Venue: ICLR 2026 Category: VLA Architecture — Discrete diffusion / Autonomous driving Trend tag: Diffusion · Reflective inference · Safety constraints Affiliations: LiAuto, Tsinghua University (Li, Zheng, Wang, Wang, Zhao, Liu, Zhan, Zhan, Lang)

flowchart LR
  Obs[Multi-view cameras<br/>front, front-left, front-right<br/>+ ego state + nav instruction] --> Tok[Trajectory tokenizer<br/>uniform 1D codebook per axis]
  Tok --> DLM[Discrete Diffusion LM<br/>LLaDA-V backbone, fine-tuned]
  DLM --> Goal[Goal-Conditioned Generation<br/>Top-K' goals, NMS dNMS = 0.9 m, K = 3]
  Goal --> Cand[K candidate trajectories]
  Cand --> Sg[Global Scorer Sglobal]
  Sg --> Best[Best initial trajectory τ*]
  Best --> Loop{Safety-Guided<br/>Regeneration Loop<br/>≤ 10 iterations}
  Loop -->|Safety Scorer Ssafe| Viol[Identify violating waypoints V]
  Viol -->|Local Scorer Slocal| Anchor[Manhattan local search δ ≤ 10]
  Anchor --> Inpaint[Diffusion inpainting<br/>around safe anchors]
  Inpaint --> Loop
  Loop -->|safe| Out[Final trajectory]

End-to-end driving stacks struggle to encode physical / safety rules during training. Existing fixes are either (1) rule-based post-processing that demands human engineering, (2) RL with simulation-only training and unsafe online rollouts, or (3) gradient-based diffusion guidance that is computationally expensive and brittle around safety constraints. The paper targets a gradient-free way to inject safety into a learned VLA-style planner.

Each 2D waypoint (x, y) is quantized independently with a uniform 1D codebook A = {a₁, a₂, …} over spatial range [-M, M] = [-100, 100] m (Table 4) at resolution Δg; the paper defines Δg symbolically but does not report a numeric value for it or the resulting codebook cardinality. A trajectory of N waypoints is flattened into a length-2N token sequence

y = Q(τ) = (y1,x, y1,y, ..., yN,x, yN,y) ∈ A^{2N}

This makes BEV-space search efficient and supports the inpainting story of discrete diffusion.

Forward process masks tokens with a cosine-style noise schedule. Reverse process is trained with the standard masked-token NLL

L(θ) = E[ -Σ_{i: m_i^(s) = 1} log p_θ(y_i | ỹ^(s), c, s) ]

conditioned on context c = three camera views + ego state + language navigation command. Backbone is LLaDA-V (You et al., 2025), a pretrained Diffusion Language Model. Training uses classifier-free guidance.

Training hyperparameters (Table 4):

Inference config (Table 3): 5 diffusion steps, answer length 32, block length 32, low-confidence remasking, K = 3 goal candidates, NMS distance threshold 0.9 m, max refinement iterations 10.

Three scoring functions (implementation in Appendix C):

• Sglobal(τ): overall trajectory quality including safety + coherence; returns 0 if any critical rule is violated.
• Ssafe(τ): safety oracle that scores each waypoint by worst-violation-in-local-window.
• Slocal(ax, ay): per-token-pair scorer used during local search.

Stage 1 — Goal-Conditioned Generation. The model produces a distribution over the terminal waypoint pθ(yN | c, s). Top-K' goal candidates are extracted, then Non-Maximum Suppression with threshold dNMS = 0.9 m returns K = 3 spatially diverse goals G = {G1, G2, G3}. For each goal Gk, a full trajectory is generated by inpainting the intermediate waypoints conditioned on Gk. The K trajectories are scored with Sglobal; the highest-scoring τ* proceeds to Stage 2.

Stage 2 — Safety-Guided Regeneration. Iterative loop:

1. Ssafe identifies violating waypoint indices V = {t | Ssafe(τ*)t < τ_safe}.
2. For each t ∈ T* ⊆ V, search the Manhattan neighborhood Nδ (δ ≤ 10 tokens) for the corrected pair maximizing Slocal — these become safety anchors.
3. Diffusion inpainting regenerates the surrounding tokens conditioned on the safety anchors.
4. Repeat until V is empty or budget = 10 iterations is hit; fall back to highest-Ssafe candidate otherwise.

Empirically most violations are resolved in 1–3 iterations (the paper does not report a numeric average).

PDMS aggregates NC (no-collision), DAC (drivable-area compliance), TTC (time-to-collision), Comfort, EP (ego progress).

Camera-only ReflectDrive surpasses Camera+LiDAR Augmented planners on PDMS, with the largest margin on DAC = 99.3 (drivable-area compliance), where reflection injects hard safety. With ground-truth-agent oracle (†), the system matches human performance within noise.

Reflection improves over the no-RI baseline by +3.9 DAC, +1.3 TTC, +0.8 NC, +7.9 EP → +6.3 PDMS.

Goal-Conditioned generation primarily lifts EP (progress); Safety-Guided primarily lifts DAC and TTC. They are complementary.

• Generation steps (Fig. 4a). Non-monotonic: PDMS peaks at 5 steps, then declines.
• Goal candidates K and NMS dNMS (Fig. 4b). Multi-modal goal generation gives a wider option set; scaling K beyond 3 yields diminishing returns.
• Exploration steps and max iterations (Fig. 4c). PDMS scales positively with both → "inference-time scaling" exists for discrete-diffusion planners.

1. Single-frame inputs. Three-view images of the current frame only; velocity / motion of surrounding agents is unobserved. Future: incorporate history and predict obstacle trajectories jointly.
2. Reflection design.
   • Goal-Conditioned Generation reuses the PDM scorer without task-specific tuning of high-level navigation/efficiency objectives.
   • Safety-Guided Regeneration: more iterations don't always help. Failure modes:
     • Boundary oscillation in narrow drivable space (discrete-token rounding error compounds).
     • Navigation correctness not encoded in the reward.
     • Goal-point selection is suboptimal in some scenarios; constrained search range can't recover.
3. Sample efficiency. Limited engineering optimization so far; the authors note substantial room to improve inference efficiency (no concrete latency figures are reported in the paper).

Provenance note (important): This is the only "discrete-diffusion VLA" actually at ICLR 2026, and it lives in the autonomous driving family — distinct from the manipulation-focused DDVLA paper at NeurIPS 2025 which the community sometimes conflates with it. The contribution is not "discrete diffusion for VLA" per se; it is the gradient-free reflection / inpainting loop that uses the discrete-token structure to perform local search and regeneration around safety anchors.

Compared to neighbors:

• AutoVLA (Post-RFT): same VLA-for-driving paradigm, autoregressive instead of discrete-diffusion, slightly weaker on PDMS (89.1 vs 91.1) and notably weaker on EP (81.9 vs 86.9). Reflection lets ReflectDrive correct without RL fine-tuning.
• Hydra-MDP / DiffusionDrive / GoalFlow: rely on trajectory anchors or rule-based guidance and Camera+LiDAR. ReflectDrive matches/beats them with camera only.
• Diffusion Planner / classifier-guided diffusion: gradient-based guidance during denoising — expensive and parameter-sensitive. ReflectDrive's discrete inpainting is gradient-free.
• Manipulation discrete-diffusion VLAs (DDVLA NeurIPS 2025): shares the discrete-action-codebook idea but optimizes for sampling efficiency, not safety constraint enforcement.

The broader bet is that discrete action spaces + masked diffusion make it natural to compose learned generators with rule-based search — the same recipe that worked for code (FIM, infilling) now applied to safety-critical planning.

• OpenReview: https://openreview.net/forum?id=XJxXSMLDoZ
• PDF: https://openreview.net/pdf?id=XJxXSMLDoZ

• ICLR-2026

← Back to ICLR-2026