PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Zixuan Huang, Yang Zhou, Kaixuan Wang, Guli Zhang, Hongyan Xie, Yakun Zhu, Hao Geng, Xiaozhi Chen, Yikun Ban, Deqing Wang

arXiv:2608.01755v2cs.AI

TL;DR

GT-conditioned CoT supervision can anchor autonomous-driving VLMs to rationalize revealed future trajectories rather than infer decisions from scene evidence. The paper introduces AD-MCQ and DEFT-RLVR, which defer trajectory exposure until after a scene-derived decision and use explicit candidates for verification; DEFT-RLVR strengthens AD reasoning while preserving or slightly improving general visual capability.

  • Problem

    Existing AD VLA annotation commonly exposes teachers to GT future trajectories, leaving open whether rationales reflect scene evidence or rationalize known outcomes.

  • Method

    The paper introduces AD-MCQ for explicit candidate-trajectory selection and DEFT-RLVR, which exposes trajectories only after the policy commits to a scene-derived decision.

  • Results

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

  • Takeaways & Limitations

    AD-MCQ provides a VLM-only, scalable, and difficulty-controllable foundation for verifiable autonomous-driving reasoning through candidate construction.

  • Takeaways & Limitations

    Highly similar future trajectories and larger candidate sets remain challenging regimes for fine-grained candidate grounding.

Abstract

from arXiv · show

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

The paper identifies trajectory anchoring as a flaw in GT-conditioned CoT supervision and introduces AD-MCQ and DEFT-RLVR to make trajectory-level reasoning verifiable without open-ended synthesis. DEFT-RLVR improves autonomous-driving reasoning while preserving general visual capability.

  • Motivation: VLA annotation pipelines often expose teachers to GT trajectories, encouraging rationalization of known outcomes instead of inference from scene evidence.This is especially problematic when reasoning must resolve into a concrete driving decision.
  • Trajectory anchoring bias: GT-conditioned CoTs show lower causal faithfulness and more severe hallucinations, particularly in hard causal scenarios.The reported effect includes degradation in causal-planning quality and increased fabricated causal evidence.
  • Motivation: Hiding the GT trajectory supports a solve-then-verify process, but open-ended trajectory synthesis entangles decision making with continuous geometry and low-level dynamics.This motivates a language-model-compatible interface for AD reasoning.
  • AD-MCQ: AD-MCQ casts planning as selection among explicit scene-specific trajectories, preserving distinctions in braking time, speed profile, and lateral geometry.The benchmark makes trajectory-level decisions verifiable without open-ended geometric synthesis.
  • DEFT-RLVR: DEFT-RLVR first commits the policy to a scene-derived decision and exposes trajectories afterward as explicit grounding targets rather than pre-decision premises.This avoids replacing the GT anchor with a candidate-set anchor.
  • Results: Across multiple VLM backbones, DEFT-RLVR strengthens autonomous-driving reasoning and decision making while slightly improving aggregate general visual capability.The framework retains trajectories for exact verification and process supervision after the initial decision.

2 Trajectory Anchoring Bias in AD VLMs

The study tests whether GT future trajectories help teachers produce faithful rationales or instead anchor post-hoc explanations. Human ratings show that GT-conditioned CoTs are less causally faithful, more hallucinated, and less preferred than causal-planning CoTs.

  • Research question: The central question is whether revealing the GT trajectory improves faithful rationale inference or merely makes a known outcome easier to justify.The AD-VLM should instead infer the appropriate trajectory from historical scene evidence.
  • Trajectory anchoring: GT trajectories can anchor teachers into reverse inference, producing CoTs that match the outcome while missing the scene evidence supporting the action.Such rationales may remain geometrically consistent with the GT trajectory without being causally faithful.
  • Evaluation: Human-scored evaluation compares GT-conditioned CoTs with causal-planning CoTs using causal-faithfulness criteria.The study is summarized in Figure 2 and Table 1.
  • Findings: GT-conditioned CoTs receive lower causal-faithfulness scores, substantially more severe hallucinations, and lower pairwise preference than causal-planning CoTs.The figure breaks causal faithfulness into grounding, absence of hallucination, specificity, causal coherence, and aggregate CFS.
  • Implication: The results expose a supervision-direction mismatch: future trajectories function as reasoning shortcuts when used as premises rather than as post-decision targets.The proposed remedy is to infer a plan from the scene before grounding it in an explicit future.

3 AD-MCQ: A Verifiable Candidate-Trajectory Benchmark

AD-MCQ reformulates planning as selection from explicit trajectory candidates, making trajectory-level decisions exactly verifiable without open-ended coordinate generation. Its codebook and candidate-construction design preserve explicit geometry while enabling controllable hard negatives and scalable evaluation.

  • AD-MCQ formulation: AD-MCQ casts autonomous-driving planning as selecting one future trajectory from a small, scene-specific candidate set.Exactly one candidate corresponds to the quantized logged future, and its shuffled position is the verifiable target.
  • AD-MCQ formulation: The framework replaces open-ended coordinate generation with an exactly verifiable trajectory-selection decision.The VLM selects a shuffled option, which a deterministic verifier evaluates.
  • Trajectory codebook: K=8192 provides a favorable balance between out-of-sample reconstruction fidelity and codebook utilization.The scaling study varies clustering trajectories N and prototypes K to assess reconstruction fidelity and utilization.
  • Candidate construction: Candidate construction maps the ground-truth trajectory to a prototype, forms a hard-negative pool, and randomly shuffles M decoded candidates.An exclusion set prevents the oracle and selected distractors from duplicating one another.
  • Trajectory codebook: The codebook maps continuous futures to a finite motion vocabulary and retrieves textual waypoint trajectory candidates.This supports controlled hard-negative construction without sacrificing explicit geometry.

4 DEFT-RLVR: Deferred Exposure of Future Trajectories

DEFT-RLVR defers trajectory-candidate exposure until after scene-derived causal reasoning commits to a high-level decision. It then grounds that decision in explicit geometry and combines exact outcome verification with rubric-based supervision of correct reasoning traces.

  • Deferred exposure: DEFT-RLVR defers candidate exposure until after the policy commits to a scene-derived decision.This makes future trajectories post-decision verification targets rather than pre-decision anchors.
  • Deferred exposure: Turn 1 elicits causal reasoning solely from scene context and requires a complete high-level decision before candidate trajectories are revealed.The resulting trace connects scene evidence to driving implications.
  • Trajectory grounding: Turn 2 reveals the candidate set and uses its geometry only to identify the closest explicit realization of the recorded decision.The selected option is parsed from the second response while retaining exact trajectory-level verification.
  • Joint optimization: DEFT-RLVR jointly optimizes the two-turn interaction as a single rollout with GRPO.Each rollout serializes causal reasoning and trajectory grounding into one sequence for policy optimization.
  • Structured rubric rewards: The exact verifier supplies an outcome reward, while rubric rewards supervise reasoning traces only for rollouts with correct MCQ answers.The rubric grader evaluates normalized Turn-1 traces against instance-specific atomic criteria.
  • Structured rubric rewards: Neither the ground-truth trajectory nor candidate set enters Turn-1 causal reasoning, and the judge evaluates traces without direct access to visual inputs or trajectory options.This separates process supervision from the information used to form the initial decision.

5 Experiments

Experiments evaluate AD-MCQ and DEFT-RLVR across reasoning quality, general visual capability, transfer, ablations, and computational cost. Deferred exposure consistently improves scene-derived driving reasoning while retaining general capability and avoiding candidate-visible shortcuts.

  • DEFT-RLVR improves candidate accuracy over matched JEFT + RLVR (RMCQ) by 15.3 and 6.7 percentage points on Qwen3-VL-8B and Qwen3.5-4B, respectively.Compared with training-free JEFT, DEFT also raises ACC from 28.1% to 56.6% and from 34.0% to 65.6% across the two backbones.
  • DEFT-RLVR jointly improves MCQ accuracy, CFS, and HLD over training-free DEFT, while full-interaction distillation raises MCQ accuracy from 64.0% to 82.4%.The same distillation comparison improves CFS by 0.289 and HLD by 0.111.
  • Fine-grained trajectory grounding raises accuracy from 68.2% to 84.1%, CFS from 0.925 to 0.934, and HLD from 0.560 to 0.627 over Plan Only.The comparison uses DEFT Distillation (Mixed Targets) with full two-turn targets.
  • On OOD nuScenes scenes, DEFT-RLVR raises candidate accuracy from 39.6% to 49.5% while improving CFS by 0.114 and HLD by 0.073.
  • DEFT-RLVR raises general visual capability from 54.81% to 56.09% on Qwen3-VL-8B and from 52.56% to 53.24% on Qwen3.5-4B.The passage contrasts these RL-based gains with the capability decline observed under dense SFT imitation.
  • Compared with JEFT + RLVR (RMCQ), DEFT improves accuracy from 61.1% to 76.4% and from 72.3% to 79.0%, with higher CFS and HLD.Figure 6 also shows substantially larger policy drift for JEFT.
  • Rubric supervision improves reasoning and HLD quality by steering responses toward greater faithfulness and suppressing unproductive exploration.Without rubric supervision, DEFT + RLVR (RMCQ) produces longer, higher-entropy responses.

6 Conclusion

The paper identifies trajectory anchoring in AD VLM reasoning and introduces AD-MCQ and DEFT-RLVR to make trajectory-level decisions verifiable without open-ended coordinate generation. The resulting framework improves generalizable AD reasoning while preserving or enhancing general visual capabilities.

  • The framework improves generalizable AD reasoning while preserving and even enhancing general visual capabilities.
  • Because AD-MCQ uses VLM-only inference and controllable option construction, it provides a deployable and scalable foundation for verifiable AD reasoning research.
  • The work identifies anchoring bias when logged future trajectories become premises for post hoc driving rationales.
  • AD-MCQ uses scene-specific decoded trajectory candidates to preserve braking, speed, and lateral distinctions while converting planning into exact candidate selection.
  • AD-MCQ is a verification interface rather than a replacement for a downstream continuous planner.

A.3 Verifiable Post-Training and Faithful Reasoning

This section motivates verifiable post-training by separating scene-based decision formation from future-trajectory verification. It evaluates GT-conditioned annotation with controlled paired prompts and human scoring of causal faithfulness.

  • Reinforcement learning with automatically checkable outcomes can elicit reasoning capabilities without imitating every intermediate step.
  • Revealing an answer before reasoning can distort subsequent reasoning, causing models to rationalize cues or hallucinate support for favored conclusions.
  • The study tests whether logged GT futures improve faithful rationales or merely make known outcomes easier to justify.
  • The evaluation uses 100 strong-causal scenes, including 70 Waymo scenes and 30 internal scenes, with paired annotations from the same teacher.
  • The paired comparison changes only future-trajectory availability while holding prompts, inputs, model, decoding, and response format fixed.
  • The causal-planning template gives the teacher scene history and ego state but no logged future.
  • The GT-conditioned template differs through a highlighted raw-waypoint GT block, without adding derived action labels or rationalization-specific instructions.

B.2 Human Evaluation

Human evaluation scores causal faithfulness through four independently rated dimensions, while the appendix details trajectory-codebook construction, reconstruction behavior, and candidate coverage. The analyses establish both the evaluation protocol and the resolution–coverage trade-off.

  • B.2 Human Evaluation: Two annotators score 200 CoTs across grounding, hallucination absence, specificity, and causal coherence using a three-point ordinal scale.
  • B.2 Human Evaluation: Each CoT receives a causal-faithfulness score from 0 to 8 by summing the four averaged dimension scores.
  • B.2 Human Evaluation: Dimension criteria distinguish unsupported scene claims, fabricated details, and incomplete causal links from fully supported reasoning.
  • B.2 Human Evaluation: Annotators evaluate anonymized, randomized single-CoT items without seeing paired counterparts, labels, the anchoring hypothesis, or GT futures.
  • B.2 Human Evaluation: Pairwise evaluation presents the two scene-matched CoTs together and records which rationale is better or whether they tie.
  • C Trajectory Codebook Construction: The trajectory codebook aggregates motion trajectories independently of scene observations and is used only to retrieve candidates; its indices are never exposed to the policy.
  • C Trajectory Codebook Construction: Each logged future is a 5 s ego-frame trajectory sampled at 2 Hz with 10 waypoints, clustered by cumulative squared waypoint distance.
  • C Trajectory Codebook Construction: Symbolic codebook identifiers decode to complete waypoint sequences, but AD-MCQ serializes decoded coordinates rather than exposing identifiers to the VLM.

C.4 Data Scale and Generalization

Increasing the trajectory codebook and clustering data improves reconstruction fidelity, especially for rare motions, but larger codebooks trade utilization for finer distinctions. K=8192 is selected as a balance among fidelity, utilization, and data support.

  • Increasing the number of clustering trajectories closes the gap between in-sample and held-out reconstruction error.The gap narrows as each prototype receives more trajectories, with convergence once trajectory support is sufficiently large.
  • For K=1024, the reconstruction gap is effectively closed near N≈10^5, while K=8192 reaches a 0.007 m gap at full data scale.
  • At full data scale, ADE p99 falls from 2.74 m at K=256 to 1.14 m at K=16384, while FDE p99 falls from 5.82 m to 2.33 m.Larger codebooks improve rare-motion reconstruction but do not eliminate long-tail error.
  • K=8192 achieves 0.290 m out-of-sample ADE, 0.516 m out-of-sample FDE, and 92.0% utilization, balancing fidelity with codebook support.Larger codebooks improve displacement error but reduce out-of-sample prototype utilization by more than 18%.

D.1 Data Sources and AD-MCQ Construction

AD-MCQ combines scene-derived inputs with codebook-based trajectory candidates and structured hard negatives, emphasizing causally difficult scenes. Its evaluation separates candidate-blind reasoning from later trajectory verification and reports robustness across settings.

  • Data Sources and Benchmark: AD-MCQ uses four historical frames at 2 Hz from three camera streams, together with ego state, navigation context, and logged future trajectory.The visual history, ego history, motion state, and navigation instruction provide complementary spatial, temporal, motion, and route-level signals.
  • Candidate Construction: Candidates are sampled deterministically from fixed codebook similarities, with the oracle nearest the logged future and distractors constrained to be distinct hard negatives.Repeated construction with the same seed is reproducible, while broad sampling avoids a fixed distractor template.
  • Candidate Construction: Train uses five random hard negatives, whereas Dev and Test use two scale-matched, one constant-velocity, and two additional hard negatives.The structured candidates preserve explicit distinctions in trajectory shape and difficulty across splits.
  • Hard-Causal Scene Construction: Hard-causal Dev and Test scenes are selected through an ordered classifier using speed, braking, path-length, and heading-change cues.The resulting sets include stop-from-motion, hard-braking, and sharp-turn cases with large logged-future versus constant-velocity gaps.
  • Oracle Fidelity and Candidate Separation: The oracle prototype reconstructs logged motion more closely than every nearest distractor, with a median oracle–distractor gap of 6.74 m across 500 instances.AD-MCQ evaluates discrimination among explicit plans relative to demonstrated behavior rather than exhaustive certification of every feasible future.
  • Training and Evaluation: Candidate-blind Turn-1 traces are normalized before grading, excluding options, committed maneuvers, final answers, and structural wrappers from the reasoning-quality input.The diagnostic also reports 6.8% truncated raw CoTs and 16.1% without parseable action fields, while DEFT-RLVR outperforms DEFT (Training-Free) by 4.4%–24.8% across 20 settings.
  • Evaluation Diagnostic: Candidate restriction reduces output search difficulty: six-way selection reaches 0.61 greedy accuracy and 0.88 pass@64, unlike full-vocabulary prediction with approximately zero strict accuracy and 2.5 m ADE.This comparison is an interface diagnostic rather than a matched downstream evaluation.

E Human Validation of AD CoT Evaluation

A human audit finds strong agreement between automatic AD reasoning scores and human annotations, supporting the automatic judge for CoT metrics while excluding exact-match candidate accuracy from this validation.

  • The audit samples 200 candidate-blind AD reasoning outputs pooled across methods and assigns disjoint subsets to two independent annotators.Annotators evaluate outputs without seeing model identities or automatic scores.
  • Human annotations show strong rank agreement and small absolute error for CFS, alongside high exact and chance-corrected agreement for HLD.The results support using the automatic judge for the AD CoT metrics in the main table.
  • Candidate-selection accuracy is outside the audit’s scope because it is evaluated by exact match rather than the validated human–judge scoring process.

F Prompt Templates

The appendix specifies DEFT-RLVR’s two-turn candidate-blind-then-candidate-exposed interface, offline scene-only rubric generation, text-only grading, controlled image-conditioned grading, and JEFT’s joint-exposure comparison.

  • Turn-1 Candidate-Blind Policy Prompt: Turn 1 presents scene frames, ego state, and navigation command without candidate trajectories, preserving the production message order.Turn 2 reveals six deterministically shuffled candidates only after the policy commits its first-turn plan, and parsing uses the last FINAL_CHOICE field.
  • Offline Question-Specific Rubric-Generation Prompt: Offline rubric generation uses only history-visible scene context and produces concrete scene constraints with driving implications for later grading.The generator receives no logged future, candidates, oracle label, or future-derived statistic.
  • Online Text-Only Rubric-Grader Prompt: The online text-only grader sees stored criteria and the normalized first-turn CoT, returning binary criterion decisions whose weighted present rate forms the process score.It receives no image, oracle trajectory, option list, or high-level-decision line.
  • Image-Conditioned Rubric-Judge Prompt: The controlled DEFT+RLVR variant replaces question-specific text grading with an image-conditioned four-axis judge using weights 0.30/0.30/0.10/0.30.The judge evaluates the scene and one candidate CoT without being told the correct maneuver.
  • JEFT Joint-Exposure Policy Prompt: JEFT exposes all six candidates before both reasoning and the high-level decision while matching DEFT’s scene context, requirements, and total generation budget.This provides a joint-exposure comparison against deferred exposure.
  • CFS and HLD Evaluation-Judge Prompts: CFS is GT-blind and scores grounding, hallucination absence, specificity, and coherence, whereas HLD checks predicted direction and speed against the ground-truth maneuver.The HLD judge evaluates agreement with ground truth rather than safe or legally preferred driving.

G Detailed Candidate-Trajectory MCQ Case Studies

Figures 17–18 present two representative DEFT-RLVR rollouts at signal- and stop-controlled intersections. In each case, the policy commits to the visible control before selecting the matching stopping trajectory.

  • Figures 17–18 show representative DEFT-RLVR rollouts at signal- and stop-controlled intersections.
  • In the red-light case, the policy identifies the signal before seeing options and matches that commitment to stopping trajectory A.
  • In the stop-controlled case, the candidate-blind plan follows the visible stop control despite a constant-speed history, then selects trajectory C.

H Full General Visual Capability Results

Tables 13–14 provide the complete 12-benchmark evaluation, comparing task-matched JEFT-based conventional baselines with DEFT-based framework variants. DEFT-RLVR generally preserves or improves general visual capabilities, with gains varying across backbones and capability groups.

  • Tables 13–14 expand the four category aggregates in Table 2 into a complete 12-benchmark evaluation of JEFT and DEFT methods.
  • Basic Visual and Embodied Spatial capabilities: DEFT-RLVR improves five of six Basic Visual and Embodied Spatial benchmarks for both backbones, with mean gains of 0.26% for Qwen3-VL-8B and 1.39% for Qwen3.5-4B.The largest gains are 2.00% on ERQA and 4.29% on RoboSpat, respectively.
  • 3D/Multi-View and RefSpatial capabilities: For 3D/Multi-View and RefSpatial capabilities, DEFT-RLVR improves Qwen3-VL-8B’s RefSpatial mean by 4.83% while largely preserving Qwen3.5-4B’s six-benchmark mean at −0.03%.Qwen3-VL-8B’s 3D/Multi-View mean changes by −0.21%, while Qwen3.5-4B changes by +0.20% on 3D/Multi-View and −0.26% on RefSpatial.
Loading 2608.01755v2…