Fewer Steps, Better Actions: Rethinking Flow-Matching Inference for VLA Policies

Zhipeng Tang; Xinda Chen; Weining Rao; Xiao Li; Wenting Tan; Yuning Wang; Xiao Shi; Xiaofang Zhao

cs.ROSep 18, 2026

Fewer Steps, Better Actions: Rethinking Flow-Matching Inference for VLA Policies

Authors: Zhipeng Tang, Xinda Chen, Weining Rao, Xiao Li, Wenting Tan, Yuning Wang, Xiao Shi, Xiaofang Zhao

Organizations: Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China · University of Chinese Academy of Sciences, Beijing, China · Zhongke Haichuan Intelligent

Abstract

Vision-language-action (VLA) policies based on flow matching generate action chunks through repeated evaluations of an action expert. Increasing the number of integration steps raises inference cost, but does not necessarily improve closed-loop success. We propose Coda, which reallocates part of this integration budget to a single learned endpoint correction. A frozen policy first completes a few-step noise-to-action trajectory; a lightweight Transformer then predicts a demonstration-supervised residual using the candidate action, source noise, and shared observation-prefix cache. Only the corrector is trained. On 50 RoboTwin Easy tasks, five-step Coda improves success from 71.64% to 74.68% over the matched five-step baseline, while reducing forward latency by 30.2% relative to the default ten-step policy. A two-step configuration achieves 71.88% success with a 2.12 $\times$ speedup. An independent 13-task control shows a 5.69-percentage-point gain at nearly equal latency, supporting correction as an effective alternative to additional integration. The same design also improves frozen official SmolVLA, raising two-step success from 60.8% to 69.4%. These results show that endpoint correction improves the quality-latency trade-off of frozen flow-matching policies.

Figures & tables

Figure 1: Compute allocation. Ten-step and few-step generation both cover the full noise-to-action interval. Coda reallocates part of the serial integration budget to one learned endpoint correction; the target and offsets are schematic.

Figure 2: Coda inference. A frozen VLM encodes the three camera views and instruction once; in $\pi_{0.5}$ , discretized robot state is also included in the text prefix. Both modules reuse the prefix KV cache. The frozen AE completes $N$ Euler steps, and Coda uses the candidate and the same source noise to predict a residual in one forward. The residual is added to the candidate before execution. A line bridge denotes a crossing without a connection. Only Coda is trainable.

AE-NFE	1	2	3	4	5	10
$\pi_{0.5}$ -FT	59.76	67.40	71.28	69.96	71.64	71.52
Coda	69.80	71.88	73.64	73.12	74.68	72.60

Table I: Success on 50 Easy tasks (%). Each method has 50 episodes per task at every NFE. Bold indicates the higher mean at each NFE.

Figure 3: Diminishing returns of frozen

\pi_{0.5}

on the 50-task panel. (a) Success versus AE steps. (b) Measured forward latency versus AE steps. The highlighted 5-to-10-step segment changes success by

-0.12

percentage points while increasing latency by 61.0%.

Method	AE-NFE	Success (%)	Latency (ms)
$\pi_{0.5}$ -FT	5	70.40	193.39
$\pi_{0.5}$ -FT	6	71.69	218.14
$\pi_{0.5}$ -FT	10	73.02	318.98
Coda-MLP	5	73.63	195.63
Coda-Transformer	5	77.38	218.59

Table II: Independent 13-task compute-allocation panel. Each method has 621 valid episodes; timing uses an A100 40GB.

Figure 4: Where to spend the inference budget. The baseline line connects measured 5-, 6-, and 10-step policies. At approximately 219 ms, Coda improves success by 5.69 points over six-step generation. Upper-left is better; this panel uses the independent 13-task protocol in Table II .

Figure 5: Deployment landscapes on A100 80GB: (a)

\pi_{0.5}

, 50 tasks; (b) frozen official SmolVLA, 10 tasks. Upper-left is better. Point labels (2, 5, 10) indicate the number of AE integration steps (NFE); Coda adds one correction forward at each setting. Lines connect measured points at NFE

\geq 2

, and arrows highlight improvements over the default ten-step policy. One-step success is retained in Tables I and IV . The panels use separate host policies and evaluation protocols.

GPU	Method	N	P	A	R	Total	Speedup
A100	$\pi_{0.5}$ -FT	10	55.8	175.6	0.0	233.0	$1.00\times$
	$\pi_{0.5}$ -FT	5	55.8	88.0	0.0	144.7	$1.61\times$
	Coda	5	55.7	88.0	17.8	162.6	$1.43\times$
	$\pi_{0.5}$ -FT	3	57.1	54.3	0.0	112.7	$2.07\times$
	Coda	2	55.9	35.3	17.9	109.9	$2.12\times$
4090	$\pi_{0.5}$ -FT	10	62.7	182.1	0.0	246.5	$1.00\times$

Table III: Measured median forward latency (ms). P/A/R denote prefix, AE, and residual stage times. Speedup is relative to the same-GPU ten-step baseline. Bold highlights Coda deployment points.

AE-NFE	1	2	5	10
Official frozen	35.0	60.8	60.8	64.4
AE-cont	48.2	64.6	70.8	69.2
Coda ( $+\varepsilon$ )	55.8	69.4	72.0	72.6
Coda (no $\varepsilon$ )	58.2	71.4	75.2	74.8

Table IV: SmolVLA success on 10 Easy tasks (%). Each configuration completes 500 episodes. AE-cont and Coda (

+\varepsilon

) use checkpoint 18716; the no-

\varepsilon

variant uses checkpoint 18000.

Figure 6: Endpoint diagnostics. (a) Hanging-mug success with 95% Wilson intervals; the dashed divider separates the privileged GT-mix probe from deployable policies. (b) Reduction in mean open-loop action error relative to FT5; Mix denotes Mix-AE. Direct Coda execution preserves the largest reduction.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Input: observation

o

, AE-NFE

N

, recipe

\in\{\pi_{0.5}\text{-FT},\mathrm{Coda},\mathrm{Mix\text{-}AE}\}

Frozen:

f_{\mathrm{VLM}}

v_{\theta}

. On

\pi_{0.5}

, Coda uses residual

r_{\phi}

t_{\mathrm{res}}=0.1

; SmolVLA does not use that time conditioning.

1. Encode

o

once:

z_{o}=f_{\mathrm{VLM}}(o)

. Cache prefix key-value tensors.

2. Draw

\varepsilon\sim\mathcal{N}(0,I)

(use the host policy’s evaluation noise stream). Set

x^{[0]}=\varepsilon

3. For

i=0,\ldots,N-1

x^{[i+1]}=x^{[i]}-\frac{1}{N}v_{\theta}(x^{[i]},t_{i},z_{o})

t_{i}=1-i/N

4. Set

a_{\theta}=x^{[N]}

\pi_{0.5}

-FT: execute

a_{\theta}

and stop.

Appendix

Table V: Inference at AE-NFE

N

\pi_{0.5}

-FT stops at

a_{\theta}

; Coda adds one residual; Mix-AE is ablation only.

	$\boldsymbol{\pi}_{0.5}$ -FT	Coda	Cont	SmolRes
Init	$\pi_{0.5}$ FT	FT@60k	FT@60k	official
Trainable	VLM $+$ AE	residual only	AE	residual only
Global batch	$48$	$960$	$48$	$256$
Steps	$60$ k	$3$ k	$21$ k	$4$ ep
Frames	$2.88$ M	$2.88$ M	$1.01$ M	$4$ epochs
Peak lr	$2.5{\times}10^{-5}$	$5{\times}10^{-5}$	$2.5{\times}10^{-5}$	$1{\times}10^{-4}$

Appendix

Table VI: Training hyperparameters.

\pi_{0.5}

-FT / Coda / Cont use the

50

-task RoboTwin clean50 mixture. SmolVLA residual uses the official

10

-task set; official SmolVLA stays frozen.

	Easy	Hard	Smol Easy
Suite	RoboTwin 2.0	RoboTwin 2.0	RoboTwin 2.0
Split	demo_clean_50	demo_randomized_50	demo_clean_50
Tasks	$50$	$50$	official $10$
Shader	default (rt32)	minimal	default
Target $n$	$50$	mid- $20$ s– $49$	$50$
Seeds	$100000{+}i$	$100000{+}i$	$100000{+}i$

Appendix

Table VII: Environments for every closed-loop and latency number in the main text. Easy and Hard use

\pi_{0.5}

; Smol Easy uses official SmolVLA.

Explore similar work

Guided Action Flow: Value-Guided Sampling for Frozen Vision-Language-Action Policies
Jul 2, 2026Liuhaichen Yang, Zhuang Jiang, Chenchao Sheng +8Robotic Manipulation Vision-Language-Action Models
Department of Computer Science, University College London. · Department of Mechanical Engineering, University College London.
SEAM: Smooth Execution of Action-Chunked Motion for Vision-Language-Action Policies
Jul 6, 2026Dijia Zhan, Xuemiao Xu, Jinyi Li +1Vision-Language-Action Models Action Chunking
South China University of Technology
SnapFlow: One-Step Action Generation for Flow-Matching VLAs via Progressive Self-Distillation
Apr 7, 2026Wuyang Luan, Junhui Li, Weiguang Zhao +3Efficient VLA Model Inference Vision-Language-Action Models
Jilin University · Chongqing University · University of Liverpool +1

Jul 2, 2026J/K move · Enter open · S save

Guided Action Flow: Value-Guided Sampling for Frozen Vision-Language-Action Policies

Liuhaichen Yang, Zhuang Jiang, Chenchao Sheng, Ningwei Bai+7

Department of Computer Science, University College London. · Department of Mechanical Engineering, University College London.

Robotic Manipulation Vision-Language-Action Models Robustness of VLA Models RL for VLMs

Open paper AskPDF arXiv

Reinforcement learning can improve vision-language-action (VLA) policies beyond supervised fine-tuning, although this typically involves further updates to the policy parameters. For flow-matching policies, iterative action generation provides an additional opportunity to incorporate task information during inference. We introduce Guided Action Flow (GAF), which learns a compact, observation-conditioned action-value critic from robot task rollouts and applies its action gradient to steer reverse-time flow sampling. The supervised-fine-tuned VLA remains frozen throughout critic learning and deployment. Physical-robot experiments show an increase in aggregate success from 60.0% to 82.5% across six nominal manipulation tasks. Under six altered-lighting and object-distractor conditions evaluated on three of these tasks, aggregate success improves from 34.2% to 49.2%. Ablations and rollout analyses support the importance of the learned guidance direction and the critic's visual and proprioceptive inputs. With approximately 2.735M trainable critic parameters alongside a 0.45B-parameter VLA, GAF enables task outcomes to inform action generation through a compact inference-time guidance module.

Abstract

Explore similar work

Guided Action Flow: Value-Guided Sampling for Frozen Vision-Language-Action Policies

SEAM: Smooth Execution of Action-Chunked Motion for Vision-Language-Action Policies

SnapFlow: One-Step Action Generation for Flow-Matching VLAs via Progressive Self-Distillation