Research map / Vision-language and multimodal / Robot action and navigation

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

NeurIPS 2026 · Vision-language and multimodal › Robot action and navigation · approach: Vision-language-action foundation models

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes OneVL, a latent chain-of-thought VLA for driving that uses language and visual (future-frame world-model) auxiliary decoders over latent tokens and prefill inference to match answer-only latency while beating explicit CoT on several benchmarks.

Problem

Explicit autoregressive chain-of-thought improves VLA trajectory prediction but adds prohibitive latency for real-time driving, and existing latent CoT methods underperform explicit CoT on driving tasks.

Approach

Compress reasoning into latent tokens supervised by both a language decoder (reconstructing CoT text) and a visual world-model decoder (predicting future-frame visual tokens), train with a three-stage recipe, and prefill all latent tokens in one pass at inference so latency matches answer-only prediction.

Key results

• On NAVSIM, OneVL reaches 88.84 PDM-score versus 88.29 for explicit AR CoT+Answer and 87.47 for AR Answer, at 4.46s latency versus 6.58s for AR CoT.
• On ROADWork, OneVL achieves 12.49 ADE and 28.80 FDE, versus 13.18/29.98 for AR CoT+Answer, at 4.71s latency.
• Removing staged training drops NAVSIM PDM-score from 88.84 to 67.13; an MLP regression head variant reaches 86.83 at 0.24s latency.

Measured effects reported in the paper

Surprising findings

• Latent CoT methods (COCONUT, CODI, SIM-CoT) underperform answer-only AR prediction on driving benchmarks, and OneVL is the only latent method beating explicit CoT.
• Visual future-frame auxiliary supervision contributes more (+0.87 PDM) than language auxiliary supervision (+0.31 PDM).

Limitations

• Training requires roughly 3x memory since three 4B model instances are held in memory.
• Latent token count was chosen empirically without a systematic study.
• Trajectory tokens are still generated autoregressively, so latency remains bounded by AR decoding.

Experimental setup

Datasets: NAVSIM ROADWork Impromptu Alpamayo-R1

Benchmarks: NAVSIM ROADWork Impromptu Alpamayo-R1

Metrics: PDM-score ADE FDE L2 error latency meta action accuracy STS LLM judge score

Baselines: AR Answer AR CoT+Answer COCONUT CODI SIM-CoT AdaThinkDrive LaST-VLA YNet Impromptu VLA Cosmos-Reason

Builds on: COCONUT CODI SIM-CoT Qwen3-VL IBQ

Closest papers by meaning

• LaST-VLA: Thinking in Latent Spatio-Temporal Space for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.94)
• Drive-R1: Bridging Reasoning and Planning in VLMs for Autonomous Driving with Reinforcement Learning (AAAI 2026 · similarity 0.93)
• $AutoDrive\text{-}P^3$: Unified Chain of Perception–Prediction–Planning Thought via Reinforcement Fine-Tuning (ICLR 2026 · similarity 0.91)
• AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.91)
• Explicit Geometric Chain-of-Thought for Vision-Language-Action in Autonomous Driving (NeurIPS 2026 · similarity 0.91)
• More Than Meets the Eye? Uncovering the Reasoning-Planning Disconnect in Training Vision-Language Driving Models (NeurIPS 2026 · similarity 0.91)