Discrete Forcing: Infusing Discrete Guidance into Continuous Denoising for Few-Step Action Experts

Jingbo Wang; Wenxuan Song; Wenhao Yu; Han Zhao; Xi Wang; Jiayi Chen; Donglin Wang; Yan Wang

cs.ROSep 30, 2026

Discrete Forcing: Infusing Discrete Guidance into Continuous Denoising for Few-Step Action Experts

Authors: Jingbo Wang, Wenxuan Song, Wenhao Yu, Han Zhao, Xi Wang, Jiayi Chen, Donglin Wang, Yan Wang, +1 more

Organizations: The Hong Kong University of Science and Technology (Guangzhou) · South China University of Technology · University of Science and Technology of China · Westlake University · Zhejiang University · Tsinghua University

Abstract

Efficient action generation in vision-language-action (VLA) models requires capturing both coarse action structure and fine-grained details. Discrete action tokens provide compact structural representations but sacrifice precision, while continuous action tokens offer high precision but often require multiple denoising steps. We introduce Discrete Forcing, a flow-matching framework that combines these representations through an explicit coarse-to-fine generation process. It first predicts discrete action tokens to establish a coarse action structure, then uses them to guide continuous action refinement. The discrete and continuous components share a common diffusion transformer backbone with specialized branches, maintaining a parameter count comparable to a conventional single-branch model while requiring only one forward pass per branch. Extensive evaluations across multiple benchmarks demonstrate improved performance and faster inference over a parameter-matched continuous action expert, with consistent performance gains as model capacity increases. Real-world experiments further demonstrate improvements on high-precision and dynamic manipulation tasks.

Figures & tables

Figure 1 : Overview of the Discrete Forcing (DF). (a) Discrete Forcing Expert (DFE) performs discrete action prediction followed by continuous refinement within a partially shared Action DiT conditioned on VLM features. (b) This two-stage design reduces inference cost while consistently improving performance across simulation and real-world benchmarks.

Figure 2 : Comparison of action generation paradigms. Existing methods rely on $K$ -step iterative denoising in either (a) discrete or (b) continuous action spaces, whereas (c) Discrete Forcing (Ours) first predicts a coarse discrete action and then performs a single continuous refinement.

Method	Parameters	NFE $\downarrow$	Spatial	Object	Goal	Long	Avg.
Iterative Generation (NFE $>2$ )
DreamVLA ( Zhang et al., 2026 )	1.6B	10	97.5	94.0	89.5	89.5	92.6
MemoryVLA ( Shi et al., 2026 )	7B	10	98.4	98.4	96.4	93.4	96.5
$\pi_{0.5}$ ( Physical Intelligence et al., 2025 )	3B	10	98.8	98.2	98.0	92.4	96.9
FlowerVLA ( Reuss et al., 2025 )	1B	8	97.5	99.1	96.1	94.9	96.9
Fast Generation (NFE $\leq 2$ )

Table 1 : Comparison on the LIBERO benchmark. Prior methods are grouped according to their number of function evaluations (NFE): iterative generation with more than two NFE, and fast generation with NFE

\leq 2

. Best results are highlighted in bold.

Table 2 : Mean task success rate (%) on RoboTwin 2.0 and RoboCasa-GR1.

Method	Success Rate (%) $\uparrow$	NFE $\downarrow$	Action Latency (ms) $\downarrow$		End-to-End Latency (ms) $\downarrow$
Method	Success Rate (%) $\uparrow$	NFE $\downarrow$	P50	P95	P50	P95
StarVLA (NFE=10)	96.7	10	145.46	149.36	200.43	206.06
StarVLA (NFE=4)	96.1	4	58.05	59.88	119.86	123.57
StarVLA (NFE=2)	95.1	2	29.20	30.44	93.29	96.26
Discrete Forcing (Ours)	97.6	2	24.91	29.69	78.95	83.89

Table 3 : Inference efficiency comparison on LIBERO. Action latency measures the action-generation module only, while End-to-End Latency additionally includes the VLM forward pass. P50 and P95 denote the median and 95th-percentile latency, respectively.

Figure 3 : Model scaling experiments. Marker size denotes model parameter count, while the value in parentheses indicates the parameter count of the action expert. Our Discrete Forcing consistently improves with model scale while maintaining lower inference latency than the baseline.

Figure 4 : Comparison of the action generation mechanisms between our method and the continuous-generation baseline. DF denotes Discrete Forcing. DF@1 denotes the output after the first discrete forward pass, whereas DF@2 denotes the output after the subsequent continuous refinement. StarVLA@

n

denotes the action output after the

n

-th iterative forward pass.

Figure 5 : Real-world experiments across four tasks.

Variant	Success Rate (%)	$\bm{\Delta}$
Discrete Forcing (Ours)	95.2	–
Generation Order
$Discrete\!\rightarrow\!Discrete$	93.6	$-1.6$
$Continuous\!\rightarrow\!Continuous$	91.8	$-3.4$
Architecture and Training Design
w/o Discrete-guided Source	92.6	$-2.6$

Table 4 : Ablation studies on LIBERO-Long. We analyze the effect of generation order, architectural designs, and discrete conditioning strategies.

Configuration	Setting
Model parameters	1.55B total
Action DiT	24 layers per execution path, with 2 shared and 22 branch-specific layers; hidden dim. 1024
Action horizon	8
Optimizer	AdamW, $\beta=(0.9,0.95)$ , $\epsilon=10^{-8}$ , weight decay $10^{-8}$
Learning rate	$1\times 10^{-5}$ for both VLM and Action Expert
Batch size	16 per GPU on 4 H100 GPUs; global batch size 64

Table 5 : Implementation details on LIBERO.

Task	Success Rate	Task	Success Rate
adjust_bottle	100	place_can_basket	56
beat_block_hammer	78	place_cans_plasticbox	32
blocks_ranking_rgb	26	place_container_plate	98
blocks_ranking_size	20	place_dual_shoes	60
click_alarmclock	64	place_empty_cup	96
click_bell	60	place_fan	34

Table 6: Task-level success rates (%) on RoboTwin 2.0. We report the success rate for each task, together with the average performance across all tasks.

Task	DiT4DiT	GR00T-N1.6	StarVLA- $\pi$	LangForce	DualCoT-VLA	LDA	Ours
BottleToCabinetClose	48.0	51.5	26.0	72.0	66.0	76.0	74.0
CanToDrawerClose	74.0	13.0	62.0	78.0	64.0	71.0	82.0
CupToDrawerClose	52.0	8.5	42.0	46.0	46.0	41.0	20.0
MilkToMicrowaveClose	50.0	14.0	50.0	56.0	58.0	52.0	62.0
PotatoToMicrowaveClose	36.0	41.5	42.0	36.0	30.0	41.0	40.0
WineToCabinetClose	42.0	16.5	32.0	46.0	38.0	57.0	56.0

Table 7 : Per-task success rate (%) on the RoboCasa-GR1 benchmark.

# Shared Layers	Success Rate (%)
0	88.2
2	95.2
4	95.0
6	93.6
8	93.4
10	90.0

Table 8 : Ablation on the number of shared Action DiT layers on LIBERO-Long.

Discrete Action ( $\alpha$ )	Gaussian Noise ( $1-\alpha$ )	Success Rate (%)
0.3	0.7	95.2
0.5	0.5	93.2
0.7	0.3	91.0

Table 9 : Ablation on the composition of the discrete-guided continuous source on LIBERO-Long.

\alpha

controls the contribution of the dequantized discrete action, while

1-\alpha

controls the Gaussian noise.

Figure 6 : Real-world experimental setup.

Explore similar work

CF-VLA: Efficient Coarse-to-Fine Action Generation for Vision-Language-Action Policies
Apr 27, 2026Fan Du, Feng Yan, Jianxiong Wu +8Action Generation
1Southern University of Science and Technology · 2Xi’an Jiaotong University · 3United Nova Technology +1
TBD-VLA: Temporal Block Diffusion Vision Language Action Model
Jun 5, 2026Sung-Wook Lee, Xuhui Kang, Yen-Ling KuoDiffusion-Based Vision-Language-Actions Action Generation
University of Virginia
dVLA-RL: Reinforcement Learning over Denoising Trajectories for Discrete Diffusion Vision-Language-Action Models
Jun 22, 2026Yuhao Wu, Yitian Liu, Weijie Shen +13Generative Models
ScaleLab at SJTU · 1Shanghai Jiao Tong University · 3Baidu AI Cloud +3

Apr 27, 2026, cs.CVJ/K move · Enter open · S save

CF-VLA: Efficient Coarse-to-Fine Action Generation for Vision-Language-Action Policies

Fan Du, Feng Yan, Jianxiong Wu, Xinrun Xu+7

1Southern University of Science and Technology · 2Xi’an Jiaotong University · 3United Nova Technology+1

Action Generation

Open paper AskPDF arXiv

Flow-based vision-language-action (VLA) policies offer strong expressivity for action generation, but suffer from a fundamental inefficiency: multi-step inference is required to recover action structure from uninformative Gaussian noise, leading to a poor efficiency-quality trade-off under real-time constraints. We address this issue by rethinking the role of the starting point in generative action modeling. Instead of shortening the sampling trajectory, we propose CF-VLA, a coarse-to-fine two-stage formulation that restructures action generation into a coarse initialization step that constructs an action-aware starting point, followed by a single-step local refinement that corrects residual errors. Concretely, the coarse stage learns a conditional posterior over endpoint velocity to transform Gaussian noise into a structured initialization, while the fine stage performs a fixed-time refinement from this initialization. To stabilize training, we introduce a stepwise strategy that first learns a controlled coarse predictor and then performs joint optimization. Experiments on CALVIN and LIBERO show that our method establishes a strong efficiency-performance frontier under low-NFE (Number of Function Evaluations) regimes: it consistently outperforms existing NFE=2 methods, matches or surpasses the NFE=10

π_{0.5}

baseline on several metrics, reduces action sampling latency by 75.4%, and achieves the best average real-robot success rate of 83.0%, outperforming MIP by 19.5 points and

π_{0.5}

by 4.0 points. These results suggest that structured, coarse-to-fine generation enables both strong performance and efficient inference. Our code is available at https://github.com/EmbodiedAI-RoboTron/CF-VLA.

Abstract

Explore similar work

CF-VLA: Efficient Coarse-to-Fine Action Generation for Vision-Language-Action Policies

TBD-VLA: Temporal Block Diffusion Vision Language Action Model

dVLA-RL: Reinforcement Learning over Denoising Trajectories for Discrete Diffusion Vision-Language-Action Models