DriveFine: Refining-Augmented Masked Diffusion VLA for Accurate and Robust Driving

Conference: ECCV 2026
Paper: ECCV Official
Area: Multimodal VLM
Keywords: Masked Diffusion LLM, Vision-Language-Action, Block MoE, Trajectory Refinement, Reinforcement Learning

TL;DR

DriveFine explores masked diffusion LLMs (dLLMs) for autonomous driving trajectory planning, introducing a lightweight block-level MoE to decouple parallel trajectory generation from iterative self-refinement and establishing state-of-the-art accuracy and generalization on NAVSIM and Navhard benchmarks.

Background & Motivation

End-to-end autonomous driving systems increasingly depend on generative planners to capture multimodal behaviors and explore human-preferred driving strategies. Current state-of-the-art generative planners predominantly bifurcate into two paradigms: continuous diffusion policies and discrete token-based vision-language-action models (VLAs). Diffusion policies iteratively denoise continuous action representations in parallel to produce accurate and smooth trajectories; however, their standalone diffusion decoder heads remain loosely coupled with upstream visual-language backbones. This disconnect not only demands hundreds of training epochs but also causes severe reward hacking during reinforcement fine-tuning (RFT), where optimizing targeted metrics (e.g., PDMS) precipitates catastrophic drops on extended generalization metrics (e.g., EPDMS).

In contrast, token-based autoregressive VLAs discretize continuous actions into language tokens, unifying perception, reasoning, and planning within a shared vocabulary. While this shared representation guarantees superior generalization and robustness against reward hacking, token-based VLAs lag behind diffusion planners in both inference efficiency and planning accuracy. Constrained by strict causal masking and token-by-token decoding, autoregressive generation suffers from cumulative drift errors and substantial inference latency. More critically, discrete token decoding is strictly irreversible: once an errant token (such as a slight lateral deviation or a collision-prone waypoint) is emitted, the model has no mechanism to revisit or adjust it, leading to unrecoverable trajectory failures.

The central tension lies in the complementary trade-offs between continuous diffusion's iterative refinement and discrete VLAs' robust generalizability. Masked diffusion language models (dLLMs, such as LLaDA) provide an ideal architectural bridge by combining discrete action tokenization with bidirectional attention and parallel unmasking. However, standard dLLMs remain vulnerable to irreversible decoding after unmasking. Core idea: adopt a masked diffusion LLM as the planning foundation, decouple trajectory generation and refinement via a plug-and-play block-level MoE with isolated gradients, and train the refiner using a hybrid online-offline advantage matrix under reinforcement learning to enable zero-overhead error correction and robust trajectory smoothing.

Method

Overall Architecture

The architecture of DriveFine comprises a multimodal tokenization front-end, 28 shared Transformer backbone blocks from LLaDA-8B, and a decoupled 4-block Mixture-of-Experts module at the top (partitioned into a Generation Expert and a Refinement Expert). Front-view camera images are processed by a SigLIP-384 vision encoder into continuous image tokens, which are concatenated with tokenized textual navigation instructions in a unified language representation space. Trajectory coordinates are discretized into a high-resolution action vocabulary. During planning, a fully masked sequence \([M]\) is iteratively unmasked in parallel by the Generation Expert to produce an initial candidate trajectory. Subsequently, the generated action tokens are routed to the Refinement Expert for a single-step self-correction that eliminates outliers and collision risks before execution.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Front-View Image + Textual Driving Instructions"] --> B["Multimodal Tokenization & Shared Backbone<br/>SigLIP visual encoding + First 28 shared LLaDA blocks"]
    B --> C["Block-level MoE Decoupled Architecture<br/>Final 4 blocks cloned into Generation & Refinement Experts"]
    C --> D["Generation Expert: Multi-Step Parallel Unmasking<br/>Iteratively denoises fully masked sequence [M] into candidates"]
    D --> E["Refinement Expert: Single-Step Self-Correction<br/>Directly ingests generated tokens to correct outliers & collisions"]
    E --> F["Final Robust Executable Trajectory"]

Key Designs

1. Discrete Action Tokenization with Masked Diffusion Planning: High Resolution Meets Bidirectional Context

Continuous action heads weaken foundation model alignment, while causal autoregressive decoding accumulates sequential drift. DriveFine discretizes longitudinal and lateral positions over \([-100\text{m}, +100\text{m}]\) into 4,000 bins at a fine resolution of \(0.05\text{m}\), and the heading angle over \([-90^\circ, +90^\circ]\) into 1,800 bins at \(0.1^\circ\) resolution, directly appending them to the LLM vocabulary. During training, clean action sequences are randomly masked with token \([M]\) at rate \(t\) and trained via masked cross-entropy. During inference, rather than decoding sequentially left-to-right, DriveFine applies adaptive confidence-based unmasking over multiple parallel steps, allowing every waypoint to attend to both preceding and succeeding waypoints simultaneously and eliminating directional cumulative errors.

2. Block-Level MoE Architecture: Non-Interfering Plug-and-Play Refinement

Adding conventional token-level MoE layers or forcing a single dLLM head to both unmask and refine corrupts the pretrained masked generative prior and induces severe gradient conflict between distinct objectives. Recognizing that generation and refinement share identical contextual representations, DriveFine duplicates only the final 4 transformer blocks of LLaDA while freezing the first 28 blocks as shared feature extractors. The Generation Expert strictly decodes masked positions, whereas the Refinement Expert ingests completed trajectory tokens to compute holistic sequence correction loss. Gradients from refinement are isolated within the refinement blocks. This physical decoupling preserves foundational dLLM knowledge while providing an efficient, plug-and-play trajectory editor.

3. Hybrid Online-Offline Reinforcement Fine-Tuning: Dense Advantage Optimization Without Value Estimators

The objective of the refinement expert is to reward corrections that improve driving scores while penalizing those that degrade safety, irrespective of the anchor trajectory's baseline difficulty. DriveFine eliminates the need for separate Critic value networks or baseline estimators by using the sampled trajectories themselves as references. In the RFT stage, the generation expert samples \(G\) trajectories per scenario and is optimized via standard GRPO. For the refinement expert, DriveFine computes pairwise reward differences across the \(G\) sampled trajectories to construct an offline advantage matrix \(\hat{A}^{\text{of}} \in \mathbb{R}^{G \times G}\): $$ \hat{A}{ij}^{\text{of}} = r_i - r_j, \quad \forall i, j \in {1, \dots, G} $$ This matrix has zero mean, creates symmetric gradient signals, and costs zero extra sampling. To prevent the refiner from being capped by the generator's offline distribution, the refinement expert additionally samples \(K\) refined paths online for each generated trajectory \(x_i\), computing online exploration advantages: $$ \hat{A} - r_i, \quad \forall i \in {1, \dots, G}, k \in {1, \dots, K} $$ Combining offline paired advantages with online exploratory updates trains the refiner to proactively pull boundary-violating waypoints back into safe drivable corridors.}^{\text{on}} = \hat{r}_{ik

Loss & Training

Training proceeds across two sequential stages: 1. Supervised Fine-Tuning (SFT): Trained on QA pairs and textualized trajectories from ReCogDrive for 12 epochs with batch size 64 using AdamW (initial learning rate \(4 \times 10^{-5}\) with cosine decay). The generation expert optimizes masked token cross-entropy, while the refinement expert undergoes warm-start decoding. 2. Reinforcement Fine-Tuning (RFT): Executed in the NAVSIM closed-loop simulator for 1 epoch with batch size 16 and learning rate \(1 \times 10^{-6}\). Generation group size is set to \(G = 10\), and online refinement exploration samples \(K = 6\) paths in parallel.

Key Experimental Results

Main Results

On the NAVSIM-v1 and NAVSIM-v2 benchmarks, DriveFine sets a new state of the art among VLA and end-to-end autonomous driving planners (Table 1 and Table 2 in original paper):

Method	Ref.	Sensors	NC↑	DAC↑	TTC↑	C.↑	EP↑	PDMS↑ (v1)	EPDMS↑ (v2)
Human Expert	-	-	100.0	100.0	100.0	99.9	87.5	94.8	-
UniAD	CVPR'23	Camera	97.8	91.9	92.9	100.0	78.8	83.4	-
DiffusionDrive	CVPR'25	Cam+LiDAR	98.2	96.2	94.7	100.0	82.2	88.1	88.1
AutoVLA	NeurIPS'25	Camera	98.4	95.6	98.0	99.9	81.9	89.1	-
ReCogDrive	ICLR'26	Camera	97.9	97.3	95.2	99.8	86.7	90.4	83.6
AdaThinkDrive	ICRA'26	Camera	98.4	97.8	95.2	100.0	84.4	90.3	-
CuriousVLA	CVPR'26	Camera	99.1	98.8	95.2	98.1	88.5	90.3	-
DriveFine (Base)	Ours	Camera	98.6	98.1	95.2	99.9	85.5	90.8	89.7
DriveFine* (Score-RFT)	Ours	Camera	98.8	98.6	96.2	100.0	86.9	91.8	-
DriveFine† (Best-of-6)	Ours	Camera	99.3	99.2	97.9	100.0	89.1	94.2	-

(Abbreviations: NC: No Collision, DAC: Drivable Area Compliance, TTC: Time-to-Collision, EP: Ego Progress, C.: Comfort)

On the challenging out-of-distribution Navhard benchmark (Table 3 in original paper): - Stage 1 EPDMS: ReCogDrive scores 68.9 and DiffusionDrive scores 66.7, whereas DriveFine achieves 74.4 (+5.5 points over ReCogDrive, with DAC surging from 80.2% to 90.0%). - Stage 2 EPDMS: DriveFine achieves 41.0 (vs. 37.8 for ReCogDrive and 40.5 for DiffusionDrive), demonstrating superior generalizability under extreme corner cases.

Ablation Study

Component ablations reveal steady, cumulative improvements across each module (Table 4 in original paper):

Config (ID)	SFT	GRPO	Offline-RFT	Online-RFT	NC↑	DAC↑	TTC↑	Conf.↑	EP↑	PDMS↑ (Gain)
1. SFT Baseline	✓	✗	✗	✗	98.0	95.2	94.9	99.4	81.6	86.7 (Base)
2. + GRPO Generator	✓	✓	✗	✗	98.3	96.7	95.1	99.2	85.0	89.7 (+3.0)
3. + Offline Refinement	✓	✓	✓	✗	98.5	97.3	95.2	99.9	85.4	90.4 (+0.7)
4. + Online Exploration (Full)	✓	✓	✓	✓	98.6	97.9	95.2	99.9	85.5	90.8 (+0.4)

Analysis on decoding mechanisms and efficiency (Table 6 & Fig. 7 in original paper): - Decoding Strategy: Adaptive confidence unmasking yields 89.7 PDMS, outperforming causal left-to-right (89.0 PDMS) and inverse-causal right-to-left decoding (89.2 PDMS). - Latency vs. Accuracy: With 4 diffusion steps and 1 refinement step, DriveFine achieves 90.51 PDMS with an average latency of only 280ms, vastly faster than ReCogDrive-8B (~450ms+).

Key Findings

• Error Correction via Discrete Refinement: Ablation and qualitative traces confirm that the refinement expert primarily boosts DAC and trajectory smoothness. Rogue tokens that veer into oncoming traffic or curbs during early diffusion unmasking are seamlessly shifted back into drivable corridors during the single refinement pass.
• Robust Resistance to Reward Hacking: In Table 5, diffusion planners suffer an EPDMS drop from 86.3 to 83.4 when tuned against PDMS rewards. In contrast, DriveFine exhibits aligned gains: optimizing PDMS (+3.2) concurrently raises EPDMS (+2.3 from 86.8 to 89.1), highlighting the intrinsic regularization of discrete VLA representations.

Highlights & Insights

• Masked Diffusion as a Unified Planning Backbone: Bypasses the historical dichotomy between autoregressive language models and separate continuous diffusion heads, reaping bidirectional global attention and fast parallel sampling.
• Elegant Block-MoE Task Separation: Allocating only 4 duplicated blocks to specialize in full generation vs. token-level editing creates a plug-and-play refiner without contaminating the foundation model's base language prior.
• Self-Supervised Advantage Pairing: Formulating offline relative advantages directly from generator rollouts eliminates Critic networks, providing an effective blueprint for post-training refinement policies in physical agents.

Limitations & Future Work

• Single-Step Refinement Ceiling: Refinement is empirically constrained to 1 step during inference; iterative multi-step refinement introduces diminishing returns and added latency.
• Monocular Visual Input: Relying solely on front-view camera images restricts full 360-degree situational awareness in complex intersections, which could be expanded via unified multi-camera tokenization.

Related Work & Insights

• vs ReCogDrive: ReCogDrive uses an autoregressive VLM connected to an external continuous diffusion head, which suffers from loose cross-modal coupling, training inefficiency, and generalization degradation during RFT; DriveFine maintains discrete end-to-end alignment inside a dLLM.
• vs AutoVLA / AdaThinkDrive: Autoregressive VLAs suffer from irreversible decoding and higher latency; DriveFine achieves nearly \(2\times\) faster inference via parallel unmasking and incorporates explicit post-hoc trajectory refinement.

Rating

• Novelty: ⭐⭐⭐⭐⭐ Pioneering application of masked diffusion LLMs and block-level MoE refinement to autonomous driving planners.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluations across NAVSIM v1, v2, and the challenging out-of-distribution Navhard benchmark.
• Writing Quality: ⭐⭐⭐⭐⭐ Clear exposition, thorough ablation studies, and insightful analyses of generative planning trade-offs.
• Value: ⭐⭐⭐⭐⭐ Offers a practical, robust paradigm for resolving irreversible token decoding in embodied agent planning.

Related Papers

• [ECCV 2026] ReGRPO: Reflection-Augmented Policy Optimization for Tool-Using Agents
• [ICML 2026] VLANeXt: A Recipe for Building Robust VLA Models
• [ICLR 2026] Lavida-O: Elastic Large Masked Diffusion Models for Unified Multimodal Understanding and Generation
• [ECCV 2026] MinerU-Diffusion: Rethinking Document OCR as Inverse Rendering via Diffusion Decoding
• [ECCV 2026] UniDDT: Unifying Multimodal Understanding and Generation with Decoupled Diffusion Transformer