CEED-VLA: Consistency Vision-Language-Action Model with Early-Exit Decoding

• VLA speed bottlenecks constrain high-frequency dexterous manipulation; Jacobi decoding offers parallel refinement, but standard VLAs achieve only a modest 1.28× speedup [13].Source: §1 · pp. 1–2
• Figure 1 reports 3.6× and 2.0× speedups for CEED-VLA on OpenVLA and LLaVA-VLA, respectively, with negligible performance degradation.Source: Figure 1 · p. 2
• Consistency distillation maps arbitrary points on a teacher’s Jacobi trajectory to its fixed point, enabling a student to predict several correct action tokens per iteration.Source: §1 · p. 2
• Dual supervision pairs an auxiliary autoregressive loss with mixed-label supervision that selects teacher outputs above a learned accuracy threshold and ground truth otherwise.Source: §1 · p. 2
• Early-exit decoding relaxes strict convergence conditions to bypass inefficient iterations, with minimal effect on success rates according to the authors’ analysis and observations.Source: §1 · pp. 2–3
• In simulation, CEED-VLA achieves 2-4.1× acceleration with comparable success rates across baselines; real-world deployment reaches 4× frequency with improved success on high-frequency dexterous tasks.Source: §1 · p. 2

• VLA acceleration approaches include adaptive inference depth, quantization-aware training, lightweight architectures, caching, tokenization, layer skipping, and parallel fixed-point decoding.Source: §2 · p. 3
• Consistency policies accelerate manipulation by mapping intermediate ODE states toward solutions; CEED-VLA instead maps intermediate Jacobi-trajectory points to a fixed point.Source: §2 · p. 3

• Greedy autoregressive decoding predicts each token from the prompt and preceding-token prefix, requiring n forward passes to produce n tokens.Source: §3 · p. 3
• Jacobi decoding reformulates token prediction as a system of nonlinear equations, then updates every token during each iteration.Source: §3 · pp. 3–4
• Jacobi decoding starts from a random token sequence and iterates until successive sequences match at convergence.The converged sequence is the fixed point.Source: §3 · p. 4
• Because Jacobi decoding can predict multiple correct tokens in parallel, it can require fewer iterations than autoregressive decoding.Source: §3 · p. 4

4.1 Teacher Model

• Teacher VLAs predict chunked 7-dimensional actions; LLaVA-VLA [13] [19] and OpenVLA [8] serve as teachers, while autoregressive training limits multi-token Jacobi accuracy.Source: §4.1 · p. 4

4.2 Student Model

• The student model is initialized from the target VLA and finetuned to predict multiple correct actions despite incorrect preceding tokens.Source: §4.2 · p. 4

4.2.1 Jacobi Trajectory Collection

• The target VLA generates Jacobi trajectories on the robot dataset, which are recorded as the consistency-distillation dataset; Figure 3 measures their L1 distance from ground truth.Source: §4.2.1 · p. 4, Figure 3 · p. 4

4.2.2 Consistency Training

• Consistency training maps a sampled Jacobi-trajectory state to its fixed point while an inherited autoregressive loss constrains drift from the teacher distribution.The objectives guide simultaneous prediction of multiple correct tokens and preserve manipulation skills.Source: §4.2.2 · p. 5
• Mixed-label supervision replaces autoregressive labels with ground-truth actions when a trajectory sample’s L1 distance exceeds the correctness threshold δmax.This addresses teacher action-prediction inaccuracies and mitigates error propagation during distillation.Source: §4.2.2 · p. 5

4.2.3 Inference

• Early exit stops Jacobi decoding at an exit point and emits the intermediate output; Table 1 compares AR, Jacobi, and early-exit inference speeds.Strict convergence can require over 30 iterations, while a small exit point offers higher speedup with an acceptable performance drop as CEED-VLA-Turbo.Source: §4.2.3 · p. 6, Table 1 · p. 6

5.1 Simulation Experiments

• On CALVIN, mixed-label consistency training with early-exit decoding yields 3.2× speedup and 4.4× frequency, with average length reduced 0.03; on LIBERO-Long, CEED-VLA achieves 3.6× speedup and 4.1× frequency.Integration with action chunking improves success rates by 15.8%.Source: §5.1 · pp. 7–8
• CEED-VLA can predict fixed action tokens correctly despite incorrect preceding tokens, enabling multiple token predictions per iteration.Figure 4 depicts the Jacobi trajectory from initialization to an exit point before the top-row fixed point.Source: Figure 4 · p. 6, §5.1 · p. 8

5.2 Ablation Experiments

• Consistency training and early exit accelerate OpenVLA, mixed-label supervision best balances speed and performance, and increasing AR-loss weight improves accuracy at a speedup cost.The ablations evaluate the designs on OpenVLA with action chunk=3.Source: §5.2 · pp. 8–9, Table 3 · p. 8
• An exit point of 16 gives noticeable speedup with negligible performance drop; 8 accelerates further with slight degradation, while 4 makes task success difficult.AR decoding takes 37 iterations to complete one action.Source: §5.2 · pp. 8–9, Figure 5 · p. 9
• 60k trajectories suffice for significant acceleration, gains become marginal beyond 120k, and one training epoch on 60k trajectories takes 10 hours.Increasing data volume has limited impact on average episode length.Source: §5.2 · p. 9, Figure 5 · p. 9

5.3 Real-world Eexperiments

• CEED-VLA raises dexterous-task success above 70% while boosting inference speed and control frequency for high-frequency actions in real-world experiments.The experiments use a dual-arm robotic system and include high-frequency towel-folding and water-pouring tasks.Source: Table 7 · p. 10, §5.3 · p. 10

6 Conclusions

• CEED-VLA combines mixed-label supervision during consistency training with early-exit decoding during inference.The paper describes it as a universal acceleration method intended to maintain manipulation performance.Source: §6 · p. 10
• Across several simulated and real-world environments, CEED-VLA improves inference speed and execution frequency while maintaining comparative manipulation performance.The conclusion also reports better management of high-frequency dexterous tasks.Source: §6 · p. 10
• The paper concludes that CEED-VLA better manages high-frequency dexterous tasks while accelerating inference and execution.Source: §6 · p. 10

7 Future work

• Combining CEED-VLA with tokenization schemes that represent multiple action components per token is proposed as a direction toward real-time inference.Source: §7 · p. 10
• The authors identify potential applications to embodied chain-of-thought methods that produce extensive intermediate reasoning tokens.Source: §7 · p. 10
• The authors also identify potential for knowledge insulating methods that generate linguistic responses and action tokens simultaneously.Source: §7 · p. 10

A Limitations

• CEED-VLA is less suited to continuous action spaces, where convergence conditions are harder to define.Source: A Limitations · p. 14
• The authors identify more effective and unified convergence conditions as a direction for future work.Source: A Limitations · p. 14

B Generating Jacobi Trajectory Datasets

• Algorithm 2 generates training examples by sampling image, instruction, and proprioception data, running teacher-model Jacobi decoding, and appending each trajectory to the training set.Source: B Generating Jacobi Trajectory Datasets · p. 14

C Benchmark

• CALVIN evaluation uses the ABCD→D setup across 34 tasks and four environments, with 500 rollouts and success-rate and sequential-task-completion metrics.The simulator uses a Franka Panda arm.Source: C Benchmark · p. 14
• LIBERO-Long evaluates multi-step, long-horizon manipulation over 100 rollouts per task with varying initial states, reporting per-task and average success rates.Its tasks require coherent, temporally aligned completion of subgoals.Source: C Benchmark · p. 14

D Baseline

• LLaVA-VLA is the main baseline, adding output-stage action chunking for temporally coherent actions; most CEED-VLA experiments use it on CALVIN ABC-D.The model is reported to perform stably in simulated and real-world environments and has appeared in prior work [13].Source: D Baseline · p. 15
• OpenVLA, an additional baseline for validating general applicability, combines an LLaMA 2 language model with DINOv2 and SigLIP visual features.The passage describes it as a widely used open-source VLA model.Source: D Baseline · p. 15

E Adaptability of Our Proposed Method

• The method requires no backbone redesign, pretrained-model modifications, or additional GPU memory overhead.The passage characterizes it as lightweight and practical for computationally constrained deployment.Source: E Adaptability of Our Proposed Method · p. 15

F Visualization of the Decoding Process

• AR decoding generates tokens sequentially, whereas Jacobi decoding updates all tokens in parallel but converges slowly without specialized training.The figure caption identifies Figure 8 as a comparison of four decoding methods.Source: F Visualization of the Decoding Process · pp. 15–16, Figure 8 · p. 15
• Table 8 evaluates early-exit speed and accuracy for LLaVA-VLA on CALVIN ABC→D using an NVIDIA RTX 4090 GPU.The supplied caption gives the evaluation scope but no numerical table values.Source: Table 8 · p. 16
• Consistency training improves Jacobi convergence from arbitrary initial predictions, reducing the steps needed while leaving some tokens inefficient to stabilize.Those outlier tokens constrain minimum achievable decoding latency.Source: F Visualization of the Decoding Process · p. 16
• Early exit terminates tokens whose predictions converge or change negligibly, or imposes a global iteration cap; consistency-trained models retain robust performance under strict limits.The passage says limiting iterations at certain decoding steps does not significantly affect overall task completion rate and characterizes early-exit Jacobi decoding as balancing speed and performance.Source: F Visualization of the Decoding Process · p. 16

G Ablation experiment

• Consistency models show more pronounced acceleration while maintaining high success rates at early exit points.The comparison covers different exit points.Source: G Ablation experiment · p. 16
• Consistency models converge most action tokens with minimal iterations, supporting their performance under early exit.Source: G Ablation experiment · p. 16
• Vanilla models fail to converge adequately under limited iterations and suffer significant performance degradation.Source: G Ablation experiment · p. 16