Organizations: Beijing University of Posts and Telecommunications · University of Electronic Science and Technology of China · National University of Defense Technology · Nanyang Technological University · Harbin Institute of Technology · Beihang University
When a vision-and-language navigation policy generates an action sequence autoregressively, the robot waits for a chain of dependent predictions. A small discrete action vocabulary makes it possible to score all short-horizon plans in parallel. We introduce VLN-Flash, a training-free decoder that evaluates the complete action-prefix tree in a single forward pass. Branch-specific attention and matched token positions preserve the original policy's conditional action distribution for each prefix in exact arithmetic. We use the score margin between the stepwise-greedy prefix and its strongest alternative at each depth to decide how many actions to execute. The robot then executes the plan with the highest joint score at the chosen depth before observing again. On matched inputs, VLN-Flash achieves a nearly 3× speedup in action decoding over native autoregression on Uni-NaVid and StreamVLN. Compared with the native autoregressive decoders, it improves success rate and success weighted by path length (SPL) on R2R-CE and RxR-CE, with success gains up to 2.7 percentage points. On a Unitree GO2, VLN-Flash improves navigation success and cuts the mean wait between actions by about 160 ms after subtracting network transmission time.
Figures & tables
Figure 1: Overview of parallel action decoding. The instruction and observation history are formatted as the model query x , followed by candidate action tokens. An attention mask restricts each candidate action token to attending only to the query, its branch ancestors, and itself. The VLN backbone processes the resulting sequence in a single forward pass. The hidden state of the last query token and those of all candidate tokens are then decoded into action distributions through the shared LM head. The last query token predicts the first action, so K=3 uses two candidate levels. Scoring K=4 adds a third level with 27 candidate tokens.
Figure 2: Comparison between autoregressive and parallel decoding. Autoregressive decoding (left) requires three sequential forward passes to predict a1 , a2 , and a3 , whereas parallel decoding (right) computes all candidate action distributions in a single forward pass. For each action prediction, the visible tokens (highlighted) are identical along the corresponding branch in both methods, yielding equivalent conditional action distributions in exact arithmetic.
Figure 3: Adaptive execution via joint path scoring. Nodes are labeled with conditional action probabilities, while L(path) denotes cumulative joint log-probabilities. The greedy path can be outscored by the selected path in joint probability. The scan compares the fixed greedy prefix with its strongest alternative at each depth. When their score margin falls below δ , the highest-scoring branch at that depth is executed and deeper actions are skipped. Here, m1=0.58>δ continues the scan, whereas m2=−0.18<δ at δ=0.2 selects a left turn followed by a forward step ( L→F ).
Uni-NaVid
StreamVLN
Decoder
Latency (ms) ↓
Speedup ↑
Latency (ms) ↓
Speedup ↑
Native autoregressive [ 27 , 23 ]
150.8
1.00×
182.8
1.00×
Speculative [ 8 ]
218.2
0.69×
242.6
0.75×
Draft & Verify [ 29 ]
136.6
1.10×
174.4
1.05×
Prompt Lookup [ 17 ]
141.4
1.07×
145.9
1.25×
Lookahead [ 4 ]
146.6
1.03×
126.5
1.45×
Table 1: Complete-reply latency across decoders. Speedup is relative to native autoregression on the same backbone.
Uni-NaVid
StreamVLN
R2R-CE
RxR-CE
R2R-CE
RxR-CE
Method
SR
SPL
NE
OS
SR
SPL
NE
SR
SPL
NE
OS
SR
SPL
NE
Reproduced
52.3
48.3
4.98
57.7
55.1
44.0
6.01
55.5
49.7
4.81
63.6
54.6
45.8
5.68
Published
51.8
47.7
4.96
59.4
56.1
44.5
5.67
56.4
50.2
4.90
63.6
54.4
45.4
5.65
Ours
53.2
49.2
4.98
57.8
57.8
46.7
5.71
57.5
51.0
4.73
65.7
56.2
46.6
5.57
Table 2: Navigation performance on the val-unseen splits. The Published row lists results from the original papers; Reproduced evaluates the same checkpoints as Ours with their native decoders. SR, SPL, and OS are percentages; NE is in meters. Bold and underline mark the best and second-best results within each backbone and dataset.
Uni-NaVid
StreamVLN
K
Autoregressive
VLN-Flash
Autoregressive
VLN-Flash
2
107.0
50.1
149.4
68.4
3
132.5
51.5
168.3
69.4
4
152.5
55.3
186.4
71.6
Table 3: Mean latency (ms) to generate a complete reply at each prediction horizon K .
Cached history (positions)
Native AR (ms)
VLN-Flash (ms)
Saved (ms)
533
85.8
32.0
53.8
752
86.8
32.5
54.3
1190
87.9
34.0
53.9
1627
89.6
35.6
54.0
Table 4: StreamVLN action-decoding latency with a preloaded cache. Timing starts with the cached history and current query ready, excluding visual preprocessing and cache construction. For each trajectory, the median of ten timed repetitions is used; values average the three trajectories.
R2R-CE
RxR-CE
Variant
SR
SPL
NE
OS
SR
SPL
NE
Uni-NaVid
Reproduced
52.3
48.3
4.98
57.7
55.1
44.0
6.01
Joint only
51.8
48.6
4.96
57.5
56.2
46.1
5.92
Ours
53.2
49.2
4.98
57.8
57.8
46.7
5.71
StreamVLN
Table 5: Joint selection and adaptive execution. Joint only selects the plan with the highest joint score at a fixed depth (two actions for Uni-NaVid and four for StreamVLN); Ours selects the depth adaptively. Metrics and rank markers follow Table 2 .
Uni-NaVid
StreamVLN
Execution
R2R-CE
RxR-CE
R2R-CE
RxR-CE
k=2
51.8/48.6
56.2/46.1
54.3/48.2
53.4/44.7
k=3
51.6/48.4
55.3/44.6
54.5/49.1
54.6/45.2
k=4
51.2/48.1
54.9/44.4
56.9/50.8
55.9/46.5
Adaptive
53.2/49.2
57.8/46.7
57.5/51.0
56.2/46.6
Avg. k
3.1
3.4
3.8
3.8
Table 6: Adaptive versus fixed execution (SR/SPL, %). All variants select plans by joint score. Fixed variants execute up to k actions and end early on stop . Avg. k refers to adaptive execution and excludes the final decision ending in stop . Bold and underline mark the best and second-best SR/SPL scores.
Figure 4: Success rates on the GO2 over 20 trials per environment, using the same StreamVLN policy, instructions, and start poses. Bar labels show success rates and successful-trial counts.
Figure 5: Real-world deployment trajectories on a Unitree GO2. The left and right scenes show a corridor and a classroom, respectively. In each scene, both methods start from the same position and receive the same instruction shown above. Baseline and Ours correspond to the autoregressive and parallel decoding variants of StreamVLN, respectively. Across the decision steps shown in these two scenes, our method reduces the average wait by 172 ms.
Figure 6: Additional Unitree GO2 deployments. (a) Office. “Turn left ahead, walk to the door, and stop in front of it.” (b) Outdoors. “Go straight along this road, climb the stairs, and stop before the glass railing.” (c) Classroom. “Go straight down this road to the end, and stop beside the table at the podium on your left.”
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Backbone
Precision
Cache
N
max∣Δℓ∣
Top-1
Uni-NaVid
fp16
Off
800
4.88×10−2
800/800
Uni-NaVid
fp16
On
800
4.30×10−2
800/800
StreamVLN
bf16
On
780
2.74×10−1
778/780
StreamVLN
fp32
On
780
9.16×10−3
780/780
Appendix
Table S1 : Agreement between tree and causal readouts at K=4 . N is the number of action distributions, max∣Δℓ∣ is the largest absolute difference across the four action logits, and Top-1 counts matching most likely actions.
Decoder
Target
Draft
GiB
StreamVLN (autoregressive)
–
–
15.44
Speculative (0.5B draft)
5.73
5.93
16.39
Draft & Verify
5.00
6.13
15.52
Prompt Lookup
5.60
0.00
15.49
Lookahead
4.41
0.00
15.51
Jacobi
2.93
0.00
15.60
Appendix
Table S2 : StreamVLN replay costs. Target and Draft are mean target and draft forward counts per reply. A dash marks an uninstrumented count. Memory is peak GPU allocation in GiB.
Our aim is to accelerate autoregressive vision-language-action models (VLAs), which generate robot actions one token at a time. Speculative decoding makes this possible: a lightweight draft model proposes several tokens at once, and the VLA verifies them in a single pass, accepting only those that match its own predictions, so decoding remains lossless. The speedup, however, is limited in robot control, where many different tokens encode nearly the same motion and a draft close to the correct action is still rejected. To address this, our approach, VSpark, is built upon three design choices: (i) rollouts collected with calibrated action noise, which expose the draft to a wider range of robot states; (ii) feature alignment, which trains the draft to match the VLA's hidden features rather than its output tokens alone; and (iii) continuation reuse, which keeps the predictions the VLA has already computed along rejected drafts as candidates for the next verification step, instead of discarding them. On four VLAs with different action tokenizers and four manipulation benchmarks, VSpark achieves 1.28x to 2.10x speedup while keeping the VLA's decoding lossless.
Existing vision-language navigation methods often couple a VLM with waypoint decoders to produce multi-step action plans, but they typically lack an explicit closed-loop mechanism for tracking semantic progress, diagnosing execution failures, and recovering from error accumulation in long-horizon navigation. To address this gap, we propose ReflectVLN, an agentic VLN framework that organizes decision-making through bidirectionally interactive intention and execution agents. The intention agent performs subtask decomposition and reflection, generating executable subtask descriptions as corrective plans. Conditioned on these descriptions, the execution agent grounds them into short-horizon actions under current observations while monitoring sub-goal progress and detecting off-track behavior. Crucially, ReflectVLN enables closed-loop bidirectional communication: the execution agent emits progress and deviation signals to trigger reflection and subtask updates on demand, and the intention agent returns structured guidance that reconditions subsequent actions for recovery. To encourage temporally coherent decisions with interpretable intermediate rationales, we introduce Action Chain-of-Thought (Action-CoT), a path-conditioned dual-query training scheme for action generation. Experiments on standard VLN benchmarks show that ReflectVLN improves success rates and path efficiency under a constrained data budget, with favorable training cost and fewer high-level intention calls at inference time, while providing interpretable intermediate decisions for analysis and collaboration. Code is available at: https://github.com/AIprogrammer/ReflectVLN
Vision-Language Navigation (VLN) requires long-horizon instruction-consistent planning and short-horizon path planning grounded in the currently observed geometry. Anticipating how future observations evolve along the instructed route provides predictive guidance beyond the current field of view, motivating video-action models (VAMs) that learn representations of future visual states. However, many existing VAMs couple future visual prediction with action generation during inference, despite their distinct temporal requirements: instruction-conditioned route guidance can be reused across multiple control steps, whereas local trajectories must continuously adapt to fresh observations. To bridge this gap, we introduce HAVANA, a dual-system VAM that decouples predictive reasoning from reactive local planning. A slow system produces a compact language-conditioned latent plan from observation history, while a fast system combines the latest plan with the current observation to generate local trajectories. Their asynchronous execution enables rapid adaptation without repeated future prediction, and no future video is generated during inference. Experiments on OmniNavBench demonstrate state-of-the-art navigation performance over prior baselines, while zero-shot real-world deployments further show robust generalization to unseen environments.
Dongzhihan Wang, Yi Du, Chen Wang +2
Tongji University · SAIR Lab · Shanghai University