Venue: ICLR 2026 Category: VLA Architecture — Efficiency Trend tag: Efficiency / real-time deployment Affiliation: Tsinghua University · CUHK · UIUC · Beihang University

flowchart LR
  Hist[Action buffer S_A<br/>last n=6 actions] --> SCH[Action-aware Scheduler<br/>velocity check + intuitive ratio τ=0.5]
  SCH -- "Deliberative path<br/>low speed or low intuitive ratio" --> TP[Spatio-Semantic Token Pruning<br/>velocity-adaptive prune ratio]
  TP --> ENC[ViT Encoder]
  ENC --> LLM[7B LLM<br/>e.g. Llama 2 / OpenVLA / CogACT]
  LLM --> AH[Action Head<br/>D-tokenizer or diffusion]
  SCH -- "Intuitive path<br/>v in v_min..v_max and N_G/N_A above τ" --> RIDGE["Lightweight Generator<br/>Ridge Regression on action buffer<br/>under 1K params"]
  RIDGE --> CHK[Validity check]
  AH --> ACT[Robot action]
  CHK --> ACT
  ACT -. feedback .-> Hist

VLA models are too compute-heavy for real-time deployment. OpenVLA is 7B+, RT-X is up to 55B; even on a 4090 inference runs at ~4 Hz. Existing acceleration work attacks single-step compute via quantization (QAIL), early-exit (DeeR-VLA), token caching (VLA-Cache), parallel decoding (PD-VLA, OpenVLA-OFT), or speculative decoding — but all treat one frame in isolation.

SP-VLA argues two redundancies are systematically ignored:

• Temporal redundancy in the sequential action stream (most steps don't need full reasoning — humans alternate "deliberate" vs "intuitive" motor control).
• Spatial redundancy in visual tokens (most patches are background — but VLA-targeted token pruning has unique constraints).

Behavioral analysis. Across 50 pick-and-place trials, the authors observe the manipulator follows a 4-phase velocity profile: target → grasp → move → place. Slow alignment, then high-speed translation, then careful action, etc. They argue VLA models have implicitly learned this kinematic pattern, so the action stream is naturally split into:

• Deliberative actions — slow, precise (grasping, turning, placement). Need the full 7B VLA.
• Intuitive actions — fast, ballistic (point-to-point translation). Can be approximated by a tiny model.

Scheduler logic. Let a^t_d = (a_x, a_y, a_z) be the per-step end-effector translational velocity. Define:

• An action a is intuitive if all components |a_i| > v_min (speed threshold).
• The lightweight model is allowed when (a) a_{t-1} ∈ [v_min, v_max] and (b) the ratio of recent VLA-generated actions in the buffer N_G / N_A > τ (default τ = 0.5).

LWM = 1 if both hold, else 0. (Equation 1.) This drives small-step, high-frequency model switching — even within an "intuitive" segment, the VLA is invoked periodically to correct drift.

Lightweight generator: Ridge Regression on the action buffer S_A = {a_{t-n}, …, a_{t-1}} (n=6).

• X = [T, 1] ∈ R^{n×2}, T = [0,…,n−1]ᵀ; Y = action buffer; β ∈ R^{2×ℓ}.
• Minimize J(β) = ‖Xβ − Y‖² + λ‖β‖² (Tikhonov / ridge).
• Closed-form: β = (Xᵀ X + λI)⁻¹ Xᵀ Y.
• Predict a_t = x_t β*, x_t = [t 1]ᵀ.
• Gripper state is NOT regressed (it's binary) — instead reuse the t−1 value, leaving binary state transitions to the VLA. Predicted intuitive actions go through a validity check before execution.

Insight from controlled experiments (Fig. 2b): Random token pruning degrades but doesn't destroy task performance — there is spatial redundancy. But two surprising failure modes:

1. Reordering tokens by semantic importance (no actual pruning) causes complete task failure → relative position of tokens carries spatial meaning the auto-regressive VLA depends on.
2. Pruning purely by semantic attention scores removes object-contour background tokens and also fails → object contours are critical for spatial grounding.

So pruning must (a) preserve relative ordering and (b) explicitly retain edge / contour tokens.

Semantic-aware token importance. From last-encoder-layer attention: Q,K,V = X W_q,k,v Attn = Softmax(QKᵀ/√d_k) V AccuAttn = ½ (eᵀ ⊗ I_M) vec(Attn). Select T_se = {x_i | AccuAttn_i > t_ks}.

Spatial-aware token importance. Apply Canny edge detector to the input image: X_s = Canny(X). Then T_sp = f_E(X_s) is the ordered set of edge-region tokens.

Order-preserving union: T_select = U(T_se, T_sp) — the union, with original token positions retained.

Velocity-adaptive prune rate. Pruning is disabled for low-speed (deliberative) actions to protect precision. For higher speeds: T_r(v) = 1 if v < v_pmin, else 1 − (v − v_pmin) / (v_pmax − v_pmin).

This couples the spatial dimension (how aggressively to prune) to the temporal mode — fast intuitive actions can tolerate aggressive pruning; precise grasps cannot.

• Buffer size n = 6.
• Deliberation/intuition ratio threshold τ = 0.5.
• Velocity thresholds: v_min = 0.2, v_max = 0.5 (paper settings); token-pruning velocity threshold v_pmin = 0.5. Sensitivity study (Table 7) varies these ±25% and finds accuracy robust to speed but sensitive to n and τ.
• Hardware: NVIDIA A100 GPUs for experiments/training; NVIDIA RTX 4090 (40GB) for frequency/latency measurements (paper Sec. on Frequency and Latency).

"+R" = preserve relative token positions. "+S" = add Canny edges. SP-VLA is the only method to deliver acceleration with ≥ baseline accuracy. PruMerge is interesting — it gets a 1.36× speedup but drops to 0% on every suite (the model collapses).

Visual Matching split:

Visual Aggregation split:

Note +27 pp on Drawer in Visual-Aggregation while still 1.81× faster — interpreted as error-correction (the scheduler/lightweight generator smooths trajectories where CogACT alone falters).

WidowX split (bottom of Table 2; paper labels it "WindowX"):

Stack Block jumps from 15 → 29 with 2.54× acceleration.

≈ 2.2× frequency improvement, ~50% latency reduction.

The ICLR camera-ready / arXiv v3 paper (2506.12723v3, Oct 2025) reports no real-robot experiment table — all benchmarks are simulation (LIBERO + SimplerEnv). The project's GitHub README separately states a real Franka Panda result of ~2.5× end-to-end inference acceleration with only a 1% success-rate drop, but provides no per-task breakdown, FLOPs, or latency numbers. Treat any detailed real-robot figures with caution until the source table is located.

Audit note: a previously listed "Real Franka Research 3 (Table 4)" table with per-task success rates (80/74/77 vs 78/74/76), FLOPs 35.55, 0.27/0.13 s latency, "150 trajectories per task," and "20 morning / 10 noon / 20 evening" lighting splits was not found in any reachable source and has been removed as a likely fabrication.

Without Canny edge tokens the model effectively collapses (74.9 → 23.9), confirming the central insight that VLA spatial perception relies on object-contour tokens.

• Token pruning identifies ~22.7% redundancy on LIBERO-Spatial.
• Model scheduling identifies ~21% redundancy on LIBERO-Object.
• Intuitive-action proportion grows with task length: 18% on LIBERO-Spatial → 28% on LIBERO-Long → corresponding speedups 1.18× → 1.39×.

The temporal and spatial dimensions are not orthogonal — different tasks expose redundancy along different axes — so combining both is multiplicative, not additive.

The authors flag a single explicit limitation:

1. Intuitive-action generation is preliminary. Currently they only "lightweight-ify" the VLA via Ridge Regression; they have not achieved a complete behavioral separation between deliberative and intuitive generation. Making the VLA more human-like by structurally separating these two modes (rather than scheduling between a big and tiny model) is flagged as the key future direction.

Other implicit caveats:

• Hyperparameter v_min, v_max are device-dependent. The 1/4-3/4 max-task-speed heuristic (v_min = 0.2, v_max = 0.5 in the paper's settings) is validated only in simulation; the sensitivity study (Table 7) shows accuracy is robust to ±25% speed perturbation but sensitive to buffer size n and the intuitive-action proportion τ.
• Lossless acceleration claim is benchmark-specific. "1.5× lossless on LIBERO" is the Speed-priority preset with 71.9 vs 74.6 baseline (2.7 pp drop) — not strictly lossless. Authors' "Acc" preset gets 74.9 (no drop) at 1.35×.
• No real-robot evaluation in the paper. v3 reports only LIBERO + SimplerEnv (simulation); the lone real Franka claim lives in the repo README without a results table.
• Lightweight generator works because intuitive segments are approximately linear; on highly nonlinear / contact-rich intuitive segments (rare in pick-and-place but common in dexterous manipulation), Ridge Regression may break down.

• vs token-pruning-only methods (Action-aware Dynamic Pruning, FastV, VLA-Cache, EfficientVLA, SparseVLM, FoPru, FastVLM, VisionZip, PruMerge): SP-VLA's spatio-semantic-with-edge pruning prevents the catastrophic failure the paper documents in pure semantic pruners (PruMerge → 0%, VisionZip → 33%). Combined with scheduling it gets 2.4× SimplerEnv speedup at +0.25 pp accuracy, which no single-axis pruner achieves.
• vs quantization (AutoQVLA, QAIL): orthogonal — could be stacked.
• vs parallel decoding (OpenVLA-OFT, PD-VLA, FASTER): these accelerate inside one VLA forward pass; SP-VLA accelerates between forward passes by skipping them. Composable.
• vs hierarchical dual-system VLAs (π0.5, Hi-Robot, Helix, OneTwoVLA): dual-system designs route reasoning to a separate model; SP-VLA borrows the System-1 / System-2 metaphor but applies it to action-typing rather than reasoning. The "lightweight generator" is essentially a System-1 motor primitive (ridge regression) for intuitive arm trajectories, while the full VLA acts as System-2 for grasp-class moments.
• vs early-exit (DeeR-VLA): DeeR exits earlier in the LLM stack on easy frames; SP-VLA skips the LLM entirely on intuitive frames, which is a stronger speedup ceiling.
• First systematic treatment of temporal × spatial redundancy in VLA inference — the action-type indicator + edge-aware pruning ablation reveals previously unappreciated structural facts about VLA computation.

• OpenReview: https://openreview.net/forum?id=RwdGIIjPlC
• Code: https://github.com/ChildTang/SP-VLA

• Action-aware Dynamic Pruning — token-level efficiency
• AutoQVLA — channel-aware quantization
• FASTER
• OneTwoVLA — adaptive reasoning, kindred conceptual structure
• Survey: VLA & Manipulation

← Back to ICLR-2026