Pith. sign in

REVIEW 3 major objections 5 minor 22 references

A two-stage fine-tuning pipeline teaches a 2B vision-language model to read and reason over a physics-encoded bird's-eye view, enabling it to beat a zero-shot 8B baseline on seven of eight driving-reasoning question types and lift emergency

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 07:06 UTC pith:MWJB7PJS

load-bearing objection MoRAL is a genuinely useful applied paper with a credible direction and honest limitations; the headline 2B-beats-8B result is plausible but currently rests on a single LLM judge, so it needs a second metric before the claim is solid. the 3 major comments →

arxiv 2608.02449 v1 pith:MWJB7PJS submitted 2026-08-03 cs.CV cs.RO

MoRAL: Sensor-Grounded BEV Reasoning for Compact VLMs toward Edge-Oriented Autonomous Driving

Ambarish Govindarajulu Kaliamurthi (San Jose State University) , Kaikai Liu (San Jose State University) This is my paper
classification cs.CV cs.RO
keywords autonomous drivingvision-language modelsbird's-eye viewsensor groundingmetric spatial reasoningLoRA fine-tuningLLM-as-judgenuScenes
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

MoRAL is a two-stage fine-tuning pipeline that teaches a 2B vision-language model to read and reason over a deterministic bird's-eye-view image in which LiDAR distance is encoded as color bands, object class as cluster shape, and radar Doppler velocity as directional wedges. The paper's claim is that this representational move—externalizing metric spatial perception into the input image rather than learning a 3D backbone—is what makes compact models capable of safety-critical spatial reasoning, and that a fine-tuned 2B model can therefore beat a zero-shot 8B model on this benchmark. The evidence is a four-condition comparison on 2,304 held-out nuScenes frames: the fine-tuned 2B scores 0.565 composite, 29% above the 8B baseline, wins seven of eight question types, and lifts emergency-braking recall from 10.8% to 47.8% while cutting output degeneration from 94.1% to 20.8%.

Core claim

The central discovery is that the bottleneck in VLM driving reasoning is not model size but the readability of the sensor representation. When a BEV image explicitly encodes metric distance, class, and Doppler velocity as visual features, and the model is first taught to decode that vocabulary in a dedicated grounding stage, a 2B model can reason over it more reliably than a zero-shot 8B model that receives the same image but cannot parse it. The paper shows that both zero-shot baselines fail to decode the BEV vocabulary—the 2B produces no parseable outputs, the 8B defaults to empty responses on 79.2% of frames—whereas the grounded 2B reaches a Zone F1 of 0.89 and then converts that reading

What carries the argument

The key machinery is the deterministic BEV renderer plus a two-stage LoRA fine-tuning schedule. The 896x896 px BEV image encodes: LiDAR returns colored by metric distance into six bands (yellow 0–5 m through purple 40–50 m), object class as cluster morphology (wide-dense for vehicles, elongated for barriers, compact-sparse for pedestrians), and radar Doppler velocity as filled directional wedges whose size encodes closing speed. This externalizes perception into the input, eliminating the need for a learned 3D backbone at inference. Stage 1 fine-tunes only the vision encoder (14.4M parameters) on 60,000 scan/verify/compare grounding records to teach BEV vocabulary; Stage 2 fine-tunes the ful

Load-bearing premise

The load-bearing premise is that the LLM-as-judge scores, calibrated on just 40–80 frames against human review, measure genuine reasoning quality for all eight question types rather than rewarding the structured output format the student model was explicitly trained to produce.

What would settle it

Re-run the 2,304-record evaluation with a second independent multimodal judge and a stratified human-annotated sample of at least 200 frames per condition; if the fine-tuned 2B no longer wins seven of eight question types, or if the margin collapses when the judge is blind to the [CONDITIONS]/[OBSERVE_BEV]/[PHYSICS_WORK] format, the central claim would be undermined. Also, a temporal or closed-loop test that counts phantom emergency-brake predictions (37–48% on non-emergency rows per the paper's own confusion matrices) would show whether the recall gain translates to safe behavior.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • A fine-tuned 2B model can outperform a zero-shot 8B model on an open-loop spatial-reasoning benchmark, so parameter count is not the binding constraint for this class of driving questions.
  • The pipeline fits a consumer 8 GB GPU at 42 tok/s without quantization, making this style of sensor-grounded reasoning accessible on edge hardware.
  • Emergency-braking recall rises from 10.8% to 47.8% and degeneration falls from 94.1% to 20.8%, so grounding the sensor interface changes safety-relevant behavior, not just fluency.
  • The two-stage design—vision-encoder grounding before full-model reasoning—prevents the model from copying text supplements instead of reading the image, which the authors report as an empirical failure of single-stage training.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the representational bottleneck claim generalizes, then improving the BEV rendering—e.g., denser far-range LiDAR or encoding elevation—should yield accuracy gains that scaling model size would not.
  • The reported phantom-braking pattern (37–48% false emergency-brake predictions on non-emergency frames) implies that any deployment would need temporal confirmation and threshold calibration; the authors suggest this too, but it is a direct consequence of their confusion matrices.
  • Because the evaluation relies on a single LLM judge calibrated on 40–80 frames, re-running with a second independent judge and a larger human-annotated sample would test whether the 7-of-8 win is stable across evaluators—a check the paper flags as a priority.
  • The determinism and vehicle-agnosticism of the renderer suggests the two-stage grounding recipe could transfer to other LiDAR-equipped platforms (robots, humanoids), a direction the authors mention as future work.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper presents MoRAL, a two-stage LoRA fine-tuning pipeline for Cosmos-Reason2-2B that reads a deterministic physics-encoded BEV image (LiDAR distance bands, radar Doppler wedges, ego telemetry) and answers eight open-loop driving reasoning question types. Stage 1 trains the vision encoder to decode the BEV vocabulary; Stage 2 trains the full model on 57,696 teacher-generated chain-of-thought records. On 2,304 held-out nuScenes records, a four-condition evaluation reports that MoRAL wins seven of eight question types over a zero-shot 8B baseline, raises emergency-braking recall from 10.8% to 47.8%, cuts degeneration from 94.1% to 20.8%, and runs on a consumer 8 GB GPU at 42 tok/s. The authors are candid about limitations, including the dependence on a single LLM judge and the system's unsuitability for unsupervised deployment.

Significance. If the headline results hold, this is a valuable demonstration that externalizing metric spatial perception into a deterministic BEV renderer can enable a compact 2B model to beat a zero-shot 8B model on an open-loop spatial-reasoning benchmark. The paper has notable strengths: a deterministic, reproducible BEV renderer; a four-condition controlled design; an anti-shortcut schema; objective metrics beyond the LLM judge (emergency-braking recall, degeneration, Zone F1); edge deployment profiling; and public code/checkpoint artifacts. However, the central 'seven of eight' claim rests on a single LLM judge calibrated on only 40–80 frames, with no statistical inference or second judge. The result is therefore plausible but not yet established at the level a journal evaluation requires.

major comments (3)
  1. [V] The 7-of-8 claim is carried entirely by normalized Gemma 4 judge scores, per Eq. (1), calibrated on a 40–80 frame pilot. No confidence intervals, significance tests, second judge, or inter-rater reliability are reported. The judge prompt explicitly penalizes formatting failures, and Condition C is fine-tuned to emit the ordered [CONDITIONS]...[UNCERTAINTY] structure that Conditions A/B/D largely lack; the largest claimed margins (Q6 +0.134, Q8 +0.298) are on question types where structured multi-step output is most distinctive. The score advantage may partly be a format artifact. I request item-level score distributions, paired comparisons, a second independent multimodal judge, and a scoring-side ablation that strips or controls for the tag structure before the 7-of-8 headline is supportable.
  2. [V] Even assuming the judge is valid, per-question differences are reported without uncertainty: Q2 is 0.432 vs 0.353 and Q7 is 0.509 vs 0.481, yet both are counted as wins. The 2,304 records come from 2,042 unique frames, so repeated frames create non-independence that is not addressed. Without paired statistics, clustered bootstrapping, or effect sizes with intervals, 'wins seven of eight' is not established. Please provide per-item scores and inferential tests on the question-level comparisons.
  3. [V] The safety-behavior claim is incomplete. Condition C's 47.8% emergency-braking recall is accompanied by 37–48% of non-emergency ground-truth rows being predicted as EMERGENCY BRAKE, versus 9–13% for Condition D. The paper acknowledges this phantom-braking pattern and correctly states that the system is not deployment-ready, but the highlighted 'recall improves from 10.8% to 47.8%' framing needs precision, false-positive rate, and a precision–recall or threshold curve to be interpretable as a safety improvement. As published, the recall number alone is not a usable safety claim.
minor comments (5)
  1. [Abstract] The abstract says zero-shot baselines produce no parseable BEV outputs, but Table II reports the 8B baseline parses 645/808 frames, albeit with 79.2% empty responses and only 15 distinct outputs. Clarify as 'no usable/non-degenerate outputs' to avoid contradiction.
  2. [V] The statement that Gemma's stricter rubric makes scores 'conservative' is not substantiated by a direct comparison of judge and human scores. A calibration table with score agreement and error patterns would strengthen this claim.
  3. [V] The 42-condition ablation is referenced to the companion thesis [19] but not summarized in the paper. Since the design choices (filled wedges, anti-shortcut schema) are justified by this ablation, include at least a compact summary table in the main text or an appendix.
  4. [V] Eq. (1) would benefit from a precise definition of the raw judge score scale and the rubric mapping, including whether each question type is scored with equal item weights. Currently the composite mixes question types and repeated frames without documenting weighting.
  5. [V] The within-20pct gate is 0.58, marginally below the stated 0.60 target, and training proceeds. State whether 0.60 was a pre-specified threshold and how the 0.58 result affects the interpretation of Stage 1 readiness.

Circularity Check

1 steps flagged

Primary reasoning-quality result is partially self-confirming: the Gemma judge was calibrated on a 40-80 frame pilot whose ranking already favored Condition C, and it explicitly penalizes the exact ordered-tag format the student was fine-tuned to emit.

specific steps
  1. fitted input called prediction [Section V.A, Evaluation Protocol (Table IV, Eq. 1)]
    "Gemma 4’s separated thinking mode was key to calibration: the judge’s reasoning trace was inspectable, allowing iterative rubric refinement over 40–80 frames until judge rankings matched human review. ... Because Gemma penalizes formatting failures more consistently than a human reviewer under time pressure, reported scores are conservative relative to human judgment."

    The judge is iteratively fitted until its rankings match human review on a 40-80 frame pilot, and the pilot already ranks Condition C first (Table IV: MoRAL 3.18 vs D 1.66). The same judge then produces the headline 7-of-8 and 29% composite results on 2,304 records. The rubric also penalizes formatting failures, while Stage 2 explicitly trains the model to emit the ordered [CONDITIONS]...[PHYSICS_WORK] tags; so part of the judge-based margin measures compliance with the training target, not independent physics reasoning. The headline judge-based comparison is therefore partially built into the evaluation instrument.

full rationale

The two-stage training/BEV pipeline itself is not circular: Stage 1 is evaluated on held-out grounding records, Stage 2 on held-out nuScenes validation frames, and the EB-recall, degeneration, and Zone F1 metrics are objective. The circularity is confined to the headline reasoning-quality comparison, which rests on a single LLM judge calibrated on a tiny human pilot that already favored Condition C, and whose formatting penalty aligns with the tags the student was trained to produce. The paper discloses this limitation in Section VI, but it does not eliminate the self-confirming element. Independent objective metrics support a weaker version of the claim (better EB recall, less degeneration), so the score is moderate, not extreme.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

No new physical entities are introduced. The ledger mostly records hand-chosen evaluation thresholds, hyperparameters, data-selection weights, and unverified assumptions about teacher-label correctness and judge fidelity.

free parameters (6)
  • Gemma judge rubric/calibration = 40–80 pilot frames
    Guardrails and scoring thresholds were iteratively adjusted until judge rankings matched human review; the entire 2,304-record main evaluation inherits this small-sample calibration (Section V.A).
  • Action thresholds (TTC) = EMERGENCY BRAKE <1.5s, BRAKE <3s, MONITOR <5s
    These author-chosen thresholds convert physics values into ground-truth action labels; EB recall and severity error are computed against them (Section IV.B).
  • Braking-distance coefficients = d_dry = v^2/8, d_wet = v^2/4
    A simplified vehicle model assumed for teacher labels and evaluation; no vehicle-specific calibration is provided (Section IV.B).
  • LoRA ranks / learning rates / repetition penalty = r=16 (S1); r=8/16 (S2); LR 2e-4/1e-4; penalty 1.3
    Hyperparameters chosen for validation performance; no sensitivity analysis is reported (Section IV).
  • Stage 1 frame quality weights = 3.0, 1.5, 1.0, 1.0, 1.0, 0.5
    Author-chosen weights in the quality scoring function select the top 10,000 training frames, biasing the training distribution (Section IV.A).
  • CoT rejection criteria = 11 hard criteria; 57,696 of ~70,000 kept
    Structural filters chosen by hand; no factual verification of teacher physics correctness is applied (Section IV.B).
axioms (5)
  • domain assumption nuScenes GT detections and radar quality flags are accurate enough to define ground truth for physics precomputation and EB recall labels.
    Teacher labels and evaluation action labels derive from the nuScenes 15-field schema; errors in GT propagate into both training and evaluation (Section IV.B).
  • domain assumption Gemma 4 judge scores are a valid proxy for human reasoning quality after 40–80 frame calibration.
    The main 'wins 7 of 8' claim is measured by this judge; no full multi-rater study or second judge is provided (Sections V.A, VI).
  • domain assumption Teacher Cosmos-Reason2-8B chain-of-thought is factually correct when it passes structural filters.
    The student is trained on 57,696 teacher records; filters check format, order, and leaks, not physics correctness (Section IV.B).
  • domain assumption The deterministic BEV rendering preserves the metric information needed for reasoning despite discarding the vertical axis and far-range sparsity.
    Ring accuracy collapses beyond 30 m (0.227 at 30–40 m, 0.039 at 40–50 m); the model is still evaluated on claims requiring far-range reading (Sections III.B, V.B).
  • domain assumption Author-defined TTC thresholds correctly capture safety-critical decision boundaries.
    EMERGENCY BRAKE ground truth depends on the 1.5s threshold; alternative thresholds would change all EB recall numbers (Section IV.B).

pith-pipeline@v1.3.0-daily-deepseek · 10430 in / 14302 out tokens · 152892 ms · 2026-08-04T07:06:32.184416+00:00 · methodology

0 comments
read the original abstract

Deploying vision-language models (VLMs) for safety-critical spatial reasoning on resource-constrained autonomous driving platforms requires both compact model size and reliable metric grounding. We present MoRAL (Multimodal Reasoning for Autonomous Language Models), a two-stage fine-tuning pipeline that teaches Cosmos-Reason2-2B to first read a physics-encoded Bird's Eye View (BEV) representation and then reason over it for driving decisions. The BEV image encodes LiDAR metric distance as color bands, object class as cluster morphology, and radar Doppler velocity as directional wedge overlays, externalizing spatial perception into the input image so that no learned 3D backbone is required at inference. Stage 1 fine-tunes the vision encoder on 60,000 grounding records; zero-shot baselines produce no parseable BEV outputs, confirming the vocabulary requires explicit training. Stage 2 fine-tunes the full model (52M parameters, 2.4% of total) on 57,696 chain-of-thought records generated by Cosmos-Reason2-8B as teacher, spanning eight driving question types. On 2,304 held-out nuScenes frames evaluated by Gemma 4 (31B) calibrated against human review, MoRAL wins seven of eight question types over a zero-shot 8B baseline despite using four times fewer parameters, with the largest margins on question types requiring structured multi-step physics reasoning. Emergency braking recall improves from 10.8% to 47.8%, output degeneration falls from 94.1% to 20.8%, and the full pipeline fits a consumer 8 GB GPU at 42 tok/s without quantization. These results establish a reproducible foundation for compact, physics-grounded VLM reasoning on mobile edge platforms.

Figures

Figures reproduced from arXiv: 2608.02449 by Ambarish Govindarajulu Kaliamurthi (San Jose State University), Kaikai Liu (San Jose State University).

Figure 1
Figure 1. Figure 1: MoRAL two-stage pipeline. A deterministic BEV renderer fuses [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: BEV sensor representation. LiDAR color bands encode metric distance [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Per-question score profile. MoRAL (Cond. C) dominates on Q1, Q3, [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Composite scores across four conditions. MoRAL (Cond. C, 0.565) [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: EMERGENCY BRAKE confusion matrices. Cond. D under-reacts on 60% of Q5 records (severity −0.79); Cond. C shifts toward over-reaction with 47.8% EB recall. TABLE VI EDGE DEPLOYMENT METRICS — NVIDIA RTX 4070 LAPTOP GPU Metric Value GPU VRAM 8.2 GB Precision bfloat16 (no quantization) Trainable parameters 52M / 2,152M (2.4%) Model load time 1.3 s VRAM at load / peak 4.27 GB / 4.61 GB Avg inference latency 5.7 … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

22 extracted references · 1 canonical work pages

  1. [1]

    Vision language models in autonomous driving: A survey and outlook,

    X. Zhouet al., “Vision language models in autonomous driving: A survey and outlook,”IEEE Trans. Intell. Veh., early access, doi: 10.1109/TIV .2024.3402136, 2024

    arXiv 2024

  2. [2]

    Are VLMs ready for autonomous driving? An empirical study from the reliability, data, and metric perspectives,

    S. Xieet al., “Are VLMs ready for autonomous driving? An empirical study from the reliability, data, and metric perspectives,” inProc. IEEE/CVF Int. Conf. Comput. Vis. (ICCV), Honolulu, HI, USA, Oct. 2025, pp. 6585–6597

    2025

  3. [3]

    DriveGPT4: Interpretable end-to-end autonomous driving via large language model,

    Z. Xuet al., “DriveGPT4: Interpretable end-to-end autonomous driving via large language model,”IEEE Robot. Autom. Lett., vol. 9, no. 10, pp. 8186–8193, Oct. 2024

    2024

  4. [4]

    DriveVLM: The convergence of autonomous driving and large vision-language models,

    X. Tian, J. Gu, B. Li, Y . Liu, Y . Wang, Z. Zhao, K. Zhan, P. Jia, X. Lang, and H. Zhao, “DriveVLM: The convergence of autonomous driving and large vision-language models,” inProc. Conf. Robot Learn. (CoRL), Munich, Germany, Nov. 2024, pp. 4698–4726

    2024

  5. [5]

    DriveLM: Driving with graph visual question answer- ing,

    C. Simaet al., “DriveLM: Driving with graph visual question answer- ing,” inProc. Eur. Conf. Comput. Vis. (ECCV), Milan, Italy, 2024

    2024

  6. [6]

    Open- DriveVLA: Towards end-to-end autonomous driving with large vi- sion language action model,

    X. Zhou, X. Han, F. Yang, Y . Ma, V . Tresp, and A. Knoll, “Open- DriveVLA: Towards end-to-end autonomous driving with large vi- sion language action model,” inProc. AAAI Conf. Artif. Intell., vol. 40, no. 16, Philadelphia, PA, USA, Mar. 2026, pp. 13782–13790, doi: 10.1609/aaai.v40i16.38386

    doi:10.1609/aaai.v40i16.38386 2026

  7. [7]

    DriveWorld-VLA: Unified latent-space world modeling with vision-language-action for autonomous driving,

    F. Jiaet al., “DriveWorld-VLA: Unified latent-space world modeling with vision-language-action for autonomous driving,” arXiv:2602.06521, 2026

    arXiv 2026

  8. [8]

    Talk2BEV: Language-enhanced bird’s-eye view maps for autonomous driving,

    V . Dewanganet al., “Talk2BEV: Language-enhanced bird’s-eye view maps for autonomous driving,” inProc. IEEE Int. Conf. Robot. Autom. (ICRA), Yokohama, Japan, May 2024, pp. 16345–16352

    2024

  9. [9]

    BEVDriver: Leveraging BEV maps in LLMs for robust closed-loop driving,

    K. Winter, M. Azer, and F. B. Flohr, “BEVDriver: Leveraging BEV maps in LLMs for robust closed-loop driving,” inProc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS), Abu Dhabi, UAE, Oct. 2025. 1https://huggingface.co/AmbarishGK/moral-v4-nuscenes

    2025

  10. [10]

    VLA-MP: A vision-language-action framework for multimodal perception and physics-constrained action generation in autonomous driving,

    M. Geet al., “VLA-MP: A vision-language-action framework for multimodal perception and physics-constrained action generation in autonomous driving,”Sensors, vol. 25, no. 19, p. 6163, Sep. 2025, doi: 10.3390/s25196163

    work page doi:10.3390/s25196163 2025

  11. [11]

    Reason2Drive: Towards interpretable and chain-based reasoning for autonomous driving,

    M. Nieet al., “Reason2Drive: Towards interpretable and chain-based reasoning for autonomous driving,” inProc. Eur. Conf. Comput. Vis. (ECCV), Milan, Italy, 2024, pp. 292–308

    2024

  12. [12]

    LoRA: Low-rank adaptation of large language models,

    E. J. Huet al., “LoRA: Low-rank adaptation of large language models,” inProc. Int. Conf. Learn. Represent. (ICLR), 2022

    2022

  13. [13]

    Cosmos-Reason2: Physical-AI reasoning models,

    NVIDIA, “Cosmos-Reason2: Physical-AI reasoning models,”

  14. [14]

    nuScenes: A multimodal dataset for autonomous driving,

    H. Caesaret al., “nuScenes: A multimodal dataset for autonomous driving,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Seattle, W A, USA, Jun. 2020, pp. 11618–11628

    2020

  15. [15]

    Gemma 4: Open models for reasoning,

    Google DeepMind, “Gemma 4: Open models for reasoning,” Tech. Rep., Apr. 2026. [Online]. Available: https://ai.google.dev/gemma

    2026

  16. [16]

    BEVFusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation,

    Z. Liuet al., “BEVFusion: Multi-task multi-sensor fusion with unified bird’s-eye view representation,” inProc. IEEE Int. Conf. Robot. Autom. (ICRA), London, UK, May 2023, pp. 2774–2781

    2023

  17. [17]

    OmniDrive: A holistic vision-language dataset for autonomous driving with counterfactual reasoning,

    S. Wanget al., “OmniDrive: A holistic vision-language dataset for autonomous driving with counterfactual reasoning,” inProc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), Nashville, TN, USA, Jun. 2025, pp. 22442–22452

    2025

  18. [18]

    A survey on vision-language-action models for au- tonomous driving,

    S. Jianget al., “A survey on vision-language-action models for au- tonomous driving,” arXiv:2506.24044, 2025

    Pith/arXiv arXiv 2025

  19. [19]

    MoRAL: Multimodal Reasoning for Autonomous Language Models with Sensor-Grounded Spatial BEV Rendering,

    A. G. Kaliamurthi, “MoRAL: Multimodal Reasoning for Autonomous Language Models with Sensor-Grounded Spatial BEV Rendering,” M.S. thesis, Dept. Comput. Eng., San Jos ´e State Univ., San Jos ´e, CA, USA, 2026

    2026

  20. [20]

    [Online]

    Anthropic, “Claude,” 2025. [Online]. Available: https://www.anthropic. com/claude

    2025

  21. [21]

    BEVWorld: A multimodal world simulator for au- tonomous driving via scene-level BEV latents,

    Y . Zhanget al., “BEVWorld: A multimodal world simulator for au- tonomous driving via scene-level BEV latents,” arXiv:2407.05679, 2025

    Pith/arXiv arXiv 2025

  22. [2026]

    Available: https://huggingface.co/collections/nvidia/ cosmos-reason2

    [Online]. Available: https://huggingface.co/collections/nvidia/ cosmos-reason2