object representation, 2026.https://arxiv

    Weijian Ma, Shizhao Sun, Tianyu Yu, Ruiyu Wang, Tat-Seng Chua, and Jiang Bian. Thinking with blueprints: Assisting vision-language models in spatial reasoning via structured object representation, 2026.https://arxiv. org/abs/2601.01984

    2026

  29. [37]

    3dsr- bench: A comprehensive 3d spatial reasoning benchmark

    Wufei Ma, Haoyu Chen, Guofeng Zhang, Yu-Cheng Chou, Jieneng Chen, Celso de Melo, and Alan Yuille. 3dsr- bench: A comprehensive 3d spatial reasoning benchmark. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 6924–6934, 2025

    2025

  30. [38]

    Gpt-driver: Learning to drive with gpt.arXiv preprint arXiv:2310.01415, 2023

    Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt.arXiv preprint arXiv:2310.01415, 2023

    2023 arXiv

  31. [39]

    Lingoqa: Visual question answering for au- tonomous driving

    Ana-Maria Marcu, Long Chen, Jan Hünermann, Alice Karnsund, Benoit Hanotte, Prajwal Chidananda, Saurabh Nair, Vijay Badrinarayanan, Alex Kendall, Jamie Shotton, et al. Lingoqa: Visual question answering for au- tonomous driving. InEuropean Conference on Computer Vision, pages 2...

    2024

  32. [40]

    Reason2drive: Towards interpretable and chain-based reasoning for autonomous driving

    Ming Nie, Renyuan Peng, Chunwei Wang, Xinyue Cai, Jianhua Han, Hang Xu, and Li Zhang. Reason2drive: Towards interpretable and chain-based reasoning for autonomous driving. InEuropean Conference on Computer Vision, pages 292–308. Springer, 2024

    2024

  33. [41]

    Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sand- hini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:277...

    2022

  34. [42]

    Nuplanqa: A large-scale dataset and benchmark for multi-view driving scene understanding in multi-modal large language models.arXiv preprint arXiv:2503.12772, 2025

    Sung-Yeon Park, Can Cui, Yunsheng Ma, Ahmadreza Moradipari, Rohit Gupta, Kyungtae Han, and Ziran Wang. Nuplanqa: A large-scale dataset and benchmark for multi-view driving scene understanding in multi-modal large language models.arXiv preprint arXiv:2503.12772, 2025

    2025 arXiv

  35. [43]

    Fast: Efficient action tokenization for vision-language-action models.arXiv preprint arXiv:2501.09747, 2025

    Karl Pertsch, Kyle Stachowicz, Brian Ichter, Danny Driess, Suraj Nair, Quan Vuong, Oier Mees, Chelsea Finn, and Sergey Levine. Fast: Efficient action tokenization for vision-language-action models.arXiv preprint arXiv:2501.09747, 2025

    2025 arXiv

  36. [44]

    Trajeglish: Traffic modeling as next-token prediction.arXiv preprint arXiv:2312.04535, 2023

    Jonah Philion, Xue Bin Peng, and Sanja Fidler. Trajeglish: Traffic modeling as next-token prediction.arXiv preprint arXiv:2312.04535, 2023

    2023 arXiv

  37. [45]

    Nuscenes-qa: A multi-modal visual question answering benchmark for autonomous driving scenario

    Tianwen Qian, Jingjing Chen, Linhai Zhuo, Yang Jiao, and Yu-Gang Jiang. Nuscenes-qa: A multi-modal visual question answering benchmark for autonomous driving scenario. InProceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4542–4550, 2024

    2024

  38. [46]

    Qwen3.5: Towards native multimodal agents, February 2026.https://qwen.ai/blog?id=qwen3.5

    Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026.https://qwen.ai/blog?id=qwen3.5

    2026

  39. [47]

    Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Information Processing Systems, 36:53728–53741, 2023

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.Advances in Neural Information Processing Systems, 36:53728–53741, 2023

    2023

  40. [48]

    Autorubric: Unifying rubric-based llm evaluation.arXiv preprint arXiv:2603.00077, 2026

    Delip Rao and Chris Callison-Burch. Autorubric: Unifying rubric-based llm evaluation.arXiv preprint arXiv:2603.00077, 2026

    2026 arXiv

  41. [49]

    Online rubrics elicitation from pairwise comparisons.arXiv preprint arXiv:2510.07284, 2025

    MohammadHossein Rezaei, Robert Vacareanu, Zihao Wang, Clinton Wang, Bing Liu, Yunzhong He, and Afra Feyza Akyürek. Online rubrics elicitation from pairwise comparisons.arXiv preprint arXiv:2510.07284, 2025

    2025

  42. [50]

    Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms.arXiv preprint arXiv:1707.06347, 2017

    2017 arXiv

  43. [51]

    Deepseekmath: Pushing the limits of mathematical reasoning in open language models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Y Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024. 14

    2024 arXiv

  44. [52]

    Drivelm: Driving with graph visual question answering

    Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beißwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. InEuropean Conference on Computer Vision, pages 256–274. Springer, 2024

    2024

  45. [53]

    Robospatial: Teach- ing spatial understanding to 2d and 3d vision-language models for robotics.arXiv preprint arXiv:2411.16537, 2024

    ChanHeeSong, ValtsBlukis, JonathanTremblay, StephenTyree, YuSu, andStanBirchfield. Robospatial: Teach- ing spatial understanding to 2d and 3d vision-language models for robotics.arXiv preprint arXiv:2411.16537, 2024

    2024

  46. [54]

    Probing perceptual constancy in large vision-language models, 2026.https://arxiv.org/abs/2502.10273

    Haoran Sun, Bingyang Wang, Suyang Yu, Yijiang Li, Qingying Gao, Haiyun Lyu, Lianyu Huang, Zelong Hong, Jiahui Ge, Qianli Ma, Hang He, Yifan Zhou, Lingzi Guo, Lantao Mei, Maijunxian Wang, Dezhi Luo, and Hokin Deng. Probing perceptual constancy in large vision-language models, 2...

    2026

  47. [55]

    Sparsedrive: End-to-end autonomous driving via sparse scene representation.arXiv preprint arXiv:2405.19620, 2024

    Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Haoran Wu, and Sifa Zheng. Sparsedrive: End-to-end autonomous driving via sparse scene representation.arXiv preprint arXiv:2405.19620, 2024

    2024 arXiv

  48. [56]

    Flow matching-based autonomous driving planning with advanced interactive behavior modeling.Advances in Neural Information Processing Systems, 38:38310–38335, 2025

    Tianyi Tan, Yinan Zheng, Ruiming Liang, Zexu Wang, Kexin Zheng, Jinliang Zheng, Jianxiong Li, Xianyuan Zhan, and Jingjing Liu. Flow matching-based autonomous driving planning with advanced interactive behavior modeling.Advances in Neural Information Processing Systems, 38:3831...

    2025

  49. [57]

    Gemini robotics: Bringing ai into the physical world.arXiv preprint arXiv:2503.20020, 2025

    Gemini Robotics Team, Saminda Abeyruwan, Joshua Ainslie, Jean-Baptiste Alayrac, Montserrat Gonzalez Are- nas, Travis Armstrong, Ashwin Balakrishna, Robert Baruch, Maria Bauza, Michiel Blokzijl, et al. Gemini robotics: Bringing ai into the physical world.arXiv preprint arXiv:25...

    2025 arXiv

  50. [58]

    Nuscenes-spatialqa: A spatial understanding and reasoning benchmark for vision-language models in autonomous driving.arXiv preprint arXiv:2504.03164, 2025

    Kexin Tian, Jingrui Mao, Yunlong Zhang, Jiwan Jiang, Yang Zhou, and Zhengzhong Tu. Nuscenes-spatialqa: A spatial understanding and reasoning benchmark for vision-language models in autonomous driving.arXiv preprint arXiv:2504.03164, 2025

    2025 arXiv

  51. [59]

    Drivevlm: The convergence of autonomous driving and large vision-language models.arXiv preprint arXiv:2402.12289, 2024

    Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models.arXiv preprint arXiv:2402.12289, 2024

    2024 arXiv

  52. [60]

    Cambrian-1: A fully open, vision-centric exploration of multimodal llms

    Shengbang Tong, Ellis L Brown II, Penghao Wu, Sanghyun Woo, Adithya Jairam Iyer, Sai Charitha Akula, Shusheng Yang, Jihan Yang, Manoj Middepogu, Ziteng Wang, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. InThe Thirty-eighth Annual Conference o...

    2024

  53. [61]

    A very big video reasoning suite

    Maijunxian Wang, Ruisi Wang, Juyi Lin, Ran Ji, Thaddäus Wiedemer, Qingying Gao, Dezhi Luo, Yaoyao Qian, Lianyu Huang, Zelong Hong, Jiahui Ge, Qianli Ma, Hang He, et al. A very big video reasoning suite. In Proceedings of the 43rd International Conference on Machine Learning, 2...

    2026

  54. [62]

    Drivecot: Integrating chain-of-thought reasoning with end-to-end driving.arXiv preprint arXiv:2403.16996, 2024

    Tianqi Wang, Enze Xie, Ruihang Chu, Zhenguo Li, and Ping Luo. Drivecot: Integrating chain-of-thought reasoning with end-to-end driving.arXiv preprint arXiv:2403.16996, 2024

    2024 arXiv

  55. [63]

    Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving.arXiv preprint arXiv:2312.09245, 2023

    Wenhai Wang, Jiangwei Xie, ChuanYang Hu, Haoming Zou, Jianan Fan, Wenwen Tong, Yang Wen, Silei Wu, Hanming Deng, Zhiqi Li, et al. Drivemlm: Aligning multi-modal large language models with behavioral planning states for autonomous driving.arXiv preprint arXiv:2312.09245, 2023

    2023

  56. [64]

    Adˆ 2-bench: A hierar- chical cot benchmark for mllm in autonomous driving under adverse conditions.arXiv preprint arXiv:2506.09557, 2025

    Zhaoyang Wei, Chenhui Qiang, Bowen Jiang, Xumeng Han, Xuehui Yu, and Zhenjun Han. Adˆ 2-bench: A hierar- chical cot benchmark for mllm in autonomous driving under adverse conditions.arXiv preprint arXiv:2506.09557, 2025

    2025 arXiv

  57. [65]

    Separate the wheat from the chaff: A post-hoc approach to safety re-alignment for fine-tuned language models.arXiv preprint arXiv:2412.11041, 2024

    Di Wu, Xin Lu, Yanyan Zhao, and Bing Qin. Separate the wheat from the chaff: A post-hoc approach to safety re-alignment for fine-tuned language models.arXiv preprint arXiv:2412.11041, 2024

    2024 arXiv

  58. [66]

    Star-s: Improving safety alignment through self-taught reasoning on safety rules.arXiv preprint arXiv:2601.03537, 2026

    Di Wu, Yanyan Zhao, Xin Lu, Mingzhe Li, and Bing Qin. Star-s: Improving safety alignment through self-taught reasoning on safety rules.arXiv preprint arXiv:2601.03537, 2026

    2026

  59. [67]

    Smart: scalable multi-agent real-time motion generation via next-token prediction.Advances in Neural Information Processing Systems, 37:114048–114071, 2024

    Wei Wu, Xiaoxin Feng, Ziyan Gao, and Yuheng Kan. Smart: scalable multi-agent real-time motion generation via next-token prediction.Advances in Neural Information Processing Systems, 37:114048–114071, 2024

    2024

  60. [68]

    Preemptive answer “attacks” on chain-of-thought reasoning

    Rongwu Xu, Zehan Qi, and Wei Xu. Preemptive answer “attacks” on chain-of-thought reasoning. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar, editors,Findings of the Association for Computational Linguistics: ACL 2024, pages 14708–14726, Bangkok, Thailand, August 2024. Associa...

    2024 doi

  61. [69]

    Wod-e2e: Waymo open dataset for end-to-end driving in challenging long-tail scenarios

    Runsheng Xu, Hubert Lin, Wonseok Jeon, Hao Feng, Yuliang Zou, Liting Sun, John Gorman, Kate Tolstaya, Sarah Tang, Brandyn White, et al. Wod-e2e: Waymo open dataset for end-to-end driving in challenging long-tail scenarios. InProceedings of the IEEE/CVF Conference on Computer V...

    2026

  62. [70]

    Drivegpt4: Interpretable end-to-end autonomous driving via large language model.IEEE Robotics and Automation Letters, 2024

    Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model.IEEE Robotics and Automation Letters, 2024

    2024

  63. [71]

    More thinking, less seeing? assessing amplified hallucination in multimodal reasoning models.Advances in Neural Information Processing Systems, 38:82878–82905, 2025

    Zhongxing Xu, Chengzhi Liu, Qingyue Wei, Juncheng Wu, James Zou, Xin Wang, Yuyin Zhou, and Sheng Liu. More thinking, less seeing? assessing amplified hallucination in multimodal reasoning models.Advances in Neural Information Processing Systems, 38:82878–82905, 2025

    2025

  64. [72]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    2025 arXiv

  65. [73]

    Batched contextual reinforcement

    Bangji Yang, Hongbo Ma, Jiajun Fan, and Ge Liu. Batched contextual reinforcement. InForty-third International Conference on Machine Learning, 2026.https://openreview.net/forum?id=8Oc3Mx754M

    2026

  66. [74]

    Depth anything v2.Advances in Neural Information Processing Systems, 37:21875–21911, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2.Advances in Neural Information Processing Systems, 37:21875–21911, 2024

    2024

  67. [75]

    Mmsi-bench: A benchmark for multi-image spatial intelligence.arXiv preprint arXiv:2505.23764, 2025

    Sihan Yang, Runsen Xu, Yiman Xie, Sizhe Yang, Mo Li, Jingli Lin, Chenming Zhu, Xiaochen Chen, Haodong Duan, Xiangyu Yue, et al. Mmsi-bench: A benchmark for multi-image spatial intelligence.arXiv preprint arXiv:2505.23764, 2025

    2025 arXiv

  68. [76]

    Fooled by facts: quantifying anchoring bias through a large-scale experiment

    Taha Yasseri and Jannie Reher. Fooled by facts: quantifying anchoring bias through a large-scale experiment. Journal of Computational Social Science, 5(1):1001–1021, 2022

    2022

  69. [77]

    Vulca-bench: A multicultural vision-language benchmark for evaluating cultural understanding, 2026.https://arxiv.org/abs/2601.07986

    Haorui Yu, Diji Yang, Hang He, Fengrui Zhang, and Qiufeng Yi. Vulca-bench: A multicultural vision-language benchmark for evaluating cultural understanding, 2026.https://arxiv.org/abs/2601.07986

    2026

  70. [78]

    Promoting efficient reasoning with verifiable stepwise reward.Proceedings of the AAAI Conference on Artificial Intelligence, 40(41): 34530–34538, 2026

    Chuhuai Yue, Chengqi Dong, Yinan Gao, Hang He, Jiajun Chai, Wei Lin, and Guojun Yin. Promoting efficient reasoning with verifiable stepwise reward.Proceedings of the AAAI Conference on Artificial Intelligence, 40(41): 34530–34538, 2026. doi: 10.1609/aaai.v40i41.40752.https://d...

    2026 doi

  71. [79]

    Wisead: Knowledge augmented end-to-end autonomous driving with vision-language model.arXiv preprint arXiv:2412.09951, 2024

    Songyan Zhang, Wenhui Huang, Zihui Gao, Hao Chen, and Chen Lv. Wisead: Knowledge augmented end-to-end autonomous driving with vision-language model.arXiv preprint arXiv:2412.09951, 2024

    2024 arXiv

  72. [80]

    Closed-loop supervised fine-tuning of tokenized traffic models.arXiv preprint arXiv:2412.05334, 2024

    Zhejun Zhang, Peter Karkus, Maximilian Igl, Wenhao Ding, Yuxiao Chen, Boris Ivanovic, and Marco Pavone. Closed-loop supervised fine-tuning of tokenized traffic models.arXiv preprint arXiv:2412.05334, 2024

    2024 arXiv

  73. [81]

    Heterogeneous agent collaborative reinforcement learning.arXiv preprint arXiv:2603.02604, 2026

    Zhixia Zhang, Zixuan Huang, Xin Xia, Deqing Wang, Fuzhen Zhuang, Shuai Ma, Ning Ding, Yaodong Yang, Jianxin Li, and Yikun Ban. Heterogeneous agent collaborative reinforcement learning.arXiv preprint arXiv:2603.02604, 2026

    2026 arXiv

  74. [82]

    Cot-vla: Visual chain-of-thought reasoning for vision-language-action models.arXiv preprint arXiv:2503.22020, 2025

    Qingqing Zhao, Yao Lu, Moo Jin Kim, Zipeng Fu, Zhuoyang Zhang, Yecheng Wu, Zhaoshuo Li, Qianli Ma, Song Han, Chelsea Finn, et al. Cot-vla: Visual chain-of-thought reasoning for vision-language-action models.arXiv preprint arXiv:2503.22020, 2025

    2025 arXiv

  75. [83]

    Genad: Generative end-to-end autonomous driving

    Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. InEuropean Conference on Computer Vision, pages 87–104. Springer, 2024

    2024

  76. [84]

    Roborefer: Towards spatial referring with reasoning in vision-language models for robotics.Advances in Neural Information Processing Systems, 38:28404–28481, 2026

    Enshen Zhou, Jingkun An, Cheng Chi, Yi Han, Shanyu Rong, Chi Zhang, Pengwei Wang, Zhongyuan Wang, Tiejun Huang, Lu Sheng, et al. Roborefer: Towards spatial referring with reasoning in vision-language models for robotics.Advances in Neural Information Processing Systems, 38:284...

    2026

  77. [85]

    Opendrivevla: Towards end-to-end autonomous driving with large vision language action model.arXiv preprint arXiv:2503.23463, 2025

    Xingcheng Zhou, Xuyuan Han, Feng Yang, Yunpu Ma, and Alois C Knoll. Opendrivevla: Towards end-to-end autonomous driving with large vision language action model.arXiv preprint arXiv:2503.23463, 2025

    2025

  78. [86]

    Libero-pro: Towards robust and fair evaluation of vision-language-action models beyond memorization

    XueyangZhou, YangmingXu, GuiyaoTie, YongchaoChen, GuowenZhang, DuanfengChu, PanZhou, andLichao Sun. Libero-pro: Towards robust and fair evaluation of vision-language-action models beyond memorization. arXiv preprint arXiv:2510.03827, 2025

    arXiv

  79. [87]

    Autovla: A vision- language-action model for end-to-end autonomous driving with adaptive reasoning and reinforcement fine-tuning

    Zewei Zhou, Tianhui Cai, Seth Zhao, Yun 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. Advances in Neural Information Processing Systems, 38:27920–2...

    2026

  80. [88]

    The surprising effectiveness of negative reinforcement in llm reasoning.Advances in Neural Information Processing Systems, 38:126546–126573, 2025

    Xinyu Zhu, Mengzhou Xia, Zhepei Wei, Wei-Lin Chen, Danqi Chen, and Yu Meng. The surprising effectiveness of negative reinforcement in llm reasoning.Advances in Neural Information Processing Systems, 38:126546–126573, 2025. 17 Appendix Contents 1 Introduction 1 2 Trajectory Anc...

    2025

  81. [92]

    Do not invent or compute hypothetical trajectories here

    DIRECTION:from the history-visible road geometry, route instruction, and scene constraints, commit to a direction for the 5-second plan (straight / left / right); if the evidence is genuinely ambiguous, name the 1–2 plausible directions and pick the best-supported one, briefly...

  82. [95]

    This decision holds REGARDLESS of which way the road goes — never default to holding speed just because the path looks clear or because you are unsure where the road leads

    SPEED first, and firmly:decide whether to keep speed, slow, or stop from the hazards alone (signals, a lead or stopped vehicle, a crossing agent, a stop line, a tight curve or intersection). This decision holds REGARDLESS of which way the road goes — never default to holding s...

  83. [96]

    velocity

    DIRECTION:from the history-visible road geometry, route instruction, and scene constraints, commit to a direction for the 5-second plan (straight / left / right); if the evidence is genuinely ambiguous, name the 1–2 plausible directions and pick the best-supported one, briefly...

  84. [99]

    This decision holds REGARDLESS of which way the road goes — never default to holding speed just because the path looks clear or because you are unsure where the road leads

    SPEED first, and firmly:decide whether to keep speed, slow, or stop from the hazards alone (signals, a lead or stopped vehicle, a crossing agent, a stop line, a tight curve or intersection). This decision holds REGARDLESS of which way the road goes — never default to holding s...

  85. [100]

    Never infer an unseen future event, hidden traffic-control state, or hypothetical agent behavior

    HISTORY-ONL Y:Construct every criterion solely from facts clearly supported by the provided history-visible frames, ego state, and navigation instruction. Never infer an unseen future event, hidden traffic-control state, or hypothetical agent behavior. If a cue is visually amb...

  86. [101]

    Do NOT prescribe an oracle maneuver, exact future speed, trajectory, waypoint sequence, or option letter

    PROCESS, NOT OUTCOME:Evaluate how the CoT identifies visible constraints and reasons about their driving implications. Do NOT prescribe an oracle maneuver, exact future speed, trajectory, waypoint sequence, or option letter. Final maneuver correctness is evaluated separately b...

  87. [102]

    is specific

    SCENE-GROUNDED & CONCRETE:Each criterion must reference a concrete cue clearly visible in these frames — a specific agent, vehicle, pedestrian, lane, traffic light, road feature, navigation instruction, or the ego speed/heading. This is what makes grading mechanical and stable...

  88. [103]

    Merely mentioning an object is insufficient

    CAUSAL RELEVANCE, NOT OBJECT LISTING:Reward a cue only when the CoT connects it to a driving consequence or constraint. Merely mentioning an object is insufficient. Prefer criteria of the form visible scene fact→safety/legal/feasibil- ity/route implication; do not reward exhau...

  89. [104]

    REASONING FIDELITY:Criteria should reward reasoning that is grounded in visible cues, free of invented or unsupported facts, internally consistent, and relevant to the immediate driving decision — not just fluent or confident language

  90. [105]

    avoids asserting a traffic-light state that is not clearly visible,

    ANTI-GAMING:Include 1–2 criteria that guard against reward-hacking, phrased as GOOD behavior with a POSITIVE weight — e.g., “avoids asserting a traffic-light state that is not clearly visible,” “avoids treating current momentum as sufficient evidence to continue unchanged,” “a...

  91. [106]

    {criterion text} (weight {w1})

  92. [107]

    1":"PRESENT

    {criterion text} (weight {w2}) ... (6–10 criteria) ... Output exactly:{"1":"PRESENT","2":"NOT_PRESENT",...} <candidate chain-of-thought> {normalized turn-1 CoT} Figure 14:Online text-only grader chat template. The grader returns one binary decision per stored criterion; the we...

  93. [108]

    Evidence:from the history-visible scene, identify only the few facts that causally constrain the future motion — the road/route structure and any element whose state changes what the ego can safely, legally, or feasibly do next

  94. [109]

    Causal chain:for each constraint reason scene→consequence (what becomes unsafe, illegal, infeasible, or off-route)→ the maneuver it forces. Treat the ego’s CURRENT motion (its speed and acceleration) as only the starting condition to be acted upon, NEVER as evidence that the m...

  95. [110]

    This decision holds REGARDLESS of which way the road goes — never default to holding speed just because the path looks clear or because you are unsure where the road leads

    SPEED first, and firmly:decide whether to keep speed, slow, or stop from the hazards alone (signals, a lead or stopped vehicle, a crossing agent, a stop line, a tight curve or intersection). This decision holds REGARDLESS of which way the road goes — never default to holding s...

  96. [111]

    The current deceleration and command to go straight imply that an intersection or stopping point is imminent

    Evidence:The ego is on a wet road approaching a marked crosswalk with a bike symbol, traffic signals ahead, and construction barriers on the left. The current deceleration and command to go straight imply that an intersection or stopping point is imminent

  97. [112]

    Contin- uing at the current speed risks violating the traffic control or colliding with pedestrians or stopped vehicles

    Causal chain:The crosswalk and traffic lights require stopping or slowing to a legal stop before crossing. Contin- uing at the current speed risks violating the traffic control or colliding with pedestrians or stopped vehicles. The con- struction barriers also restrict lateral...

  98. [113]

    The crosswalk and traffic signals are mandatory stopping cues, and the wet road increases braking distance, making early deceleration critical

    SPEED first:The ego must decelerate to a stop. The crosswalk and traffic signals are mandatory stopping cues, and the wet road increases braking distance, making early deceleration critical

  99. [114]

    HIGH_LEVEL_DECISION:decelerate to a stop + direction: straight ahead—the crosswalk and traffic lights mandate stopping, and the road geometry does not permit devia- tion

    DIRECTION:The traffic-control layout, crosswalk mark- ings, and navigation command constrain the ego to con- tinue straight; no visible cue supports a turn or curve. HIGH_LEVEL_DECISION:decelerate to a stop + direction: straight ahead—the crosswalk and traffic lights mandate s...

  100. [115]

    The scene is a street intersection with possible cross-traffic or pedestrian crossings

    Evidence:The ego approaches a stop line markedstop, with a stop sign visible and vehicles near the approach. The scene is a street intersection with possible cross-traffic or pedestrian crossings

  101. [116]

    Continuing from the recent constant-speed history would violate that control and risk collision with stopped vehicles or crossing agents

    Causal chain:The sign and painted stop line legally man- date stopping before entry. Continuing from the recent constant-speed history would violate that control and risk collision with stopped vehicles or crossing agents. The recent velocity does not override this scene-impos...

  102. [117]

    SPEED first:The ego must decelerate to zero before proceeding; the stop sign and stop line force the speed decision

  103. [118]

    No visible turning lane or other cue forces a turn, so the evidence supports proceeding straight once traffic permits

    DIRECTION:The road appears to continue straight through the intersection. No visible turning lane or other cue forces a turn, so the evidence supports proceeding straight once traffic permits. HIGH_LEVEL_DECISION:decelerate to a stop + direction: straight(thestoplineandstopsig...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.