ReflectDrive: Safety-Aware Diffusion Planning

• ReflectDrive is a discrete-diffusion vision-language-action framework that tokenizes trajectories for end-to-end autonomous driving planning.
• It combines goal-conditioned trajectory generation with a local discrete search and inpainting mechanism to iteratively repair unsafe waypoints.
• Empirical results show that iterative reflection significantly improves safety metrics, including drivable-area compliance and overall performance on NAVSIM.

Searching arXiv for ReflectDrive and closely related driving-planning papers. ReflectDrive is a discrete-diffusion Vision-Language-Action planning framework for end-to-end autonomous driving that combines goal-conditioned trajectory generation with an inference-time reflection mechanism for safety-aware self-correction. It formulates planning as conditional trajectory generation $p(\tau \mid c)$, where $\tau$ is the future ego trajectory and $c$ is the scene context, and it represents trajectories as discrete coordinate tokens so that unsafe waypoints can be detected, locally repaired, and regenerated by inpainting without gradient computation (Li et al., 24 Sep 2025). Within the recent VLA literature, ReflectDrive is situated among methods that use pre-trained multimodal models for planning, but it differs from purely imitation-learned planners by introducing an explicit safety-aware reflection loop at inference time and from gradient-guided diffusion methods by using local discrete search, masking, and inpainting instead of reward gradients through denoising (Li et al., 24 Sep 2025).

1. Definition and conceptual framing

ReflectDrive is introduced as a response to a specific limitation of imitation-learned autonomous driving planners: highly probable trajectories under the learned policy need not satisfy hard driving constraints such as collision avoidance and drivable-area compliance (Li et al., 24 Sep 2025). The method is built around three linked ideas: discretization of the two-dimensional driving space into an action codebook, fine-tuning of a pre-trained Diffusion LLM for conditional trajectory-token generation, and a reflection mechanism that iteratively detects unsafe waypoints, repairs them by local search in token space, and inpaints the surrounding trajectory (Li et al., 24 Sep 2025).

In the paper’s terminology, “reflection” refers to iterative self-correction through interaction between the generative planner, an external safety evaluator, local discrete search, and inpainting-based regeneration (Li et al., 24 Sep 2025). This is distinct from standard sampling, which emits trajectories without external correction, and from gradient-guided diffusion, which modifies samples through reward gradients during denoising (Li et al., 24 Sep 2025). It is also distinct from stronger forms of reflective reasoning based on natural-language self-critique; ReflectDrive’s reflection is trajectory-level and operational rather than symbolic. This suggests a narrower but more directly actionable notion of reflection: the planner reflects on safety violations in its own output and re-enters generation with edited anchors.

ReflectDrive belongs to the broader class of discrete or masked-diffusion driving planners. The later paper “Discrete Diffusion for Reflective Vision-Language-Action Models in Autonomous Driving” explicitly positions ReflectDrive in that family and frames its contribution as a scalable and reliable solution for safety-critical trajectory generation on NAVSIM (Li et al., 24 Sep 2025). Related work extends or parallels this line in different ways. ReflectDrive-2 makes reflection native to the action-token space by using AutoEdit for token-to-token rewriting under reinforcement learning (Wang et al., 6 May 2026), while DriveFine adds a dedicated refinement expert on top of a masked diffusion VLA to perform a post-generation corrective pass (Dang et al., 16 Feb 2026). These later developments indicate that ReflectDrive’s central theme—explicit correction of an initial plan inside a discrete token space—became a reference design pattern for reflective driving planners.

2. Trajectory tokenization and discrete diffusion backbone

ReflectDrive discretizes future trajectories waypoint by waypoint. A future trajectory $\tau$ consists of $N$ waypoints in ego coordinates, and each waypoint $(x_j,y_j)$ is quantized independently in $x$ and $y$, producing a token pair $(\mathbf{y}_{j,x}, \mathbf{y}_{j,y})$ from a uniform one-dimensional codebook $\mathcal{A}$ defined over the spatial range $\tau$0 with resolution $\tau$1 (Li et al., 24 Sep 2025). The full trajectory becomes a flattened discrete sequence

$\tau$2

The appendix specifies the spatial range as $\tau$3, and reports answer length $\tau$4 and block length $\tau$5, implying a fixed-length token sequence for planning (Li et al., 24 Sep 2025).

This tokenization is foundational because it turns local safety repair into a search problem in discrete space. The paper acknowledges that discretization may reduce precision, but argues that it makes feasible alternatives searchable in Bird’s-Eye-View token space and naturally supports masking and inpainting (Li et al., 24 Sep 2025). A plausible implication is that ReflectDrive trades some continuous fidelity for editability and structured correction.

The generative backbone is a pre-trained Diffusion LLM initialized from LLaDA-V and fine-tuned by supervised denoising on autonomous-driving planning data (Li et al., 24 Sep 2025). The scene context $\tau$6 includes multi-view images, language instructions, and ego-vehicle state; in experiments, the inputs are front, front-left, and front-right camera images, a high-level navigational language command, and textual descriptions of ego status (Li et al., 24 Sep 2025). Given a clean token sequence $\tau$7, the forward discrete diffusion process progressively masks a subset of tokens over $\tau$8 timesteps, yielding a corrupted sequence $\tau$9, and the reverse model learns masked-token reconstruction: $c$0 with training objective

$c$1

Only masked positions contribute to the loss, and inference begins from a fully masked sequence and uses parallel decoding with bidirectional feature fusion and low-confidence re-masking (Li et al., 24 Sep 2025). The appendix specifies diffusion steps $c$2, remask policy $c$3 low-confidence, and reports use of classifier-free guidance, though the guidance formula is not given (Li et al., 24 Sep 2025).

3. Goal-conditioned generation and multimodal behavior modeling

ReflectDrive does not rely solely on local repair to explore behavioral alternatives. It first performs goal-conditioned trajectory generation so that the planner can model multimodal behaviors at the level of terminal intentions (Li et al., 24 Sep 2025). The paper argues that local safety correction alone cannot induce large-scale changes such as selecting a different turn direction; therefore the planner first predicts terminal waypoint tokens

$c$4

where $c$5, and then extracts several high-probability goal candidates (Li et al., 24 Sep 2025).

The selected goals are defined by Top-$c$6 selection followed by non-maximum suppression: $c$7 producing $c$8 spatially diverse goals $c$9 (Li et al., 24 Sep 2025). The appendix sets the number of goal candidates to $\tau$0 and the NMS distance threshold to $\tau$1 meters (Li et al., 24 Sep 2025). For each goal $\tau$2, the planner generates the remaining trajectory tokens by inpainting from

$\tau$3

The best candidate trajectory is then selected by a global scorer: $\tau$4 Although the notation alternates between $\tau$5 and $\tau$6, the intended role is a trajectory-level evaluator combining safety and quality terms (Li et al., 24 Sep 2025).

This stage provides multimodal exploration before reflection begins. Ablation results show that goal conditioning alone improves ego progress and overall PDMS relative to the base discrete diffusion planner without reflective inference: PDMS rises from $\tau$7 to $\tau$8, and EP rises from $\tau$9 to $N$0 (Li et al., 24 Sep 2025). That result supports the paper’s claim that multimodal goal hypotheses complement later safety-guided correction.

ReflectDrive-2 develops a related but more explicit decision–draft–reflect decomposition in token space, formalized as

$N$1

where sampled goals anchor drafted and post-edit trajectories (Wang et al., 6 May 2026). This later formulation suggests that ReflectDrive’s goal-first design anticipated a more general decomposition of driving behavior into intent selection followed by token-level revision.

4. Safety-aware reflection mechanism

The reflection loop is the defining mechanism of ReflectDrive. After a goal-conditioned candidate trajectory is selected, the planner checks whether it is safe. If it is not, ReflectDrive iteratively identifies the earliest unsafe waypoint, searches a local discrete neighborhood for a safer replacement, fixes that replacement as a safety anchor, and inpaints the surrounding trajectory while preserving the remaining context (Li et al., 24 Sep 2025).

The method uses three scorers: a global scorer $N$2 for whole-trajectory safety and coherence, a safety scorer $N$3 for identifying specific failures, and a local scorer $N$4 for evaluating candidate token-pair replacements at one waypoint (Li et al., 24 Sep 2025). The hard safety compliance term is

$N$5

where $N$6 is the no-at-fault collision metric and $N$7 is drivable-area compliance (Li et al., 24 Sep 2025). The quality term is

$N$8

with weights $N$9, $(x_j,y_j)$0, and $(x_j,y_j)$1 (Li et al., 24 Sep 2025). The paper states in prose that the final score is the product of hard safety and quality, implying

$(x_j,y_j)$2

though the complete standalone equation is not cleanly typeset in the paper (Li et al., 24 Sep 2025).

For reflection itself, the earliest unsafe waypoint $(x_j,y_j)$3 is located by the safety scorer, which assigns each waypoint a safety score based on the worst violation within a local time window (Li et al., 24 Sep 2025). The planner then searches a Manhattan neighborhood $(x_j,y_j)$4 of the offending token pair: $(x_j,y_j)$5 The paper describes $(x_j,y_j)$6 as a typical local search radius (Li et al., 24 Sep 2025). The optimized replacement becomes a safety anchor; the surrounding segment is then masked and regenerated by the diffusion model’s inpainting mechanism (Li et al., 24 Sep 2025). This loop repeats until the trajectory is safe or the iteration budget is exhausted. The appendix sets the maximum refinement iterations to $(x_j,y_j)$7, while the main text says most trajectories converge in $(x_j,y_j)$8–$(x_j,y_j)$9 iterations (Li et al., 24 Sep 2025).

This procedure is gradient-free. Rather than differentiating a reward through the denoising trajectory, ReflectDrive enumerates nearby token alternatives and uses the generative model only for coherence-preserving regeneration (Li et al., 24 Sep 2025). That is the central technical distinction from diffusion guidance based on reward gradients.

Two later works clarify the significance of this design. ReflectDrive-2 replaces local-search anchors with token-to-token rewriting inside the same discrete token space, using AutoEdit and full-rollout reinforcement learning; it reports that supervised AutoEdit improves PDMS by at most $x$0, whereas RL increases its gain to $x$1 (Wang et al., 6 May 2026). DriveFine instead adds a separate refinement expert that performs one post-generation corrective step and shows that refinement-oriented RL improves PDMS from $x$2 with offline RFT to $x$3 with online RFT on top of GRPO (Dang et al., 16 Feb 2026). These follow-ups suggest that ReflectDrive’s safety-aware reflection loop identified a reusable abstraction—proposal plus explicit corrective editing—even if later papers changed the editing mechanism.

5. Empirical performance and ablations

ReflectDrive is evaluated on NAVSIM using PDMS and its components. The base discrete diffusion planner without reflective inference, denoted “ReflectDrive (w/o R.I.),” achieves NC $x$4, DAC $x$5, TTC $x$6, Comfort $x$7, EP $x$8, and PDMS $x$9 (Li et al., 24 Sep 2025). The full model, “ReflectDrive (Ours),” improves these to NC $y$0, DAC $y$1, TTC $y$2, Comfort $y$3, EP $y$4, and PDMS $y$5 (Li et al., 24 Sep 2025). An analytical upper-bound variant using ground-truth agent states in reflection, “ReflectDrive$y$6,” reaches NC $y$7, DAC $y$8, TTC $y$9, Comfort $(\mathbf{y}_{j,x}, \mathbf{y}_{j,y})$0, EP $(\mathbf{y}_{j,x}, \mathbf{y}_{j,y})$1, and PDMS $(\mathbf{y}_{j,x}, \mathbf{y}_{j,y})$2, compared with a human reference of $(\mathbf{y}_{j,x}, \mathbf{y}_{j,y})$3 PDMS (Li et al., 24 Sep 2025).

The paper highlights the reflective inference gains over the same discrete diffusion planner without reflection: $(\mathbf{y}_{j,x}, \mathbf{y}_{j,y})$4 in DAC, $(\mathbf{y}_{j,x}, \mathbf{y}_{j,y})$5 in TTC, $(\mathbf{y}_{j,x}, \mathbf{y}_{j,y})$6 in NC, and $(\mathbf{y}_{j,x}, \mathbf{y}_{j,y})$7 in EP (Li et al., 24 Sep 2025). The especially large DAC improvement is consistent with the mechanism: the local neighborhood search and anchor-based regeneration directly target trajectory points that leave the drivable area. The more modest TTC and NC gains reflect a limitation explicitly noted by the paper: safety-guided regeneration uses constant-velocity estimates of surrounding agents, which are inaccurate in some critical scenes, and this is partly why ReflectDrive$(\mathbf{y}_{j,x}, \mathbf{y}_{j,y})$8, which uses ground-truth agent states, improves much further on TTC and NC (Li et al., 24 Sep 2025).

Ablation studies separate the effects of goal-conditioned generation and safety-guided regeneration. Without both, the model is the base planner at PDMS $(\mathbf{y}_{j,x}, \mathbf{y}_{j,y})$9. Adding goal conditioning alone yields PDMS $\mathcal{A}$0, mainly via EP improvement. Adding safety-guided regeneration alone yields PDMS $\mathcal{A}$1, with strong safety gains: NC $\mathcal{A}$2, DAC $\mathcal{A}$3, TTC $\mathcal{A}$4, EP $\mathcal{A}$5. The full combination reaches PDMS $\mathcal{A}$6, indicating complementarity between multimodal goal exploration and local safety correction (Li et al., 24 Sep 2025).

The paper also studies inference parameters. Performance improves initially with more diffusion generation steps and peaks at $\mathcal{A}$7 steps, then declines; more goal points improve performance; and more exploration steps and reflection iterations generally help, though the paper cautions that more inference opportunities do not always improve results (Li et al., 24 Sep 2025). These findings imply that ReflectDrive’s main leverage comes from modest-length diffusion generation followed by carefully budgeted corrective refinement, rather than arbitrarily long denoising or search.

6. Position in the literature, limitations, and legacy

ReflectDrive sits at a junction between imitation-based VLA planning, discrete diffusion, and explicit safety repair. Its principal contribution is not merely the use of discrete diffusion, but the use of discrete tokenization to make inference-time safety editing tractable through local search and inpainting (Li et al., 24 Sep 2025). In that respect it differs from rule-based post-refinement systems that depend on anchor libraries or hand-crafted path prototypes, from reinforcement-learning planners that optimize safety objectives primarily in simulation, and from gradient-guided diffusion methods that incur repeated gradient computation during inference (Li et al., 24 Sep 2025).

Its limitations are also explicit. The model uses only three current-frame camera views, so it lacks direct temporal motion cues for surrounding agents (Li et al., 24 Sep 2025). The reflection mechanism can oscillate between boundary avoidance and collision avoidance in tight spaces, its reward/scoring function does not adequately account for navigation correctness, and local search may fail if the initially selected goal points are poor (Li et al., 24 Sep 2025). The paper also notes an inference-time trade-off: reflection improves quality but adds local search, safety checks, and repeated inpainting, and the authors identify acceleration through mechanisms such as KV cache and reducing unnecessary inference iterations as future work (Li et al., 24 Sep 2025).

A broader limitation is that ReflectDrive’s reflection is external to the generative model’s weights. Safety scoring, local search, and inpainting are coordinated procedurally, but the planner is not directly optimized to make its drafts easier to refine. Later works address precisely this issue. ReflectDrive-2 makes drafting and editing co-adapt through full-rollout reinforcement learning and reports that supervised self-editing alone has limited effect until RL aligns the draft-and-edit pipeline (Wang et al., 6 May 2026). DriveFine likewise decouples a generation expert and a refinement expert, and uses hybrid offline/online RL so that the refinement expert learns from both relative ranking among generated trajectories and actual reward gains from refinement (Dang et al., 16 Feb 2026). These developments can be read as evidence that ReflectDrive identified the right intervention point—post-generation correction in token space—but left open how tightly correction should be integrated with training.

Even with those limitations, ReflectDrive remains a significant reference point in reflective autonomous-driving planning because it gives a concrete operational meaning to reflection: unsafe tokens are located, feasible local substitutes are searched, safe anchors are fixed, and a discrete diffusion model restores trajectory coherence around the edit (Li et al., 24 Sep 2025). That formulation connects later self-editing planners, refinement-augmented masked-diffusion VLAs, and discrete diffusion decision systems for driving. A plausible implication is that ReflectDrive’s lasting contribution lies less in any single benchmark number than in showing that, once actions are represented as editable discrete tokens, safe planning can be organized as an iterative cycle of proposal, diagnosis, local repair, and constrained regeneration.