PaperlayerView exact revision on arXiv

Source-linked AI summary

Drive the Thoughts: Runtime Monitoring of VLA Reasoning-Trajectory Consistency

Tian Yu, Lu Feng, Sebastian Elbaum

arXiv:2608.29583v1cs.SE

TL;DR

This paper asks whether driving VLA chain-of-thought traces can serve as candidate specifications for runtime verification of generated trajectories. It builds and evaluates DriveAlignBench and automated consistency monitors, finding that CoTs are useful selectively and that lane-relative F-LLM monitoring achieves the strongest reported consistency-detection performance.

  • Problem

    Driving VLA CoTs could help address the validation challenge of black-box autonomous-driving models, but they must be reliable and comparable with generated trajectories to support runtime verification.

  • Method

    The paper curates DriveAlignBench from Alpamayo 1.5, annotates 150 CoT–trajectory pairs, defines reliability criteria, and evaluates rule-based, raw-waypoint LLM, and lane-relative F-LLM monitors.

  • Results

    F1 = 0.75 for the best F-LLM monitor, improving by +0.13 absolute F1 over the strongest raw-waypoint baseline and by +0.38 over the rule-based monitor; 33.3% of CoTs were unreliable and reliable-CoT trajectories were consistent in 74% of cases.

  • Takeaways & Limitations

    CoTs can provide useful runtime specifications selectively, while CoT–trajectory inconsistency is safety-relevant but not a complete safety monitor.

  • Takeaways & Limitations

    External validity is constrained because the benchmark uses one VLA, one simulator/reconstruction pipeline, and a diversity-oriented sample rather than a naturalistic driving distribution.

Abstract

from arXiv · show

Autonomous vehicles (AVs) operate in complex environments where failures are consequential. Sophisticated machine learning models for perception and planning are key to overcoming at least part of that complexity, but their black-box nature complicates validation and verification (V&V). The recent integration of Vision-Language-Action (VLA) models into AVs introduces a unique opportunity: besides generating trajectories, these models produce an explicit Chain-of-Thought (CoT) explaining their underlying rationale. This CoT provides a rich specification to cross-check model outputs and detect inconsistencies that may expose unsafe or unintended behavior. This paper assesses whether CoTs from a recent open driving VLA can support such monitoring. We curate DriveAlignBench, a specialized dataset from NVIDIA's Alpamayo 1.5 VLA for AVs containing 150 CoT-trajectory pairs, which we manually annotate for reliability, trajectory consistency, and safety. Our analysis reveals that 33.3% of CoTs are unreliable. Among reliable CoTs, the generated trajectory is consistent with the CoT in 74% of cases. Leveraging this potential, we propose integrating a CoT-trajectory consistency check into a runtime monitor. The check is nontrivial: CoTs express open-vocabulary, scene-relative driving commitments, while trajectories are low-level ego-motion sequences whose semantics depend on road geometry and motion context. To bridge this gap, we develop a family of automated consistency monitors. Our best monitor, lane-relative F-LLM with GPT-5.5, achieves F1 = 0.75, improving over the strongest raw-waypoint LLM baseline by +0.13 absolute F1 and over a rule-based monitor by +0.38. We release DriveAlignBench, the monitor implementations, and annotation tools at https://github.com/776styjsu/drive-the-thoughts.

I. INTRODUCTION

Reasoning VLAs expose CoT traces that can serve as candidate specifications for cross-checking trajectories, but this opportunity depends on CoT reliability and semantic comparability with vehicle motion.

  • Motivation: AV black-box models complicate validation and verification because their internal representations do not use formal driving-specification concepts.These concepts include vehicles, pedestrians, lanes, right-of-way, and temporal traffic interactions.
  • Motivation: CoT traces describe intended maneuvers and scene-grounded reasons, creating a potential evidence stream for cross-checking generated trajectories.The example states a left-lane change because of slower traffic and a clear gap.
  • Research questions: The paper investigates whether CoTs are reliable and whether trajectories are semantically consistent with the actions they describe.These are the two assumptions underlying runtime use of internal reasoning for V&V.
  • Contributions: DriveAlignBench contains 150 Alpamayo 1.5 CoT–trajectory pairs annotated for reliability, trajectory consistency, and safety.The paper also releases the benchmark, monitoring architecture, and experiments.

II. BACKGROUND AND RELATED WORK

Driving VLAs combine multimodal context with explicit reasoning and future trajectories, enabling analysis of how text-described behavior relates to planned ego motion.

  • Driving VLAs: End-to-end driving systems map sensor and contextual inputs to planned future ego trajectories.Reasoning VLAs can condition on visual observations, route or language context, and ego state.
  • Driving VLAs: The studied VLA emits a text CoT describing intended driving behavior together with a trajectory whose states contain waypoint positions and headings.The monitor checks whether the generated trajectory is consistent with the CoT.
  • Related work: Recent systems use explicit reasoning for trajectory generation, including adaptive reasoning, self-reflection, physics-grounded rewards, and causal driving traces.Examples include AdaThinkDrive, AutoDrive-R2, and Alpamayo models.

C. CoT Faithfulness and Reasoning–Action Consistency

Reasoning–action consistency is difficult because textual driving commitments must be related to low-level trajectories without assuming direct predicate matching, and CoT reliability must be assessed first.

  • Faithfulness: Prior work cautions that chain-of-thought can be post-hoc rationalization, motivating direct tests of reasoning–action consistency in robotics and driving.Existing approaches examine entity fidelity, action fidelity, consistency, and perturbation sensitivity.
  • Benchmark perspective: DriveAlignBench examples distinguish reliable grounded and safe CoTs from grounding failures and unsafe plans, while withholding trajectories for reliability judgments.Reliability is evaluated from the scene and CoT alone in these examples.
  • Monitoring challenge: The proposed monitor bridges the representation gap between open-vocabulary reasoning statements and trajectory evidence instead of directly matching them to predicates or candidate actions.This complements semantic runtime monitoring and other assurance methods.
  • Specification prerequisite: A consistency monitor is meaningful only when the CoT is itself a valid specification; otherwise, inconsistency results are difficult to interpret.An unreliable rationale can be consistent with a trajectory without being a sound basis for monitoring.
  • Specification prerequisite: The paper evaluates grounding, action safety, and logical coherence as criteria for deciding whether a CoT can serve as a candidate local specification.The CoT identifies scene preconditions and justifies an action expected to satisfy safety postconditions.

B. Benchmark and Evaluation Protocol

The benchmark evaluates CoT reliability across diverse driving decisions and causal factors, finding stronger performance for routine observable behaviors and weaker performance for inference-heavy maneuvers.

  • Collection: Alpamayo 1.5 is an open 10B-parameter reasoning VLA that emits a natural-language CoT followed by a 6.4 s ego trajectory.Rollouts are collected in the released AlpaSim configuration on reconstructed scenes, with coverage-oriented scene selection.
  • Reliability findings: 100/150 CoTs (66.7%) satisfy grounding, safety, and coherence criteria, leaving 33.3% unreliable overall.Grounding failures account for 43 of the 50 unreliable CoTs, while unsafe and incoherent cases can overlap.
  • Reliability findings: CoTs are more reliable for routine, directly observable behavior than for lane topology, routing intent, or spatial and temporal gap inference.The weaker categories require additional reliability screening before CoTs are used as specifications.

IV. MONITORING FOR INCONSISTENCY DETECTION

The monitor checks whether a reliable CoT’s stated maneuver is realized by the planned trajectory, despite their different abstraction levels and semantics. This requires extracting open-vocabulary intent and incorporating context such as road geometry.

  • Reliable CoTs serve as candidate specifications whose expressed maneuvers can be checked against generated trajectories.The monitor flags plans that contradict, omit, or materially deviate from the stated maneuver.
  • Open-vocabulary maneuver extraction requires recovering the intended ego maneuver from free-form language with many possible phrasings.Examples include “move over,” “merge right,” and “position for the upcoming turn.”
  • The semantic gap arises because CoTs reference lanes, crosswalks, and obstacles, whereas trajectories provide ego-motion geometry and kinematics without those entities.Consistency may therefore require additional context, such as map information.
  • A consistency monitor compares a CoT with a transformed trajectory representation and returns CONSISTENT or INCONSISTENT.Consistency means alignment between planned ego behavior and the maneuver stated in the CoT.
  • A trajectory is inconsistent when it contradicts, omits, or fundamentally deviates from the CoT’s stated maneuver.Examples include keeping the lane instead of changing lanes or decelerating without completing a requested stop.

B. Monitoring Approaches

The monitoring approaches progressively improve comparability between CoTs and trajectories: a transparent action-label baseline, an LLM judge for open-vocabulary language, and richer trajectory evidence.

  • The three monitors progressively enrich the evidence used for CoT–trajectory consistency checking.They move from shared action labels to LLM judgment over trajectory evidence.
  • The rule-based monitor maps CoT and trajectory behaviors into a shared closed-set vocabulary and performs exact label matching.It requires every stated maneuver to appear in order as the same trajectory action label.
  • The rule-based implementation uses Alpamayo’s longitudinal and lateral meta-action vocabulary for speed and steering behaviors.The vocabulary includes actions such as stop, maintain_speed, sharp_steer_left, and go_straight.
  • The CoT parser uses 200 development-set phrases, while the trajectory parser thresholds speed, acceleration, and curvature into shared action channels.The phrase dictionary is disjoint from the evaluation set and augmented with paraphrases.

2) LLM Monitor:

The LLM and F-LLM monitors replace fixed phrase matching with semantic judgment, then add lane-relative geometry to resolve scene-relative lane intentions. This representation change is motivated by failures of raw ego-frame motion on curved roads.

  • 2) LLM Monitor:: The LLM monitor uses a prompted judge over the CoT and serialized ego-frame trajectory evidence.A fixed prompt supplies a coordinate-frame legend, few-shot examples, and a scoring rubric.
  • 2) LLM Monitor:: The LLM monitor interprets paraphrased maneuvers directly against ego-motion evidence instead of requiring predefined action labels.This addresses expressions such as “move over,” “merge right,” and “position for the turn.”
  • 3) F-LLM Monitor: Enriching the Frame of Reference:: F-LLM keeps the LLM judge unchanged and augments ego-frame evidence with lane-referenced geometry.It adds route-relative quantities for lane-centered judgments.
  • 3) F-LLM Monitor: Enriching the Frame of Reference:: Lane-relative offsets expose whether the ego stays near the lane center, drifts, nudges, or changes lanes.The route is selected from connected map candidates by minimizing projected lateral offsets.
  • 3) F-LLM Monitor: Enriching the Frame of Reference:: The lane-relative approach assumes access to high-definition maps used by the driving stack.
  • 3) F-LLM Monitor: Enriching the Frame of Reference:: In the curved-road example, raw ego-frame motion reports +35.2 m lateral displacement, whereas lane-relative offset remains within +3.7 m.The lane-relative representation more closely characterizes the behavior as lane-keeping.
  • 3) F-LLM Monitor: Enriching the Frame of Reference:: The monitor progression separates open-vocabulary extraction from the remaining representation gap for lane-centered maneuvers.F-LLM adds route-relative evidence after the LLM monitor relaxes fixed phrase extraction.
  • 3) F-LLM Monitor: Enriching the Frame of Reference:: Consistency and safety are independently labeled: a trajectory can realize a stated maneuver safely, or diverge from it and enter oncoming traffic.The figure motivates checking trajectories against the scene rather than reading consistency from CoT alone.

V. EVALUATING THE MONITORING APPROACHES

The evaluation studies the monitor approaches and the safety relevance of CoT–trajectory inconsistency in an open-loop setting.

  • The evaluation compares the monitor approaches and examines whether CoT–trajectory inconsistency is safety-relevant.

A. Research Questions

The evaluation asks how accurately monitors detect CoT–trajectory inconsistency and whether inconsistency can filter unsafe trajectories. It compares rule-based, raw-waypoint LLM, and lane-relative F-LLM monitors under a controlled evaluation protocol.

  • Research questions: The study evaluates monitor accuracy for distinguishing CONSISTENT from INCONSISTENT CoT–trajectory pairs on reliable CoTs.The reliable subset contains 100 pairs, with consistency and trajectory-safety labels annotated for each pair.
  • Research questions: The study also tests whether CoT–trajectory inconsistency is associated with trajectory safety and can filter unsafe trajectories.Safety-filter performance is measured using recall and precision for unsafe trajectories.
  • Compared monitors: Three monitors are compared: a deterministic rule-based monitor, an LLM using raw ego-frame waypoints, and F-LLM combining raw and lane-relative trajectory representations.F-LLM additionally uses a routed lane graph and lane-relative trajectory projection.
  • Compared monitors: The LLM and F-LLM monitors use Qwen3.5-4B-FP8, Kimi K2.5, or GPT-5.5 as judge models.GPT-5.5 is queried with high reasoning effort, while Qwen3.5-4B-FP8 and Kimi K2.5 are queried in thinking mode.
  • Evaluation protocol: The representative Kimi K2.5 and GPT-5.5 results use median-F1 runs across three repeats, with run-to-run F1 ranges no larger than 0.04 and 0.05, respectively.All models receive identical inputs and fixed decoding with seed = 42.
  • Evaluation protocol: Precision, recall, and F1 measure inconsistency detection, while safety filtering is evaluated through consistency–safety crosstabs.INCONSISTENT is treated as the positive class.

C. Results

Lane-relative evidence substantially improves consistency monitoring by reducing geometric false positives, while remaining errors reveal limits from ambiguous maneuver boundaries and missing scene relations. Inconsistency is safety-relevant, but consistency monitoring is not sufficient as a standalone safety monitor.

  • Detection accuracy: F1 = 0.75 for F-LLM (GPT-5.5), a +0.13 absolute gain over the strongest raw-waypoint baseline and above the rule-based monitor.The rule-based baseline achieves F1 = 0.38, detecting 9/26 inconsistencies and producing 13 false positives.
  • Representation effects: Adding lane-relative features raises F1 by +0.16 for GPT-5.5 and +0.13 for Kimi K2.5.Kimi K2.5 with lane-relative evidence reaches 0.75 versus 0.59 for GPT-5.5 with raw waypoints.
  • Judge dependence: Qwen3.5-4B-FP8 reaches recall 1.00 with F-LLM but has precision 0.37 and F1 = 0.54, showing a high-recall, low-precision tradeoff.Its lane-relative configuration identifies 26/26 raw waypoint inconsistencies, but frequent false alarms limit standalone usefulness.
  • Error analysis: F-LLM reduces false positives from 23 to 9 for Kimi K2.5 and from 24 to 7 for GPT-5.5.Lane-relative evidence distinguishes curve following, lane divergence, and clearance adjustments from sustained lane changes by incorporating route geometry.
  • Limitations: Remaining errors stem from local word matching, ambiguous maneuver boundaries, and absent agent-relative traces.Lane-relative evidence reduces geometric confusion but does not resolve surface wording, incomplete scene features, or relational claims such as maintaining distance to a lead vehicle.
  • Safety relevance: 7/9 unsafe trajectories are INCONSISTENT, while unsafe cases occur in 26.9% of inconsistent pairs versus 2.7% of consistent pairs.This supports inconsistency as a safety-relevant signal, but the remaining unsafe trajectories are consistent with their CoTs and fall outside a faithful consistency monitor’s scope.

D. Threats to Validity

The study’s validity is bounded by its data, implementation, annotation, operational definitions, and monitoring choices. Its results are conditional on this setup rather than population estimates for deployed autonomous vehicles.

  • External validity: The benchmark uses one reasoning VLA, one simulator/reconstruction pipeline, and a diversity-oriented sample rather than a naturalistic driving distribution.Simulation abstracts real-world perception noise, vehicle dynamics, and operational-design-domain constraints.
  • Internal validity: Implementation choices in trajectory parsing, route selection, lane projection, prompting, aggregation, and hosted-model serving may affect outcomes despite fixed inputs and decoding settings.Residual implementation or serving-side effects may remain.
  • Internal validity: Annotation labels remain subjective despite an explicit rubric, worked examples, and strong agreement on the doubly annotated subset.The subset shows 9/10 agreement for reliability and 10/10 agreement for consistency and safety.
  • Construct validity: Reliability, consistency, and safety are operational labels, and restricting monitoring evaluation to reliable CoTs conditions conclusions on meaningful local specifications.Alternative rubrics and binary thresholds could produce different judgments or precision–recall trade-offs.

VI. CONCLUSION

The paper finds that CoT traces can support runtime V&V only when treated as selectively reliable specifications and checked against generated trajectories. Its monitors show promise, while remaining failures motivate richer scene-relative and temporal evidence and closed-loop evaluation.

  • Conclusion: 100/150 CoTs satisfy grounding, safety, and coherence criteria, with reliability strongest for routine observable behavior and weaker for inferred topology, routing, or gaps.The conclusion recommends explicit reliability filtering before using CoTs as runtime specifications.
  • Conclusion: The authors develop rule-based, raw-waypoint LLM, and lane-relative F-LLM monitors for checking whether trajectories realize reliable CoTs.The monitors target consistency between reasoning and generated motion.
  • Conclusion: F1 = 0.75 is achieved by the best F-LLM monitor, improving over the strongest raw-waypoint LLM baseline and the rule-based monitor.The conclusion attributes the improvement to exposing scene-relative trajectory evidence to the judge.
  • Conclusion: 7/9 unsafe trajectories are INCONSISTENT, but inconsistency alone is not a complete safety monitor.Judges may over-weight action words, separate nearby maneuvers differently, or lack agent-relative traces for claims such as keeping distance.
  • Conclusion: Future monitors should add agent-relative and temporal traces, then be tested in closed loop for fallback or replanning before inconsistent trajectories become hazards.The proposed expansion targets claims such as keeping distance, yielding, and merging behind.
Loading 2608.29583v1…