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.

1 Introduction

Mainstream Vision-Language-Action (VLA) models for autonomous driving (AD) typically couple a large Vision-Language Model (VLM) with a substantially smaller action expert [59, 27]. The action expert is typically specialized for geometric prediction, whereas high-level reasoning and decision making fall to the VLM, making its AD-specific reasoning capability critical to downstream planning.

Recent work seeks to enhance this ability through CoT supervision [82, 87, 59, 62, 12]. However, when reasoning must resolve into a concrete driving decision [87], rather than scene understanding alone [25, 64, 39, 42], its CoT supervision is typically conditioned on the ground-truth (GT) trajectory. Given the logged future trajectory, the VLM CoT annotator rationalizes a known outcome rather than inferring a decision from scene evidence. This mirrors anchoring bias in cognitive psychology, whereby initially supplied information can disproportionately shape subsequent judgments [76].

We empirically validate trajectory anchoring bias through the controlled study in Figure 1. GT-conditioned CoTs exhibit lower causal faithfulness than causal-planning CoTs, with the degradation primarily concentrated in hard causal scenarios where reliable reasoning is most critical. Moreover, exposing the model to the GT trajectory substantially increases the incidence of severe hallucinations, indicating that trajectory conditioning can inject fabricated causal evidence into the CoT supervision used for subsequent training.

Given this, a natural remedy is to hide the GT trajectory while the teacher derives both its rationale and driving decision from the observed scene, and to verify the predicted future only afterward. This restores the solve-then-verify paradigm used in reasoning-model distillation [51, 72], rather than revealing the answer before constructing its rationale.

For AD, however, open-ended trajectory synthesis is poorly matched to this paradigm because it entangles high-level decision making with precise continuous geometry and low-level dynamics. We therefore seek a language-model-compatible interface through which AD reasoning can emerge from the VLM’s general reasoning capability.

To make trajectory-level driving decisions verifiable without open-ended geometric synthesis, we introduce Autonomous-Driving Multiple-Choice Question (AD-MCQ), a candidate-trajectory benchmark that casts AD planning as selection among scene-specific explicit trajectories. Unlike coarse meta actions, its candidates preserve distinctions in braking time, speed profile, and lateral geometry, grounding each answer in a concrete explicit plan.

AD-MCQ makes trajectory-level driving decisions verifiable, but revealing candidate trajectories before reasoning can simply replace the GT-trajectory anchor with a candidate-set anchor: the policy model inevitably focuses on comparing the relative quality of trajectories, thereby taking shortcuts in reasoning. We therefore propose Deferred Exposure of Future Trajectories for RLVR (DEFT-RLVR), which first commits the policy to a scene-derived decision and only then reveals candidates for explicit grounding. In this way, trajectories supervise reasoning as post-decision targets rather than pre-decision premises.

Across multiple VLM backbones, DEFT-RLVR consistently strengthens autonomous-driving reasoning and decision making while slightly improving aggregate general visual capability.

In summary, our contributions in this work are as follows:

• •

  We identify and empirically validate trajectory anchoring bias: exposing the demonstrated future trajectory produces action-consistent but causally unfaithful rationales, especially in hard causal scenes.

• •

  We introduce AD-MCQ, a verifiable candidate-trajectory benchmark that preserves explicit trajectory-level distinctions and supports both exact selection and candidate-blind reasoning evaluation without open-ended coordinate generation.

• •

  We propose DEFT-RLVR, which defers candidate-trajectory exposure until after the policy has committed to a scene-derived decision and uses exact trajectory correctness and question-specific process supervision. This design improves AD reasoning while preserving the general visual capability of the base policy.

2 Trajectory Anchoring Bias in AD VLMs

Our motivation begins with a simple research question: does revealing the GT future trajectory help a teacher infer a faithful driving rationale, or merely make an already known outcome easier to justify? The AD-VLM must infer the appropriate trajectory from the evidence available in the historical scene. By contrast, the GT trajectory can act as an anchor, allowing teachers to reverse inference and construct a post hoc explanation of the revealed outcome [68, 1]. Consequently, the resulting CoT may be geometrically consistent with the GT trajectory while failing to faithfully identify the scene evidence that genuinely supports the action [3, 71].

We examine this anchoring hypothesis through a human-scored study summarized in Figure 2 and Table 1. GT-conditioned CoTs exhibit lower causal faithfulness, a substantially higher incidence of severe hallucination, and lower pairwise preference than causal-planning CoTs. The study design and detailed analysis are provided in Appendix B.

媒体内容 · 前往原文查看

GT Exposure	Severe Halluc.	Pairwise Win
No (Causal Planning)	29.0%	60.5%
Yes (GT-Conditioned)	50.0%	24.0%

These results expose a severe supervision-direction mismatch: for post hoc chain-of-thought annotation of trajectory decisions, access to the future trajectory serves as a reasoning shortcut rather than a decision target. We therefore retain trajectories as verifiable targets while excluding them from the premises of causal reasoning: the model must first infer a plan from the scene and only then ground it in an explicit future.

3 AD-MCQ: A Verifiable Candidate-Trajectory Benchmark

As shown in Figure 3, AD-MCQ formulates autonomous-driving planning as selecting a future trajectory from a small, scene-specific candidate set. This formulation preserves trajectory-level granularity while replacing open-ended coordinate generation with an exactly verifiable decision.

Task Formulation.

Each AD-MCQ instance consists of multi-view scene-history frames , ego history and current motion state , a navigation instruction , and a shuffled set of explicit candidate trajectories . Exactly one candidate corresponds to the quantized logged future, and its shuffled position serves as the exactly verifiable target.

Discrete Trajectory Prototypes.

Let a fixed-horizon ego trajectory be , where denotes longitudinal and lateral displacement from the current ego pose. We flatten each trajectory in a corpus of logged futures and apply -means to obtain a codebook , where each prototype represents a complete future motion. We quantize by nearest-prototype assignment:

(1)

To determine an appropriate codebook configuration, we study how the number of clustering trajectories and prototypes affect reconstruction fidelity and codebook utilization. Figure 4 shows that provides a favorable balance between out-of-sample reconstruction fidelity and codebook utilization. We provide full construction and scaling analyses in Appendix C.

Candidate-trajectory Construction.

We measure the distance between two prototypes by

(2)

where larger indicates more similar decoded trajectories. For each driving scene , we first map the GT trajectory to its nearest codebook entry . Based on Eq. 2, we define the hard-negative pool as

(3)

We construct split-specific distractors from as detailed in Appendix D and summarized algorithmically in Appendix D.1, combine distinct negatives with , and randomly shuffle the candidates. Finally, we decode the candidate indices into the waypoint-option set shown to the model:

(4)

The VLM then selects one shuffled option, and its decision is evaluated by the deterministic verifier. Complete instances are provided in Appendix G.

Notably, the codebook ultimately retrieves textual waypoint trajectory candidates [29]. Unlike direct waypoint retrieval, the codebook maps continuous futures to a finite motion vocabulary [44, 67, 43], enabling controlled hard-negative construction without sacrificing explicit geometry.

4 DEFT-RLVR: Deferred Exposure of Future Trajectories

Building on AD-MCQ, DEFT-RLVR mitigates trajectory-induced anchoring bias by deferring candidate exposure until after the policy commits to a scene-derived decision, while jointly optimizing the two interaction stages with rubric-based reasoning supervision.

4.1 DEFT: Deferred Exposure of Future Trajectories

By deferring candidate exposure, DEFT reserves candidate geometry for grounding an already formed scene-derived decision rather than shaping the decision itself.

Specifically, for question , let denote the scene context and the candidate trajectories with option label set .

Turn 1: Causal Decision Reasoning. Conditioned solely on the scene context , the causal-reasoning prompt (provided in Appendix F.1) elicits causal reasoning over scene evidence before the candidate trajectories are revealed:

(5)

The resulting response explicitly traces how scene evidence leads to driving implications and commits to a complete high-level decision (HLD) before candidate exposure.

Turn 2: Explicit-Trajectory Grounding. Only after this decision has been formed do we reveal through the trajectory-matching prompt (provided in Appendix F.1), which treats the recorded decision as binding and uses candidate geometry only to identify the closest explicit realization of that decision. With , we sample:

(6)

The selected option is , where denotes an invalid output. This design prevents candidate geometry from conditioning the initial reasoning process while retaining exact trajectory-level verification.

4.2 Joint Optimization of the Two-Stage Interaction

Although DEFT separates candidate-free decision formation from trajectory grounding, we optimize them jointly as a single rollout using Group Relative Policy Optimization (GRPO) [14]. For each question , we sample two-turn rollouts under the interaction defined in Section 4.1. We serialize rollout as the complete two-turn sequence . During optimization, we perform a single forward pass over to compute the token likelihoods used for importance sampling, while masking the prompt tokens so that the policy objective is applied only to the generated tokens in and . Both turns share the rollout reward and its group-normalized advantage.

4.3 Structured Rubric Rewards for Reasoning-Trace Supervision

AD-MCQ provides a verifiable outcome reward:

(7)

where is the oracle option, but this signal alone cannot distinguish grounded reasoning from rationalization.

To prevent reinforcing reasoning trajectories that arrive at the correct answer through shortcut exploitation or random guessing [15], we apply rubric-based reasoning rewards to rollouts with correct MCQ answers. For each answer-correct rollout, we form the normalized Turn-1 reasoning trace and asynchronously submit to a text grader for evaluation. Specifically, DEFT-RLVR generates an instance-specific rubric once offline using a vision-language rubric generator conditioned on the fixed generation prompt (provided in Appendix F.2):

(8)

The resulting rubric contains atomic, positively weighted criteria that explicitly encode scene evidence verified by the offline generator and is reused across rollouts [13, 49, 48].

During RL rollouts, we prompt a shared VLM judge with (provided in Appendix F.3) to evaluate each reasoning trace against the rubric criteria set :

(9)

where indicates that the CoT satisfies criterion . With , the rubric reward is:

(10)

The final rollout reward is

(11)

Thus, trajectory correctness determines whether a rollout receives process supervision, while neither the GT trajectory nor the candidate set serves as input to the Turn-1 causal reasoning process. Crucially, is never directly provided to , or . The judge evaluates the reasoning process solely according to the predefined rubric criteria, without direct access to the visual input, candidate options, or GT trajectory. Compared with directly providing a VLM-based judge with the full visual context, this text-only grading scheme is substantially more efficient. Moreover, it allows the judge to focus on assessing the quality of the textual reasoning, without its attention being diluted by a large number of visual tokens [86, 10].

5 Experiments

媒体内容 · 前往原文查看

Method	AD-Specific Reasoning			General Visual Capability(%)
	ACC(%)	CFS	HLD	Basic	Embodied	3D/MV	RefSpatial	Avg.
Qwen3-VL-8B-Instruct	28.1	–	–	81.60	56.63	42.48	38.52	54.81
+ DEFT	56.6	0.431	0.425
+ JEFT RLVR ()	61.1	0.428	0.427	81.39	56.67	41.99	39.56	54.90
+ DEFT RLVR ()	76.4	0.442	0.462	81.36	57.15	42.80	44.26	56.39
+ DEFT RLVR ()	75.2	0.580	0.487	81.59	56.31	42.81	44.79	56.37
+ DEFT-RLVR	77.9	0.658	0.501	81.33	57.41	42.27	43.35	56.09
+ JEFT Distillation	64.0	0.620	0.480	74.20	52.80	39.60	32.60	49.80
+ DEFT Distillation (Plan Only)	68.2	0.925	0.560	78.00	53.49	40.20	33.46	51.29
+ DEFT Distillation (Full Interaction)	82.4	0.909	0.591	74.94	54.04	40.38	34.56	50.98
+ DEFT Distillation (Mixed Targets)	84.1	0.934	0.627	76.73	53.51	40.20	36.76	51.80
Qwen3.5-4B	34.0	–	–	80.31	53.00	40.73	36.19	52.56
+ DEFT	65.6	0.738	0.481
+ JEFT RLVR ()	72.3	0.740	0.511	80.23	53.34	40.16	31.32	51.26
+ DEFT RLVR ()	79.0	0.805	0.529	80.76	52.75	41.83	38.72	53.52
+ DEFT RLVR ()	79.4	0.819	0.540	80.41	53.37	40.67	38.62	53.27
+ DEFT-RLVR	82.2	0.822	0.582	80.88	55.20	40.94	35.93	53.24

5.1 Experimental Setup

Data and benchmark.

We divide scenes from Waymo Open E2E [69] and an internal driving corpus into Train, Dev, and AD-MCQ-500 Test splits, containing , , and scenes, respectively, with and . Each visual input contains four frames sampled at from three cameras: front-left, front, and front-right. Train follows the natural scene distribution, whereas Dev and Test focus on causally demanding scenes with structured hard distractors. Dev is curated as the harder of the two evaluation splits. Details are shown in Appendix D.1.

Evaluation.

For AD-specific evaluation, we report ACC (AD-MCQ-500 accuracy) and two complementary CoT metrics: CFS (Normalized Causal-Faithfulness Score) and HLD (High-Level-Decision Consistency).

To assess general-capability retention, we use vision-language benchmarks covering four capability groups [9, 77]: basic visual perception (Basic Visual) [60, 74, 54], embodied spatial reasoning (Embodied Spatial) [8, 53, 57, 36], 3D and multi-view reasoning (3D/Multi-View) [37, 75, 28, 61], and referring-expression-based spatial grounding (RefSpatial) [84]. We additionally evaluate cross-domain AD transfer on an external -scene nuScenes set. Complete evaluation settings are provided in Appendix D.2.

Models.

We use Qwen3-VL-8B-Instruct [2] and Qwen3.5-4B [46] as the base models. Qwen3.5-397B-A17B provides supervision targets for distillation, while Qwen3.6-35B-A3B generates instance-specific rubrics offline and serves as the reasoning-process judge.

Candidate-Exposure Settings.

DEFT (Deferred Exposure of Future Trajectories) first elicits a pre-exposure plan and reveals the candidate trajectories only in the subsequent selection turn. Conversely, JEFT (Joint Exposure of Future Trajectories) is the matched ablation that presents the same scene context and candidate trajectories jointly with the same prompts. For a fair comparison, both settings use , top-, and the same total token budget of tokens, with DEFT capped at tokens per turn.

RLVR Variants.

JEFT RLVR () and DEFT RLVR () use the same exact-choice reward and differ in candidate-exposure order. DEFT RLVR () uses a shared rubric whose score is assigned online by a scene-conditioned VLM grader. DEFT-RLVR uses the instance-specific reward defined in Eq. 11.

Distillation Variants.

We prompt Qwen3.5-397B-A17B under the corresponding exposure setting and use its responses as supervised fine-tuning targets for the student. JEFT Distillation imitates the single-turn reasoning-and-selection response generated with candidates exposed from the outset. DEFT Distillation (Plan Only) imitates only the Turn-1 plan generated before candidate exposure. DEFT Distillation (Full Interaction) supervises the complete two-turn plan-then-match interaction. DEFT Distillation (Mixed Targets) uses an equal mixture of plan-only and full-interaction targets. Detailed implementations of the RLVR and distillation variants are provided in Appendix D.4.

5.2 Candidate-Grounded Training Improves Generalizable AD Reasoning

DEFT mitigates candidate anchoring bias across inference and training.

As illustrated in Table 2, Compared with training-free JEFT, DEFT raises ACC from to on Qwen3-VL-8B and from to on Qwen3.5-4B. Under the same correctness-only reward, DEFT RLVR () exceeds its matched JEFT RLVR () ablation by and percentage points on Qwen3-VL-8B and Qwen3.5-4B, respectively. Across both backbones, DEFT-RLVR jointly improves MCQ accuracy, CFS, and HLD over training-free DEFT, demonstrating gains in reasoning quality and decision consistency rather than final-choice accuracy alone. Under equal-data distillation, DEFT Distillation (Full Interaction) outperforms JEFT Distillation, raising MCQ accuracy from to , while improving CFS by and HLD by .

These consistent performance gaps can be attributed to the same information-order mechanism: when future-trajectory candidates are visible during decision formation, the policy may organize the reasoning around a favored answer, creating a shortcut consistent with the anchoring bias shown in Section 2. DEFT prevents this shortcut by requiring the model to derive its decision from scene evidence before grounding it in a concrete trajectory.

Fine-grained trajectory grounding strengthens generalizable AD reasoning.

Compared with the Plan Only variant, DEFT Distillation (Mixed Targets) incorporates full two-turn targets and raises accuracy from to , CFS from to , and HLD from to . These simultaneous gains show that Turn-2 trajectory grounding is more than a mechanism for providing RL with an exact, verifiable reward: the required fine-grained discrimination among trajectory candidates also improves reasoning quality and trajectory-decision accuracy.

Candidate-grounded training delivers AD reasoning gains that generalize to an OOD driving domain.

On the out-of-distribution (OOD) nuScenes domain [4], both DEFT training variants still significantly outperform training-free DEFT on all three AD metrics. As shown in Table 3, DEFT-RLVR raises candidate accuracy from to , while improving CFS by and HLD by .

Together, these gains indicate that our training paradigm helps the policy learn transferable scene-to-decision reasoning and subsequent explicit-trajectory grounding, rather than rely on source-specific visual cues.

媒体内容 · 前往原文查看

Method	ACC	CFS	HLD
DEFT (Training-Free)	39.6	0.522	0.286
DEFT Distillation (Mixed Targets)	55.8	0.655	0.370
DEFT-RLVR	49.5	0.636	0.359

5.3 RLVR Improves AD Reasoning without Sacrificing General Visual Capability

Distillation improves AD specialization at the cost of general capability.

As shown in Table 2, JEFT Distillation reduces the average of general visual capability from to , while the equal-data DEFT variants retain –. This is because token-level SFT supervision pushes the student toward an AD-specific response distribution generated by an external teacher rather than selectively reinforcing correct behavior. Although DEFT-based distillation mitigates this policy shift relative to shortcut-prone JEFT distillation, dense teacher imitation still trades general capability for AD performance.

Cold-start SFT causes an early decline in general visual capability.

To assess whether RLVR should start from an AD-specialized policy, we first apply SFT to teacher-generated responses as a cold-start stage, with experimental details shown in Appendix D.5. As shown in Figure 5, although cold-start SFT improves Dev accuracy, all four general-capability groups decline from the first epoch. Initializing RL from this policy would additionally anchor KL regularization to an already shifted policy. Thus, we start DEFT-RLVR from the unmodified Base VLM.

DEFT-RLVR improves AD reasoning while preserving general visual capability.

As illustrated in Table 2, DEFT-RLVR raises the average of general visual capability from to on Qwen3-VL-8B and from to on Qwen3.5-4B. Unlike SFT, RL-based variants learn from responses sampled from the current or a recent policy. The resulting policy gradients merely increase or decrease the probability of each sampled token conditioned on its corresponding context [88], thereby constituting a more fine-grained form of policy optimization than SFT [11]. Meanwhile, the causal reasoning process partially exercises visual-spatial reasoning shared with the general benchmarks, which may explain the modest gains of general visual capability. The evaluation results of all 12 benchmarks are provided in Appendix H.

5.4 Ablation of the RLVR Design

JEFT RLVR () vs. DEFT RLVR (): deferred exposure avoids shortcut-driven optimization.

Figure 6 shows that the JEFT variant exhibits substantially larger policy drift while remaining less accurate than DEFT variants. As shown in Table 2, DEFT improves accuracy from to on Qwen3-VL-8B and from to on Qwen3.5-4B, while also achieving higher CFS and HLD. These results suggest that deferred exposure effectively reduces candidate-visible shortcuts and promotes more effective scene-derived reasoning.

DEFT RLVR () vs. DEFT-RLVR: rubric supervision prunes unproductive exploration.

As shown in Figure 6, DEFT RLVR () produces the longer and higher-entropy responses than rubric-supervised variants. Meanwhile, Table 2 shows that the introduction of rubric supervision consistently enhances the reasoning and HLD quality on both backbones. This improvement can be attributed to the rubric’s fine-grained supervision, which effectively steers the reasoning process toward greater faithfulness while suppressing unproductive exploration.

DEFT-RLVR vs. DEFT RLVR (): stronger reasoning capacity with marginal time cost.

As shown in Table 4, compared with DEFT RLVR (), DEFT-RLVR substantially improves reasoning quality while introducing marginal training cost, increasing total step time by only , from to seconds. DEFT-RLVR likewise achieves higher reasoning quality than the online-rubric variant, DEFT RLVR (), despite incurring substantially lower training costs. Specifically, by constructing instance-specific criteria offline and retaining only text-based grading online, DEFT-RLVR reduces total step time from to seconds ( faster).

媒体内容 · 前往原文查看

Method	Step	Rollout	Scoring
DEFT RLVR ()	424.5	171.1	0.03
DEFT RLVR ()	724.4	489.8	156.8
DEFT-RLVR	426.5	174.6	4.12

5.5 Why Formulate AD Planning as a Candidate-Grounded MCQ?

Candidate grounding avoids the dual bottleneck of trajectory error and general-capability degradation.

Using a shared codebook, we fine-tune Qwen3-VL-8B-Instruct via SFT to predict trajectory tokens with or without trajectory-conditioned CoT; full details are provided in Appendix D.6. Figure 7(a,b) shows that, even after SFT begins to overfit, both in- and out-of-distribution prediction ADEs remain substantially above the codebook’s m reconstruction ADE, identifying token generation as the primary error source. Figure 7(c) further indicates that direct trajectory generation severely degrades general visual capability even with CoT. Both failures arise because training forces the VLM to internalize a large trajectory-token inventory within its original vocabulary, substantially perturbing the pretrained token distribution. AD-MCQ instead uses the codebook only to retrieve waypoint candidates and externalizes generation as selection among scene-conditioned explicit trajectories, thereby avoiding both bottlenecks.

The gains of DEFT-RLVR generalize across diverse MCQ option constructions.

Holding scenes and oracle trajectories fixed, we randomly resample distractors across five candidate counts and four hard-negative similarity bounds. Figure 8 shows that DEFT-RLVR outperforms training-free DEFT in all settings. Thus, the capability learned by DEFT-RLVR under a fixed MCQ configuration transfers to new candidate-set constructions rather than relying on a particular distractor geometry.

Meanwhile, highly similar future trajectories and larger candidate sets remain the most challenging regimes for fine-grained candidate grounding. Through controlled option construction, AD-MCQ thus provides a simple yet difficulty-controllable experimental paradigm for future research. Appendix D.7 provides further experimental details.

6 Conclusion

We identify anchoring bias in AD VLMs, propose AD-MCQ and leverage DEFT-RLVR for training. This framework improves generalizable AD reasoning while preserving and even enhancing the model’s general visual capabilities. Since AD-MCQ relies solely on the VLM and allows difficulty to be controlled through option construction, it provides a highly deployable and scalable foundation for future research.

References

• Arcuschin et al. [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.
• Bai et al. [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.
• Balasubramanian et al. [2025] Sriram Balasubramanian, Samyadeep Basu, and Soheil Feizi. A closer look at bias and chain-of-thought faithfulness 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 2025. Association for Computational Linguistics. ISBN 979-8-89176-335-7. 10.18653/v1/2025.findings-emnlp.723. https://aclanthology.org/2025.findings-emnlp.723/.
• Caesar et al. [2020] 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.
• Chandak et al. [2025] Nikhil Chandak, Shashwat Goel, Ameya Prabhu, Moritz Hardt, and Jonas Geiping. Answer matching outperforms multiple choice for language model evaluation. arXiv preprint arXiv:2507.02856, 2025. 10.48550/arXiv.2507.02856.
• Chen et al. [2025] 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.
• Chitta et al. [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.
• Du et al. [2024] 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. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers), pages 346–355, 2024.
• Duan et al. [2024] 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.
• Fei et al. [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.
• Fu et al. [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.
• Gu et al. [2026] Yi Gu, Yan Wang, Yuxiao Chen, Yurong You, Wenjie Luo, Yue Wang, Wenhao Ding, Boyi Li, Heng Yang, Boris Ivanovic, et al. Accelerating structured chain-of-thought in autonomous vehicles. arXiv preprint arXiv:2602.02864, 2026.
• Gunjal et al. [2025] 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.
• Guo et al. [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.
• Guo et al. [2026] 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. In Findings of the Association for Computational Linguistics: ACL 2026, pages 20092–20113, 2026.
• Hashemi et al. [2024] Helia Hashemi, Jason Eisner, Corby Rosset, Benjamin Van Durme, and Chris Kedzie. Llm-rubric: A multidimensional, calibrated approach to automated evaluation of natural language texts. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 13806–13834, 2024.
• Hu et al. [2023] 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. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17853–17862, 2023.
• Huang et al. [2023a] Zhiyu Huang, Haochen Liu, and Chen Lv. Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3903–3913, 2023a.
• Huang et al. [2023b] 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, 2023b.
• Huang et al. [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.
• Huang et al. [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.
• Huang et al. [2026a] 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, 2026a.
• Huang et al. [2026b] 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. 2026b. https://api.semanticscholar.org/CorpusID:285240754.
• Hwang et al. [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.
• Ishaq et al. [2025] 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. In 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 20501–20508. IEEE, 2025.
• Jiang et al. [2023] 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. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340–8350, 2023.
• Jiang et al. [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.
• Li et al. [2025a] 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, 2025a.
• Li et al. [2025b] 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, 2025b.
• Li et al. [2025c] Yiheng Li, Cunxin Fan, Chongjian Ge, Zhihao Zhao, Chenran Li, Chenfeng Xu, Huaxiu Yao, Masayoshi Tomizuka, Bolei Zhou, Chen Tang, Mingyu Ding, and Wei Zhan. Womd-reasoning: A large-scale dataset for interaction reasoning in driving, 2025c. https://arxiv.org/abs/2407.04281.
• Liao et al. [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.
• Liu et al. [2026a] 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, 2026a.
• Liu et al. [2026b] 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, 2026b.
• Ma et al. [2025a] Hongbo Ma, Fei Shen, Hongbin Xu, Xiaoce Wang, Gang Xu, Jinkai Zheng, Liangqiong Qu, and Ming Li. Styletailor: Towards personalized fashion styling via hierarchical negative feedback, 2025a. https://arxiv.org/abs/2508.06555.
• Ma et al. [2025b] 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. 2025b. 10.1609/aaai.v39i6.32640. https://ojs.aaai.org/index.php/AAAI/article/view/32640.
• Ma et al. [2026] 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.
• Ma et al. [2025c] Wufei Ma, Haoyu Chen, Guofeng Zhang, Yu-Cheng Chou, Jieneng Chen, Celso de Melo, and Alan Yuille. 3dsrbench: A comprehensive 3d spatial reasoning benchmark. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6924–6934, 2025c.
• Mao et al. [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.
• Marcu et al. [2024] 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 autonomous driving. In European Conference on Computer Vision, pages 252–269. Springer, 2024.
• Nie et al. [2024] 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. In European Conference on Computer Vision, pages 292–308. Springer, 2024.
• Ouyang et al. [2022] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35:27730–27744, 2022.
• Park et al. [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.
• Pertsch et al. [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.
• Philion et al. [2023] Jonah Philion, Xue Bin Peng, and Sanja Fidler. Trajeglish: Traffic modeling as next-token prediction. arXiv preprint arXiv:2312.04535, 2023.
• Qian et al. [2024] 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. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 4542–4550, 2024.
• Qwen Team [2026] Qwen Team. Qwen3.5: Towards native multimodal agents, February 2026. https://qwen.ai/blog?id=qwen3.5.
• Rafailov et al. [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.
• Rao and Callison-Burch [2026] Delip Rao and Chris Callison-Burch. Autorubric: Unifying rubric-based llm evaluation. arXiv preprint arXiv:2603.00077, 2026.
• Rezaei et al. [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.
• Schulman et al. [2017] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017.
• Shao et al. [2024] 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.
• Sima et al. [2024] 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. In European Conference on Computer Vision, pages 256–274. Springer, 2024.
• Song et al. [2024] Chan Hee Song, Valts Blukis, Jonathan Tremblay, Stephen Tyree, Yu Su, and Stan Birchfield. Robospatial: Teaching spatial understanding to 2d and 3d vision-language models for robotics. arXiv preprint arXiv:2411.16537, 2024.
• Sun et al. [2026] 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, 2026. https://arxiv.org/abs/2502.10273. ES-Reasoning Workshop at ICLR 2026.
• Sun et al. [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.
• Tan et al. [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, 38:38310–38335, 2025.
• Team et al. [2025] Gemini Robotics Team, Saminda Abeyruwan, Joshua Ainslie, Jean-Baptiste Alayrac, Montserrat Gonzalez Arenas, Travis Armstrong, Ashwin Balakrishna, Robert Baruch, Maria Bauza, Michiel Blokzijl, et al. Gemini robotics: Bringing ai into the physical world. arXiv preprint arXiv:2503.20020, 2025.
• Tian et al. [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.
• Tian et al. [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.
• Tong et al. [2024] 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. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024.
• Wang et al. [2026] 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, 2026. https://openreview.net/forum?id=AwC77yHpP6.
• Wang et al. [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.
• Wang et al. [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.
• Wei et al. [2025] Zhaoyang Wei, Chenhui Qiang, Bowen Jiang, Xumeng Han, Xuehui Yu, and Zhenjun Han. Adˆ 2-bench: A hierarchical cot benchmark for mllm in autonomous driving under adverse conditions. arXiv preprint arXiv:2506.09557, 2025.
• Wu et al. [2024a] 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, 2024a.
• Wu et al. [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.
• Wu et al. [2024b] 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, 2024b.
• Xu et al. [2024a] 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 2024a. Association for Computational Linguistics. 10.18653/v1/2024.findings-acl.876. https://aclanthology.org/2024.findings-acl.876/.
• Xu et al. [2026] 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. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3709–3718, 2026.
• Xu et al. [2024b] 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, 2024b.
• Xu et al. [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.
• Yang et al. [2025a] 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, 2025a.
• Yang et al. [2026] Bangji Yang, Hongbo Ma, Jiajun Fan, and Ge Liu. Batched contextual reinforcement. In Forty-third International Conference on Machine Learning, 2026. https://openreview.net/forum?id=8Oc3Mx754M.
• Yang et al. [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.
• Yang et al. [2025b] 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, 2025b.
• Yasseri and Reher [2022] 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.
• Yu et al. [2026] 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.
• Yue et al. [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. 10.1609/aaai.v40i41.40752. https://doi.org/10.1609/aaai.v40i41.40752.
• Zhang et al. [2024a] 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, 2024a.
• Zhang et al. [2024b] 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, 2024b.
• Zhang et al. [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.
• Zhao et al. [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.
• Zheng et al. [2024] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. In European Conference on Computer Vision, pages 87–104. Springer, 2024.
• Zhou et al. [2026a] 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, 38:28404–28481, 2026a.
• Zhou et al. [2025a] 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, 2025a.
• Zhou et al. [2025b] Xueyang Zhou, Yangming Xu, Guiyao Tie, Yongchao Chen, Guowen Zhang, Duanfeng Chu, Pan Zhou, and Lichao Sun. Libero-pro: Towards robust and fair evaluation of vision-language-action models beyond memorization. arXiv preprint arXiv:2510.03827, 2025b.
• Zhou et al. [2026b] 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, 38:27920–27956, 2026b.
• Zhu et al. [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.

Appendix

Appendix A Related Work

A.1 Vision–Language Reasoning for Autonomous Driving

Language-conditioned driving models have evolved from using language as an auxiliary source of supervision to placing a VLM directly in the perception, reasoning, and planning loop. Early systems formulate driving as graph-based visual question answering, language-conditioned behavior prediction, or interpretable trajectory generation [52, 38, 63, 70]. Subsequent work expands this direction through multi-view scene reasoning, knowledge augmentation, behavioral planning states, and unified vision–language–action architectures [59, 79, 27, 24, 85]. Driving-oriented question-answering and reasoning benchmarks complement these models by measuring scene understanding, spatial reasoning, and interpretable decision making [39, 42, 45, 58, 40, 30].

More recent methods explicitly supervise chain-of-thought reasoning or combine reasoning traces with action learning [62, 82, 87, 25, 64, 12]. These approaches establish the value of explicit intermediate reasoning for driving. Our work addresses a distinct question concerning the direction of that supervision: when a rationale is generated with access to the logged future trajectory, the trajectory can become a premise from which the teacher works backward. We instead require a scene-grounded decision before exposing future-trajectory candidates, retaining trajectory-level supervision while preventing the target future from anchoring the initial reasoning process.

A.2 Trajectory Representations and End-to-End Planning

End-to-end driving has been studied through sensor-fusion policies, planning-oriented representations, sparse scene abstractions, and integrated prediction–planning architectures [7, 17, 26, 55, 19, 18]. Generative planners further model multimodal futures with autoregressive, diffusion, or flow-based objectives [83, 31, 20, 56]. In parallel, discretized action and trajectory representations make continuous behavior compatible with token-based sequence models [44, 67, 80, 43, 29]. Such representations reduce the mismatch between language-model decoding and continuous control, but direct full-vocabulary trajectory generation still requires the VLM to synthesize precise geometry and can encourage task-specific memorization.

AD-MCQ uses trajectory prototypes differently. The prototypes define a scene-specific set of decoded, explicit candidates rather than a global action vocabulary that the VLM must generate. This formulation preserves differences in lateral geometry, braking time, and speed profile while turning trajectory-level planning into exact candidate selection. It is therefore closer to a verification interface than to a replacement for a downstream continuous planner. Our candidate-construction and held-out representation analyses further separate codebook sufficiency from the difficulty of full-vocabulary trajectory-token prediction.

A.3 Verifiable Post-Training and Faithful Reasoning

Supervised instruction tuning, preference optimization, and reinforcement learning provide complementary mechanisms for adapting foundation models [41, 47, 50, 66, 65, 34, 35, 21]. For reasoning models, reinforcement learning with automatically checkable outcomes can elicit capabilities without requiring imitation of every intermediate step [51, 14, 72, 78, 33, 22, 73, 81]. Recent work extends this principle beyond exact symbolic answers through multidimensional rubrics and rubric-derived rewards [16, 48, 13, 49, 23], while studies of multiple-choice RLVR show that distractor construction and answer format materially shape the learned behavior [15, 5]. DEFT-RLVR combines exact candidate correctness with instance-specific rubric rewards, but gates process rewards on outcome correctness and grades the candidate-blind reasoning trace without visual or answer-related information.

This design is also motivated by evidence that a chain of thought need not be a faithful account of the evidence that produced an answer. Preemptively revealing an answer can distort subsequent reasoning, and both language and vision–language models may rationalize cues or hallucinate support for an already favored conclusion [68, 1, 32, 3, 6, 71]. Our controlled annotation study instantiates this issue in trajectory-level driving decisions, and our deferred-exposure formulation converts the future trajectory from a pre-reasoning cue into a post-decision verification target.

Appendix B Causal Faithfulness under Future-Trajectory Exposure

We provide the study design and detailed analysis for the comparison summarized in Section 2. We examine whether revealing the logged GT future trajectory helps a teacher infer a faithful driving rationale or merely makes an already known outcome easier to justify.

B.1 Study Design

We evaluate the effect of GT-conditioned annotation on 100 strong-causal driving scenes, including 70 Waymo scenes and 30 internal scenes. We select scenes in which the logged future trajectory substantially deviates from constant-velocity extrapolation, covering hard braking, stopping from motion, and sharp turns. We detail the construction of the hard causal evaluation set in Appendix D.1. For each scene, we construct paired annotations using the same teacher. Across the two experimental settings, we fix Qwen3.5-397B-A17B as the teacher, the same 12 visual frames, ego-state and navigation text, the complete system prompt and all user instructions outside the intervention block, the four-item causal-reasoning body, the HIGH_LEVEL_DECISION output contract, temperature , and disabled thinking. The two arms use a byte-identical system prompt and an otherwise identical user template. The sole intervention is one contiguous block containing the logged GT future as 10 raw ego-frame waypoints: the causal-planning arm omits this block, whereas the GT-conditioned arm inserts it before the shared reasoning instructions. Both arms still ask the teacher to infer and commit to a high-level decision from the scene; the GT-conditioned prompt does not ask the teacher to justify a known action, repeat an action label, or select among candidates. Consequently, the paired comparison changes only the availability of future-trajectory information while holding the task wording, inputs, model, decoding, and response format fixed. The complete role-separated templates below expose the single insertion point directly.

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.

Figure 10: Complete GT-conditioned chat template. It is identical to Figure B.1 except for the highlighted raw-waypoint GT block; no derived action label or rationalization-specific instruction is introduced.

B.2 Human Evaluation

We conduct a human evaluation of the two experimental settings. Across the 100 scenes and two settings, we obtain 200 CoTs in total. We ask two annotators to independently score every CoT along four dimensions: grounding (GND), absence of hallucination (NO-HALL), specificity (SPEC), and causal coherence (COH). Each dimension is scored on a three-point ordinal scale, where 0 denotes a clear failure with a consequential error, 1 denotes partial satisfaction with an omission or minor error, and 2 denotes full satisfaction without a substantive error. We thus collect dimension-level ratings. For each CoT, we average the two annotations on each dimension and sum the four averaged scores to obtain a causal-faithfulness score (CFS) from 0 to 8. The annotators work independently, and disagreements are retained rather than resolved through discussion or adjudication.

We instantiate the three-point scale with dimension-specific observable criteria. For GND, 0 indicates that the stated rationale conflicts with or is unsupported by the visible scene, 1 indicates that it uses some relevant scene evidence but omits or misinterprets a non-critical cue, and 2 indicates that its decision-relevant claims are supported by the observed scene. For NO-HALL, 0 indicates a consequential fabricated object, event, traffic control, or interaction, 1 indicates an unsupported but non-critical detail, and 2 indicates no unsupported factual claim. We count a CoT as a severe hallucination for an annotator when its NO-HALL score is 0; we compute the reported severe-hallucination rate by averaging this binary indicator over annotators and CoTs within each experimental setting. For SPEC, 0 indicates a generic rationale that could apply to unrelated scenes, 1 indicates limited use of scene-specific actors or geometry, and 2 indicates sufficient reference to the particular actors, spatial relations, and traffic context that determine the maneuver. For COH, 0 indicates that the conclusion does not follow from the stated evidence or contains a major contradiction, 1 indicates a broadly plausible causal chain with a missing link or minor inconsistency, and 2 indicates a complete and internally consistent connection from scene evidence to the proposed maneuver.

For dimension-level scoring, we anonymize the experimental settings by replacing their names with random identifiers and randomizing presentation order. We present the scene and a single CoT in each item and never show its paired counterpart alongside it. The annotators are not informed of our anchoring hypothesis or which setting produced an item, and we withhold the GT future trajectory during scoring. We use the same teacher model, scene inputs, reasoning template, decoding settings, and output format for the two settings, as specified in the study design above.

After completing the independent dimension-level scoring, each annotator separately evaluates all 100 scene-matched CoT pairs. For each scene, we present the two CoTs together in randomized order without experimental-setting labels, and ask the annotator to select the better rationale or record a tie using the same grounding, absence-of-hallucination, specificity, and causal-coherence criteria defined above. The two annotators therefore provide 200 independent pairwise judgments: causal planning is preferred in 121 judgments (60.5%), GT-conditioned reasoning is preferred in 48 (24.0%), and 31 judgments (15.5%) are ties.

Appendix C Trajectory Codebook Construction and Representation Analysis

In this section, we present the trajectory-codebook construction and representation analysis underlying AD-MCQ. Notably, the codebook serves a specific role in our benchmark: (1) the codebook is a purely kinematic aggregation of motion trajectories and is aggregated independently of scene observations; (2) we use it only to retrieve waypoint candidates when constructing AD-MCQ. The codebook itself and its indices are never exposed to the policy, which receives only the corresponding decoded waypoint candidates. We first formalize this representation, then describe the reconstruction evaluation setup and analyze how codebook size and data scale determine reconstruction fidelity and prototype utilization.

C.1 Trajectory Representation and Quantization

A logged future is represented as a ego-frame trajectory sampled at , with . Flattening yields a -dimensional vector. Given such trajectories, we apply -means to obtain , where each centroid represents a complete speed and lateral-motion profile. As defined in Eq. 1, each trajectory is assigned to its nearest centroid under cumulative squared waypoint distance.

For implementation, we associate each codebook index with a symbolic identifier . Encoding returns the identifier of the nearest prototype, whereas decoding retrieves the complete waypoint sequence, . These identifiers provide a discrete indexing space for candidate construction; AD-MCQ serializes the decoded coordinates rather than the identifiers themselves. Consequently, quantization remains external to the VLM and does not require modifying its vocabulary or generating dense coordinates autoregressively.

C.2 Reconstruction Evaluation Setup

We pool approximately logged trajectories from the Waymo Open E2E corpus and an internal driving corpus. Waymo ego future_states are subsampled from to ; both sources use the same , -waypoint, ego-frame, meter convention and are therefore clustered jointly. We reserve ( trajectories) from codebook fitting for same-distribution out-of-sample evaluation and use an independent Waymo validation split ( trajectories) for cross-source evaluation.

Clustering uses MiniBatchKMeans with batch size , iterations, and three initializations. For each configuration, we repeat K-means with three initialization seeds and report the mean over seeds separately for the fitting corpus, the held-out split, and the independent Waymo split. We measure average displacement error (ADE), final displacement error (FDE), their p50/p95/p99 tail statistics, and codebook utilization, defined as the fraction of prototypes assigned at least one evaluation trajectory.

C.3 Resolution as the Codebook Scales

Table 5 reports the resolution sweep at the full clustering-set size. Increasing reduces ADE and FDE smoothly rather than producing a sharp saturation point; each doubling lowers in-sample ADE by approximately . The improvement follows the approximate trend , but increasingly fine codebooks allocate prototypes to sparse motions that are not recovered on held-out trajectories.

媒体内容 · 前往原文查看

	In ADE	Out ADE	Cross ADE	In FDE	Out FDE	In Util.	Out Util.
256	0.680	0.675	0.704	1.281	1.273	100.0%	100.0%
512	0.567	0.563	0.592	1.063	1.057	100.0%	100.0%
1024	0.480	0.478	0.505	0.890	0.887	100.0%	100.0%
2048	0.399	0.399	0.432	0.729	0.729	100.0%	100.0%
4096	0.333	0.336	0.374	0.601	0.605	100.0%	98.7%
8192	0.283	0.290	0.331	0.504	0.516	99.7%	92.0%
16384	0.242	0.252	0.297	0.426	0.446	99.4%	73.7%

媒体内容 · 前往原文查看

		In ADE	Out ADE	Gap	Out Util.
	5,000	0.407	0.528		99.8%
	10,000	0.442	0.509		99.8%
	20,000	0.464	0.496		100.0%
	50,000	0.476	0.489		100.0%
	100,000	0.476	0.479		100.0%
	200,000	0.477	0.477		100.0%
	489,042	0.480	0.478		100.0%
	50,000	0.243	0.303		89.6%
	100,000	0.266	0.297		90.4%
	200,000	0.274	0.290		91.6%
	489,042	0.283	0.290		92.0%

The resolution–coverage trade-off is visible in both Table 5 and Figure 4a. Out-of-sample utilization remains complete through , is at and at , but falls to at . The independent Waymo split exhibits the same monotonic resolution trend with a consistent – ADE offset, showing that the comparison across is not specific to a single held-out split.

C.4 Data Scale and Generalization

We next vary the number of clustering trajectories while fixing . As summarized in Table 6 and Figure 4b, small clustering sets yield an artificially low in-sample error but a larger held-out error because their centroids specialize to incidental sample positions. Increasing closes this gap, with in-sample and out-of-sample reconstruction approaching convergence once the number of trajectories per prototype becomes sufficiently large.

For , the gap is effectively closed once reaches approximately ; for , it decreases to at the full data scale. The long tail remains the main source of quantization error: at full data, ADE p99 decreases from at to at , and FDE p99 decreases from to . Thus, larger codebooks improve rare-motion reconstruction but do not eliminate long-tail error by themselves.

C.5 Codebook Selection

Our choice of follows from the joint behavior of reconstruction fidelity, held-out utilization, and data support. At the full clustering scale, it achieves out-of-sample ADE and out-of-sample FDE while retaining utilization. Doubling the codebook further improves displacement error, but the fraction of prototypes exercised out of sample drops by more than . Conversely, smaller codebooks retain nearly complete utilization but provide coarser trajectory distinctions. We therefore use throughout AD-MCQ as the operating point that preserves fine-grained speed and lateral-motion patterns without allocating a large fraction of the codebook to unsupported prototypes.

Appendix D Detailed Experimental Settings

D.1 Data Sources and AD-MCQ Construction

We construct AD-MCQ from scene–trajectory pairs drawn from Waymo Open E2E and an internal driving corpus. For each scene, every front-left, front, and front-right camera stream is decoded into four historical frames sampled at , together with the ego state, navigation context, and logged future trajectory. Qwen3-VL groups each adjacent pair of frames into one temporal patch, so the rendered prompt displays two temporal-patch timestamps (<0.2 seconds><1.2 seconds>) per camera stream even though the model input contains four decoded frames; Appendix F.1 shows the resulting prompt representation.

We then construct the three scene-disjoint downstream splits in Table 7. Training follows the natural scene distribution, whereas Dev and Test emphasize causally difficult long-tail maneuvers. Following Section 3, each question contains decoded options retrieved from a -prototype codebook. Every option is a s ego-frame trajectory with waypoints sampled at Hz. The decoded candidates are logged-data-derived waypoint prototypes.

Notation and Precomputation.

We write for the visual history, ego history and current motion state, and navigation instruction of scene . This factorization preserves three complementary signals required for driving: provides spatial coverage and short-term temporal evidence about scene dynamics, supplies the ego-motion context needed to interpret those observations, and specifies route-level intent when multiple futures are geometrically feasible. It also follows the established VLA input interface of multi-view, multi-frame images, ego-vehicle states, and high-level navigation instructions [87]. For split , Algorithm D.1 maps the input scene–trajectory pairs to , where is the shuffled six-option set and is the position of the quantized logged future after shuffling.

We precompute one similarity matrix for the entire codebook and reuse it for every scene and split. Specifically, and in Eq. 2 are the minimum and maximum over all pairwise trajectory ADE values, rather than statistics of an individual split or candidate pool. We clip the resulting similarities to and set the diagonal to one. Thus, candidate sampling only indexes the fixed row associated with the oracle token; it does not renormalize similarities per instance.

媒体内容 · 前往原文查看

Split	Num	Scene Preference	Distractors
Train		Natural distribution.	Random
Dev		stop-from-motion; hard brakes; sharp turns.	Structured
Test		straight/brake/stop; left; right.	Structured

Training Split: Scene and Candidate Construction.

We randomly sample training scenes from the natural scene distribution. For each scene, we instantiate the oracle with the nearest codebook prototype to the logged future and sample five distinct distractors at random from the hard-negative pool in Eq. 3. For the main experiments, this pool uses and . We sample uniformly without replacement within this band using a deterministic per-record random-number generator whose seed is derived from the base seed and record key as . Consequently, repeated construction with the same base seed produces the same candidates. Sampling candidates broadly within this difficulty range avoids teaching the policy a fixed distractor template.

Dev and Test: Hard-Causal Scene Construction.

Dev and Test evaluate whether a model can identify the scene evidence that causally supports a driving decision. We apply the ordered classifier in Table 8. Let and denote the current and final logged speeds, the logged-future path length, the path length under constant-velocity extrapolation, and the net heading change of the logged future. We define the braking ratio as . The minimum current-speed threshold is for Waymo and for the internal corpus. Scenes classified as low-speed or routine straight driving are discarded; the remaining stop-from-motion, hard-braking, and sharp-turn scenes form the hard-causal pool. For ranking within each retained class, we use , the endpoint gap between constant-velocity extrapolation and the logged future, and retain up to the top scenes per upstream pool and class. Each resulting question includes the constant-velocity trajectory as an explicit distractor. We retain only instances for which its quantized token is distinct from the oracle and belongs to the evaluation hard-negative pool defined below.

媒体内容 · 前往原文查看

Ordered Class	Condition	Disposition
Low speed		Discard
Stop from motion	and	Retain
Hard brake	and	Retain
Sharp turn		Retain
Routine straight	Otherwise	Discard

For Dev, we manually review the candidate pool and retain the most extreme stop-from-motion, hard-braking, and sharp-turn cases. Test broadens directional coverage with straight braking or stopping scenes, left-turn scenes, and right-turn scenes; the turns are primarily sharp. The median gap between the logged future and constant-velocity extrapolation is m on Dev and m on Test, confirming that the larger test set preserves the long-tail focus. Dev is used for the motivation study and reward development; the scene-disjoint AD-MCQ-500 Test split is used only for final evaluation.

Dev and Test: Structured Candidate Construction.

For each Dev or Test scene, we instantiate the oracle with the nearest codebook prototype to the logged future and construct five distractors as a structured mixture. All five distractors come from the split-specific evaluation hard-negative pool

(12)

We select two scale-matched hard negatives; they match the oracle’s endpoint-displacement scale while differing in trajectory shape, countering shortcuts based only on displacement magnitude. For prototype , define its scale as

(13)

We define by

(14)

Among unused prototypes in satisfying this constraint, we take the two with the smallest . Thus, “scale” denotes the Euclidean displacement of the final -s waypoint from the current ego origin, not path length or speed.

For the constant-velocity candidate, we use the recorded planar ego velocity , rather than estimating velocity from the sampled ego-history positions. With , we construct

(15)

and retrieve its nearest codebook prototype using Eq. 1, countering momentum-based extrapolation. For every retained instance, this prototype is distinct from the oracle and the scale-matched candidates and lies in . Two additional distinct samples satisfying complete the five distractors. All random draws from this band are uniform without replacement.

媒体内容 · 前往原文查看

Candidate Component	Count	Eligibility Rule
Oracle		Nearest codebook prototype to the logged future.
Scale-matched		In , with endpoint-displacement mismatch at most relative to the oracle; smallest mismatch first.
Constant velocity		In , nearest to .
General hard negative		Distinct prototypes satisfying .

Candidate Validity and Deduplication.

We maintain an exclusion set containing the oracle and every selected distractor, so scale-matched, constant-velocity, and hard-negative candidates cannot duplicate one another. Every retained Train, Dev, and Test instance has enough eligible prototypes to obtain the required five distinct distractors; no sampling fallback outside its split-specific hard-negative pool is used. We randomly shuffle the oracle and five distractors and record as the oracle’s shuffled position.

The positive option is consistent with the logged future, route intent, and map constraints, whereas negative options remain plausible explicit futures. When possible, distractors match the positive option in displacement, speed range, endpoint distance, or temporal horizon but differ in lane choice, yielding behavior, braking timing, obstacle clearance, or route compliance. We randomize option order to reduce position bias. This split-specific construction deliberately tests whether behavior learned from diverse random negatives transfers to targeted endpoint-displacement- and momentum-based distractors.

Algorithm D.1 summarizes how we convert each scene–trajectory pair into an AD-MCQ instance. The split-specific sampling operator follows the settings in Appendix D.1: Train uses five samples from the broad hard-negative pool, whereas Dev and Test use two scale-matched negatives, one constant-velocity negative, and two additional hard negatives. In the algorithm, compares the endpoint displacement magnitudes of two trajectory prototypes, as defined in Appendix D.1.

媒体内容 · 前往原文查看

Quantity	Mean	P50	P75	P90	P95
Oracle ADE ()	0.45	0.227	0.483	0.933	1.630
Oracle FDE ()	0.79	0.319	0.684	1.576	2.671
Oracle ADE / (%)	3.0	2.4	3.7	5.8	7.4
All distractors ADE ()	19.77	17.94	24.93	35.27	44.25
Nearest distractor ADE ()	7.16	7.05	8.73	10.27	11.86

Oracle Fidelity and Candidate Separation.

The oracle provides a high-fidelity representation of the logged motion, while the alternative candidates constitute geometrically distinct plans rather than duplicate quantizations or small coordinate perturbations. Table 10 supports this conclusion: the oracle prototype accurately reconstructs the logged future for most AD-MCQ-500 instances, and its displacement error is typically small relative to the scale of the trajectory space. By contrast, even the nearest distractor in each instance remains substantially farther from the logged future. Its error exceeds the oracle error in all instances, with a median oracle–distractor gap of . We do not assume that the logged future is the unique safe trajectory: AD-MCQ operationalizes planning as discrimination among explicit plan hypotheses relative to demonstrated behavior, not as an exhaustive certification of every feasible future. Nevertheless, the cross-domain gains in Table 3 and the consistent gains under resampled candidate counts and similarity bands in Figure 8 and Table 11 show that the learned capability transfers across both driving domains and candidate constructions, rather than merely recovering a fixed recorded action or exploiting one particular distractor geometry.

D.2 Evaluation Settings and Benchmarks

AD-Specific Evaluation.

All AD-specific evaluations use a -token context window and the following decoding configuration: , top-, thinking enabled, and a generation limit of tokens per turn for DEFT inference and tokens in total. For JEFT, we directly set the token budget to 24,576. We evaluate candidate selection on AD-MCQ-500 using strict option accuracy. Each reported model is evaluated eight times under the same configuration, and we report the mean over these runs. On the same scenes, we assess candidate-blind Turn-1 outputs with two complementary metrics. Normed-CFS (Normalized Causal-Faithfulness Score) is a GT-blind automatic score over the same four dimensions used in the human study: grounding, absence of hallucination, specificity, and causal coherence. The judge assigns a binary value to each dimension , and we compute . This differs deliberately from the ordinal scale used by human annotators in Section 2. Human annotators can reliably distinguish partial from full satisfaction, whereas this intermediate category is less stable for a model judge; we therefore ask the automatic evaluator only for mechanically defined binary decisions [16, 48]. We report this normalized automatic score as CFS in the result tables. HLD (High-Level-Decision Consistency) measures agreement between the predicted high-level decision and the GT trajectory, requiring both direction and speed to match the oracle action induced by the GT waypoints. Qwen3.5-397B-A17B judges these two open-ended metrics; strict candidate accuracy uses exact option matching and no LLM judge. The complete CFS and HLD evaluation-judge prompts are provided in Appendix F.6.

General-Capability Evaluation.

To measure capability retention, we evaluate benchmarks with VLMEvalKit and report four category averages: basic visual perception, embodied spatial reasoning, 3D/multi-view reasoning, and RefSpatial grounding. The category averages are computed over the following benchmark triplets: Basic Visual comprises CV-Bench-2D, CV-Bench-3D, and DA-2K; Embodied Spatial comprises EmbSpatialBench, RoboSpatialHome, and ERQA; 3D/Multi-View comprises 3DSRBench, MMSIBench, and ViewSpatialBench; and RefSpatial comprises the Location, Placement, and Unseen splits of RefSpatial-Bench. Their macro-average is reported as General AVG. All scores are percentages obtained with greedy decoding and without a chain-of-thought prompt, using the same configuration across main-table evaluations.

Cross-Domain AD Evaluation.

We construct an external -scene set from nuScenes val using the same trajectory codebook, six-candidate construction, ego-motion inputs, the two-turn deferred-exposure interface, and eight independently sampled evaluation runs, matching AD-MCQ-500. The resulting evaluation preserves the task and output contract while changing the driving domain. We use Qwen3.5-397B-A17B to judge the two external-set open metrics, matching the evaluator used for the in-domain evaluation.

D.3 Models, Distillation, and RLVR Optimization

Models.

We use Qwen3-VL-8B-Instruct and Qwen3.5-4B as the base policies. Qwen3.5-397B-A17B [46] supplies CoT annotations, and Qwen3.6-35B-A3B serves as the offline vision-language rubric generator and online text-only rubric grader. The cold-start SFT configuration is detailed in Appendix D.5.

Controlled Distillation Baselines.

DEFT Distillation (Plan Only), DEFT Distillation (Full Interaction), and DEFT Distillation (Mixed Targets) each use teacher-labeled scenes and start from the unmodified Qwen3-VL-8B-Instruct policy. Their targets are respectively all Turn-1-only, all complete two-turn, or a fixed split of the two formats. All three runs use TP, global batch size , micro-batch size , maximum sequence length , and a constant learning rate of . The visual encoder is frozen and the language model is updated.

Algorithm D.3 summarizes the controlled teacher-target construction and student fine-tuning pipeline. All variants use the same teacher-labeled scene budget and student initialization; they differ only in when candidate trajectories are exposed to the teacher and which generated turns are retained as supervised targets.

RLVR Optimization.

We optimize the complete two-turn sequence with GRPO and one trajectory-level scalar advantage shared by the generated tokens of both turns. For each optimizer step we sample rollouts for each of questions, giving trajectories per step. The actor uses a constant learning rate of without warmup. The policy loss uses token-mean aggregation and gradient clipping at , with the normalized advantage clipped to . A KL loss with coefficient regularizes the policy but is not added to the reward. Training is on-policy (one update per sampled batch) and uses no critic or entropy bonus. Rollouts are sampled with , top-, and top-. We use prompt and response limits of and tokens, respectively, with a -token generation limit per turn and a -token rollout context limit. Rollout inference uses a device-memory utilization of .

The Qwen3-VL-8B and Qwen3.5-4B experiments use the same optimization configuration. All main RLVR runs start directly from their respective base models.

Reward and Grader Configuration.

The exact verifier parses the final occurrence of FINAL_CHOICE: [A--F] and assigns one only when it matches the shuffled oracle option; malformed and incorrect responses receive zero. For DEFT-RLVR, each question has – positive atomic rubric criteria with integer weights in . Qwen3.6-35B-A3B generates these criteria offline with temperature , top-, a -token output limit, and thinking disabled. The same model grades each correct normalized Turn-1 trace online without images, candidate trajectories, the oracle, or answer letters, using temperature , top-, a -token output limit, and thinking disabled. It is served in bf16 with tensor parallelism and a -token context window. If grading fails after retry handling, we fall back to the exact-correctness reward; an incorrect final choice always receives zero.

Before either rubric-based grader is called, we isolate the candidate-blind Turn-1 trace by truncating the serialized interaction at the injected Part-2 option block and removing any residual injected option text or Turn-2 response. We then remove reasoning-wrapper tags, the HIGH_LEVEL_DECISION line, and structural headers such as PART, REASONING, and ===, while retaining the substantive evidence and causal reasoning. We denote the resulting trace by . This normalization prevents the grader from using the committed maneuver, candidate options, or final answer as a proxy for reasoning quality.

Controlled SFT comparisons use equal numbers of teacher-annotated examples. Controlled RLVR comparisons share the base VLM, training scenes, prompts, rollout budget, and core optimization settings unless explicitly stated otherwise.

D.4 Baselines and Controlled Variants

We compare DEFT-RLVR against the following task-matched baselines and controlled DEFT variants, all using the same AD-MCQ task and candidate representation.

• •

  Base VLM (Direct MCQ). We run the base model in a single turn with candidates visible from the outset and evaluate only its exact-choice accuracy.

• •

  DEFT (Training-Free). We run the same base model with our two-turn evaluation prompts: it first produces a candidate-blind plan and then selects among the revealed trajectories.

• •

  JEFT RLVR (). GRPO optimizes the single-turn JEFT response using only the binary exact-choice reward in Eq. 7. Its reasoning and selection instructions match the corresponding DEFT turns. It uses the same candidate sets, training scenes, verifier, and per-step rollout group size as the other RLVR runs.

• •

  DEFT RLVR (). This baseline uses the same candidate-blind Turn 1 and candidate-revealed Turn 2 as DEFT-RLVR, jointly optimizes both generated turns, and assigns the same exact-choice reward to their tokens.

• •

  DEFT RLVR (). This variant preserves the two-turn interface and applies a shared image-conditioned rubric to outcome-correct rollouts. We provide as defined above, together with twelve scene frames, to a Qwen3.6-35B-A3B grader. The grader is decoded with temperature , top-, a -token output limit. It returns four binary indicators , , , and for grounding, absence of hallucination, specificity, and coherence, respectively. We define the shared-rubric score as

  				(16)

  and assign the rollout reward . Thus, an invalid or incorrect final choice receives zero before rubric grading. The rubric, decision rules, and weights are fixed across scenes rather than generated per instance; the exact grader prompt is shown in Figure F.4.

• •

  JEFT Distillation. We supervise the student with teacher responses produced under JEFT.

• •

  DEFT Distillation (Plan Only). The student imitates the Qwen3.5-397B-A17B teacher’s candidate-blind Turn-1 plan on every annotated scene. At evaluation, candidate matching is elicited from the resulting policy without having been included in its SFT targets.

• •

  DEFT Distillation (Full Interaction). The student imitates both the candidate-blind plan and the subsequent candidate-selection response.

• •

  DEFT Distillation (Mixed Targets). We divide the same annotated-scene budget approximately equally between Turn-1-only and complete two-turn targets.

The three completed deferred-exposure distillation runs use the same teacher-labeled scenes, raw Qwen3-VL-8B-Instruct initialization, and optimization settings described above. All principal RLVR variants start from their respective base models; their shared GRPO configuration is described above.

Runtime Accounting.

Table 4 reports median values over deduplicated main-trainer events. All policy runs use four nodes with PPU-ZW810E accelerators per node ( GB per accelerator; accelerators in total). The actor and reference model are fully sharded over these accelerators, and policy rollouts use vLLM 0.18.0 with tensor-parallel-size- engines, bf16 inference, a -token context limit, device-memory utilization, chunked prefill, and CUDA graphs. The -only run uses this policy pool alone. Both rubric-based variants use an additional, separate pool of PPU-ZW810E accelerators for Qwen3.6-35B-A3B grading (bf16, tensor parallelism ). For , the online image-conditioned grader receives the normalized trace and scene frames; for DEFT-RLVR, the online grader receives the same normalized trace without images. Thus, the table compares end-to-end wall-clock latency under our deployed configuration, rather than total accelerator-hours.

A step comprises policy rollout, reward scoring, reference-model log probabilities, the GRPO forward/backward update, and policy-weight synchronization; validation and pre-training offline rubric generation are excluded. The Rollout column measures policy generation only, whereas Scoring includes reward queuing, communication, input parsing, and grader inference. We aggregate , , and logged steps for , , and DEFT-RLVR, respectively, merging resumed logs and retaining the latest event for each duplicated global step. No logged warm-up or anomalous steps are manually removed.

Algorithm D.4 integrates the four RLVR configurations into a shared on-policy training loop. Color-coded branches isolate their candidate-exposure interfaces and reward computations, while all black steps use the same rollout grouping and GRPO update.

D.5 Cold-Start SFT

The cold-start ablation uses deduplicated examples: causal-planning examples (an equal mixture of Turn-1-only and complete two-turn targets) and general multimodal reasoning examples relabeled by the teacher. We hold the preprocessing seed fixed at . This SFT stage freezes the visual encoder and updates the language model without resizing the original -entry vocabulary. Optimization uses a constant learning rate of with warmup steps, a global batch size of , a per-rank micro batch size of , a maximum sequence length of , and tensor parallelism of . The main run trains for five epochs, comprising optimizer steps ( per epoch), and saves every steps. For the dense initialization study, we use the same configuration for one epoch and save every steps.

D.6 Direct Trajectory-Token SFT Diagnostic

This controlled experiment underlies Figure 7 and separates three possible bottlenecks: trajectory quantization, inference over the trajectory-token vocabulary, and retention of the base VLM’s general capabilities. It is distinct from the adaptation setting in Section 5.1 and uses its own matched SFT configuration.

Model, Data, and Targets.

We extend Qwen3-VL-8B-Instruct with the trajectory tokens analyzed in Appendix C. Each token decodes to a s ego-frame trajectory containing 10 waypoints at Hz. Both settings use the same examples: Waymo-E2E training scenes and internal driving scenes. SFT wo/ CoT directly emits the oracle trajectory token. SFT w/ CoT first emits a four-part rationale—scene description, critical object, reasoning, and best action—annotated by Qwen3.5-397B-A17B with the GT action available, and then emits the same oracle token. Accordingly, this setting tests trajectory-anchored rationalization rather than the scene-first reasoning used by DEFT-RLVR.

The teacher annotation limit is tokens. Of the raw CoT annotations, () end before the requested “Best Driving Action” conclusion and () lack a parseable structured action field. Preprocessing repairs the output wrapper so that every final training target contains a closed reasoning segment and answer segment; the underlying truncation remains a limitation of this diagnostic.

Optimization.

We train both settings for 14 epochs with global batch size , micro-batch size , maximum sequence length , a constant learning rate of , and warmup steps. One epoch corresponds to 782 optimizer steps. All optimization and data settings other than the target sequence are shared.

Trajectory Evaluation.

We distinguish memorization, held-out in-domain generalization, and zero-shot cross-domain transfer. The memorization tier uses seen internal moving scenes; the in-domain tier uses Waymo-E2E validation scenes excluded from training; and the OOD tier uses nuScenes, which is absent from the training mixture. We decode each predicted special token without skipping special tokens and report ADE, FDE, and parse-failure rate over the full s horizon. The epoch curves use fixed subsets. At epoch 14, larger evaluations give m ADE for NoCoT and m for CoT on Waymo-E2E validation; the available NoCoT nuScenes endpoint is m. These larger endpoints agree with the trends in Figure 7a–b.

General-Capability Evaluation.

We use greedy decoding with one sample, no CoT prompt, and model thinking disabled. We evaluate CV-Bench-2D/3D, DA-2K, ERQA, EmbSpatialBench, RoboSpatialHome, MMSIBench, RefSpatial-Bench Location/Placement/Unseen, 3DSRBench, and ViewSpatialBench, and report the arithmetic mean of their 12 primary metrics as AVG(12).

Although training loss continues to decrease, held-out Waymo-E2E ADE reaches its minimum near epoch 8 and then rises, while zero-shot nuScenes ADE plateaus after epoch 4. The best held-out ADE is m, compared with the subset-specific m codebook quantization floor (m on the larger held-out representation split in Appendix C). Adding CoT is slightly worse in domain and provides only a small out-of-domain buffer, with additional parse failures. The two settings therefore reach similar trajectory accuracy despite substantially different general-capability retention.

For SFT wo/ CoT, the -benchmark mean falls from at epoch 1 to at epoch 8 and approaches zero thereafter. CoT slows but does not prevent forgetting: its mean decreases from to by epoch 14. These curves show that rationale supervision primarily delays destructive specialization rather than improving trajectory precision.

We additionally compare the output interfaces directly using Qwen3.5-397B-A17B on the hard-causal development set. Full-vocabulary prediction obtains approximately zero strict accuracy and m ADE, whereas six-way candidate selection reaches greedy accuracy and pass@. This is not a matched downstream evaluation, but it isolates the output interface and supports the conclusion that candidate restriction substantially reduces search difficulty.

D.7 Candidate-Set Difficulty and Construction Robustness

This diagnostic tests whether the advantage of DEFT-RLVR persists when candidate-set difficulty changes. We hold the scene, question, and oracle trajectory fixed and rebuild only the distractors. Unlike the structured Dev/Test construction in Appendix D.1, which combines scale-matched, constant-velocity, and hard-negative candidates, this diagnostic disables the first two sources and samples every distractor from the hard-negative similarity band. It therefore isolates candidate count and band width rather than reproducing the main AD-MCQ-500 candidate sets.

We vary the number of candidates as and the upper similarity bound as , while fixing . Similarity is derived from the pairwise trajectory ADE within the codebook: . For each setting, we evaluate DEFT (Training-Free), the DEFT-RLVR checkpoint, and DEFT Distillation (Mixed Targets) with two candidate-blind rounds of eight samples per question. The adapted checkpoints are DEFT-RLVR step and DEFT Distillation (Mixed Targets) iteration ; decoding follows the AD-specific evaluation configuration in Appendix D.2. One malformed item is excluded consistently, leaving paired questions.

媒体内容 · 前往原文查看

		DEFT-TF	DEFT-RLVR	Mixed Distill.
2	0.50	88.3	95.8	97.6
2	0.70	88.1	95.2	96.9
2	0.85	84.8	90.8	94.5
2	0.95	75.3	79.7	90.5
4	0.50	73.3	88.7	98.8
4	0.70	73.5	89.2	98.3
4	0.85	67.8	83.4	95.0
4	0.95	51.3	61.0	84.7
6	0.50	65.6	85.7	99.0
6	0.70	71.3	88.4	97.7
6	0.85	67.3	81.0	94.3
6	0.95	45.8	55.6	81.2
8	0.50	57.3	82.1	98.7
8	0.70	68.3	85.5	98.3
8	0.85	60.6	78.1	93.7
8	0.95	39.2	49.5	77.8
10	0.50	44.3	64.1	79.5
10	0.70	51.3	67.4	78.0
10	0.85	47.0	60.7	73.3
10	0.95	29.8	36.0	59.7

DEFT-RLVR improves over DEFT (Training-Free) in all settings, with gains ranging from to . The two-candidate setting places both models near a ceiling and offers limited discrimination. At , both models are compressed by closely matched distractors, and the gain also narrows. Even at and , both models remain above the chance level, so the hardest setting remains discriminative rather than collapsing to random choice. Intermediate thresholds exhibit small non-monotonic variation, so we do not interpret as a perfectly calibrated scalar measure of realized difficulty. DEFT Distillation (Mixed Targets) attains the highest accuracy throughout this grid, but its training targets and objective differ from those of DEFT-RLVR; the comparison is therefore descriptive rather than a controlled RL-versus-SFT attribution.

Appendix E Human Validation of AD CoT Evaluation

We validate the automatic evaluation used for the two AD CoT metrics in the main results. We first pool candidate-blind Turn-1 CoTs from all methods evaluated on AD-MCQ-500 and then randomly sample outputs from this combined pool. This audit is designed to measure human–judge agreement rather than compare individual methods, so the sample is not stratified by method. We randomly partition the sampled outputs into two disjoint subsets and assign one subset to each of two human annotators. The annotators work independently on their assigned subsets without seeing the model identity or the scores produced by the automatic judge, and each output receives exactly one human annotation. For CFS, the human annotators use the four-dimension ordinal rubric defined in Appendix B, summing grounding, absence of hallucination, specificity, and causal coherence to a score in and then normalizing it to . The automatic judge evaluates the same four conceptual dimensions but makes the binary decisions defined in Appendix F.6; its four outputs are averaged to obtain the automatic Normed-CFS in . We compute CFS agreement after placing both scores on this common normalized scale. For HLD, both the human annotator and automatic judge require the predicted direction and speed decision to agree with the action induced by the GT trajectory. We pool the resulting non-overlapping human–judge pairs to compute the agreement statistics in Table 12. The Qwen3.5-397B-A17B judge and its prompt were fixed before the human labels were examined.

媒体内容 · 前往原文查看

Metric	Agreement Measure	Result
CFS	Spearman’s
CFS	Mean absolute error
CFS	Within agreement
HLD	Exact agreement
HLD	Cohen’s

Across both metrics, the human annotations are highly consistent with the automatic scores: CFS exhibits strong rank agreement and small absolute error, while HLD decisions show high exact and chance-corrected agreement. This audit supports the use of the automatic judge for the AD CoT metrics in the main table; candidate-selection accuracy remains exact-match based and is therefore outside the scope of this validation.

Appendix F Prompt Templates

F.1 Two-Turn Candidate-Grounded Policy

Figure F.1 gives the first-turn message exactly as presented to the policy, up to example-specific variables. Each of the three video streams is decoded online into four historical frames at fps. Qwen3-VL groups adjacent frames into temporal patches, so the rendered message displays two temporal-patch timestamps, and s; these two markers still correspond to four input frames. Candidate trajectories are deliberately absent from this message. The system message is shared by both turns, and generation uses temperature , top- , and a maximum of tokens.

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.

Figure 12: Turn-2 policy chat template. The environment reveals six deterministically shuffled candidate trajectories only after the policy has committed to its first-turn plan. The parser uses the last FINAL_CHOICE field.

F.2 Offline Question-Specific Rubric Generation

For DEFT-RLVR, a fixed Qwen3.6-35B-A3B rubric generator receives twelve scene frames, the first-turn task, and the ego-state summary, but no logged future, candidate trajectory, oracle label, or statistic derived from the future trajectory. The only scalar repeated outside the first-turn task is the current ego-speed norm, computed from the current planar velocity. The generator runs once offline with temperature , top- , and a maximum of tokens. The resulting six to ten scene-specific criteria are stored with the training example and subsequently applied by the online text-only grader.

Figure 13: Offline question-specific rubric-generation chat template. Criteria are constructed solely from history-visible scene context, without the logged future, candidate trajectories, or an oracle label. Each criterion explicitly encodes a concrete scene constraint and its driving implication so that the image- and answer-blind online grader can check it from the CoT alone. The abbreviated repeated Part-1 block is exactly the text in Figure F.1.

F.3 Online Text-Only Rubric Grader

At rollout time, the grader sees only the stored criteria and the normalized first-turn CoT . It receives no image, oracle trajectory, option list, or high-level-decision line. We use temperature , top- , and a maximum of tokens.

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 (), a Qwen3.6-35B-A3B judge scores from twelve scene frames using temperature , top- , and at most tokens. This controlled variant retains the two-turn DEFT interface but replaces DEFT-RLVR’s question-specific rubric and text-only grader with a shared image-conditioned four-axis rubric.

Figure 15: Image-conditioned four-axis rubric-judge template for the controlled variant (axis weights: ).

F.5 Joint-Exposure Policy Prompt (JEFT)

Figure F.5 gives JEFT’s matched single-turn prompt: the same scene context, ego state, navigation command, candidates, reasoning requirements, and outputs as DEFT, but with all six candidates preceding both reasoning and the high-level decision. JEFT uses temperature , top- , and at most tokens, matching DEFT’s total generation budget.

Figure 16: JEFT joint-exposure policy template with all candidate trajectories visible before reasoning.

F.6 CFS and HLD Evaluation-Judge Prompts

We use the following fixed prompts to evaluate the two open-ended AD metrics reported in the main results. The CFS judge is GT-blind and evaluates the candidate-blind Turn-1 CoT against the visible scene along grounding, absence of hallucination, specificity, and coherence. The HLD judge receives the predicted high-level decision and GT maneuver and requires agreement in both direction and speed.

Appendix G Detailed Candidate-Trajectory MCQ Case Studies

Figures 17–18 show two representative DEFT-RLVR rollouts at signal- and stop-controlled intersections.

媒体内容 · 前往原文查看

媒体内容 · 前往原文查看

Appendix H Full General Visual Capability Results

Tables 13–14 expand the four category aggregates in Table 2 into the complete 12-benchmark evaluation. JEFT-based methods are task-matched conventional baselines; DEFT-based rows are variants of our framework.

媒体内容 · 前往原文查看

	Basic Visual			Embodied Spatial
Method	CV2D	CV3D	DA2K	ERQA	EmbSpat	RoboSpat
Qwen3-VL-8B-Instruct	81.88	93.83	69.10	43.00	77.75	49.14
+ JEFT RLVR ()	80.97	93.92	69.29	44.00	78.30	47.71
+ DEFT RLVR ()	81.41	93.58	69.10	44.75	78.41	48.29
+ DEFT RLVR ()	81.29	94.08	69.39	42.75	78.19	48.00
+ DEFT-RLVR (ours)	80.85	94.00	69.15	45.00	77.80	49.43
+ DEFT Distillation (Plan Only)	78.12	88.67	67.21	39.25	74.09	47.14
+ DEFT Distillation (Full Interaction)	75.96	88.50	60.35	39.00	75.69	47.43
+ DEFT Distillation (Mixed Targets)	76.25	90.83	63.10	42.50	74.04	44.00
Qwen3.5-4B	82.09	91.58	67.26	47.25	74.04	37.71
+ JEFT RLVR ()	81.55	91.58	67.55	47.75	72.83	39.43
+ DEFT RLVR ()	82.27	91.50	68.52	45.75	73.65	38.86
+ DEFT RLVR ()	81.81	92.00	67.41	47.50	73.74	38.86
+ DEFT-RLVR (ours)	82.13	92.42	68.09	50.00	73.60	42.00

媒体内容 · 前往原文查看

	3D/Multi-View			RefSpatial
Method	3DSR	MMSI	ViewSpat	RefLoc	RefPlc	RefUns
Qwen3-VL-8B-Instruct	55.21	30.70	41.54	55.00	32.00	28.57
+ JEFT RLVR ()	54.88	29.70	41.39	57.00	37.00	24.68
+ DEFT RLVR ()	54.88	31.50	42.02	56.00	43.00	33.77
+ DEFT RLVR ()	55.09	32.00	41.33	57.00	41.00	36.36
+ DEFT-RLVR (ours)	55.17	30.00	41.65	54.00	41.00	35.06
+ DEFT Distillation (Plan Only)	50.67	27.50	42.42	45.00	32.00	23.38
+ DEFT Distillation (Full Interaction)	49.62	27.30	44.22	44.00	35.00	24.68
+ DEFT Distillation (Mixed Targets)	50.42	26.80	43.38	48.00	35.00	27.27
Qwen3.5-4B	45.98	33.50	42.72	51.00	29.00	28.57
+ JEFT RLVR ()	45.69	32.10	42.68	44.00	24.00	25.97
+ DEFT RLVR ()	48.30	32.80	44.38	49.00	36.00	31.17
+ DEFT RLVR ()	46.09	32.90	43.03	55.00	31.00	29.87
+ DEFT-RLVR (ours)	46.50	33.30	43.01	53.00	34.00	20.78