OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation

Abstract: Chain-of-Thought (CoT) reasoning has become a powerful driver of trajectory prediction in VLA-based autonomous driving, yet its autoregressive nature imposes a latency cost that is prohibitive for real-time deployment. Latent CoT methods attempt to close this gap by compressing reasoning into continuous hidden states, but consistently fall short of their explicit counterparts. We suggest that this is due to purely linguistic latent representations compressing a symbolic abstraction of the world, rather than the causal dynamics that actually govern driving. Thus, we present OneVL (One-step latent reasoning and planning with Vision-Language explanations), a unified VLA and World Model framework that routes reasoning through compact latent tokens supervised by dual auxiliary decoders. Alongside a language decoder that reconstructs text CoT, we introduce a visual world model decoder that predicts future-frame tokens, forcing the latent space to internalize the causal dynamics of road geometry, agent motion, and environmental change. A three-stage training pipeline progressively aligns these latents with trajectory, language, and visual objectives, ensuring stable joint optimization. At inference, the auxiliary decoders are discarded and all latent tokens are prefilled in a single parallel pass, matching the speed of answer-only prediction. Across four benchmarks, OneVL becomes the first latent CoT method to surpass explicit CoT, delivering state-of-the-art accuracy at answer-only latency, and providing direct evidence that tighter compression, when guided in both language and world-model supervision, produces more generalizable representations than verbose token-by-token reasoning. Project Page: https://xiaomi-embodied-intelligence.github.io/OneVL

• The paper demonstrates that OneVL’s dual decoder strategy for simultaneous visual and language supervision significantly improves causal reasoning and efficiency over traditional CoT methods.
• The method employs a rigorous three-stage training pipeline to stabilize latent representations, ensuring robust spatial-temporal prediction and minimizing inference latency.
• OneVL outperforms existing models on multiple autonomous driving benchmarks, achieving state-of-the-art metrics such as PDM-score, ADE, and FDE while maintaining real-time performance.

OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation

Introduction and Motivation

OneVL introduces a transformative approach to reasoning and planning in Vision-Language-Action (VLA) models tailored for autonomous driving applications. The paper addresses the main bottleneck of existing Chain-of-Thought (CoT) reasoning in vision-LLMs (VLMs): the inefficiency of explicit, autoregressive generation of reasoning tokens, which incurs substantial inference latency unsuited for real-time domains such as driving. Standard latent CoT methods, which compress reasoning into continuous representations, consistently perform worse than explicit CoT because their latents encode only symbolic abstractions of the environment, insufficient for capturing physical scene dynamics.

The central hypothesis behind OneVL is that a multimodal compression of both language reasoning and future visual states—explicitly supervising the latent space not just with language but also with visual world-model targets—forces the model to internalize causal structure instead of symbolic shortcuts. The paper's core contribution is the development and empirical validation of a single-step, jointly-supervised latent reasoning approach that offers both state-of-the-art accuracy and interpretable explanations at inference latency matching answer-only prediction.

Limitations of Prior Latent CoT Paradigms

Classic latent CoT approaches such as COCONUT, CODI, and SIM-CoT aim to distill explicit reasoning into latent vectors. However, their supervision is fundamentally text-centric; the resulting latents are unable to encode the multi-modal causal dynamics critical for trajectory prediction. The authors demonstrate—quantitatively and qualitatively—that these methods underperform explicit CoT on every tested autonomous driving benchmark. The issue is twofold: first, language-based latent supervision induces abstraction without ensuring causal scene grounding, and second, these methods retain iterative AR computation for latent token generation, thus not alleviating the latency problem.

Figure 1: Accuracy and efficiency comparison across four benchmarks. Existing latent CoT methods underperform explicit CoT. OneVL is the first to surpass it while matching answer-only prediction latency.

Figure 2: Comparison of three CoT paradigms—explicit, implicit, and OneVL—which describes how OneVL achieves both interpretability and efficiency by using dual auxiliary decoders and single-step latent reasoning.

OneVL: Unified Multimodal Latent Reasoning

OneVL's architecture is centered around dual-modal auxiliary decoders positioned after specialized language and visual latent tokens. The visual auxiliary decoder plays the role of a world-model head: latent tokens are forced to predict future visual frames, which imposes a spatial-temporal causal constraint on the latent space. The language auxiliary decoder reconstructs human-readable reasoning traces from the language latents, providing semantic alignment and interpretability. Both decoders are used only at train time. At inference, all latent tokens are prefilled in one parallel pass and only the answer (trajectory) is decoded autoregressively, matching answer-only AR latency.

Figure 3: OneVL architecture. An image and structured text prompt are processed to produce specialized latent tokens; dual auxiliary decoders train these to decode future frames and CoT text, ensuring causal and semantic coverage.

Three-Stage Training Pipeline

The model is optimized using a staged curriculum:

1. Visual Auxiliary Decoder Pretraining: The visual decoder is trained as an unconditional next-frame predictor from ViT features for action-independent world modeling.
2. Stage 0 (Main Model Warmup): The backbone VLM is trained with embedded latent tokens in the output sequence for standard trajectory prediction, encouraging meaningful latent positions.
3. Stage 1 (Auxiliary Decoder Alignment): With the backbone frozen, the auxiliary decoders learn to map latent-token features to their world-model and CoT targets.
4. Stage 2 (Joint End-to-End Fine-Tuning): All components are fine-tuned with combined trajectory, language, and visual reconstruction losses, ensuring the latent bottleneck aligns with all objectives.

This curriculum avoids gradient explosion, local minima, and catastrophic collapse of the latent representations—conditions that cripple end-to-end direct optimization, as rigorously shown in the ablation analysis.

Figure 4: Visual CoT under full training (top row) vs. direct joint optimization (bottom). Without staged training, the visual decoder collapses to memorized, irrelevant images; full training yields scene-consistent predictions used as robust latent supervision.

Experimental Results

OneVL is evaluated across four established and challenging autonomous driving benchmarks: NAVSIM, ROADWork, Impromptu, and APR1. The results are unambiguous:

• OneVL is the only latent CoT framework to consistently outperform explicit AR CoT across all metrics and benchmarks, achieving state-of-the-art PDM-score, ADE, and FDE with or below the latency of answer-only AR prediction.
• On NAVSIM, OneVL attains 88.84 PDM-score (vs. 88.29 for AR CoT+Answer and 87.47 for AR Answer) at 4.46s average latency—matching the speed of answer-only prediction and $1.5\times$ faster than explicit CoT.
• On ROADWork, OneVL yields 12.49 ADE / 28.80 FDE, dominating prior approaches; similar superiority is shown in Impromptu and APR1.

Ablation studies isolate the quantitative contributions of each module: the visual auxiliary decoder yields a +0.87 PDM-score gain, the language decoder adds +0.31, while omitting staged training catastrophically drops PDM-score by over 21 points, establishing its necessity.

Qualitative outputs from the language decoder show competitive chain-of-thought reasoning with high meta action and semantic similarity accuracy against ground truth, approaching AR CoT+Answer while maintaining low latency.

Figure 5: Visualizations of prediction on NAVSIM, overlaying ground-truth and predicted trajectories, demonstrating the spatial-temporal interpretability of OneVL.

Analysis and Theoretical Implications

The evidence validates key theoretical claims regarding multimodal compression:

• Compression with causal supervision via world-model targets yields more generalizable latent representations than purely symbolic (language-only) compression.
• Visual future-frame prediction as a supervision target is a stronger constraint for planning tasks than language reasoning annotation alone due to its direct alignment with the physical scene and agent dynamics.
• The prefill parallel latent generation mechanism harnesses modern Transformer architectures' parallelism to completely eliminate the main latency cost of explicit or implicit AR reasoning.
• Failure of prior language-only latent CoT approaches is rooted in misalignment between supervision signal and downstream task—language abstractions do not encode the necessary geometric precision for trajectory prediction.

  Figure 6: Visualizations of prediction on ROADWork, highlighting the model's performance in complex spatial arrangements (e.g., construction zones).

  

  Figure 7: Visualizations of prediction on Impromptu, showing the model’s robustness in unstructured environments and corner-case handling.

Deployment and Practical Considerations

The OneVL framework is designed for real-world deployment, offering a regression MLP head to replace AR decoding of the trajectory and further reduce end-to-end latency to 0.24s (4.16 Hz), at a modest accuracy cost. This presents a viable path towards closed-loop, real-time planning in safety-critical systems. Additional design constraints, such as memory consumption due to triple-model instantiation during training, are mitigated with infrastructure techniques but warrant continued study.

Conclusion

OneVL sets a new state-of-the-art for efficient, interpretable trajectory prediction in VLA-based autonomous driving. It is the first latent CoT framework to decisively outperform explicit AR CoT, validating that causal, world-model grounded compression is not merely an efficiency fix but crucial for robust reasoning. The integration of dual auxiliary decoders ensures that the latent bottleneck aligns with both semantic and spatial-temporal dynamics, and the staged training framework is essential for optimization stability.

The results have theoretical and practical implications for how future VLMs and VLAs internalize reasoning—suggesting that robust, efficient planning arises not from symbolic compression alone, but from world-model grounded alignment of the latent space. Extensions towards non-AR decoding of answers, 360-degree scene modeling, and richer human-machine explanations are natural next steps for safety-critical, transparent autonomous systems.

Figure 8: Visualizations of prediction on APR1, demonstrating generalization and interpretability in high-complexity causal reasoning and action prediction.