ReSync: Re-Aligning the Two Clocks of Asynchronous World-Action Models

Xi Lin; Feihong Zhang; Yulong Shi; Yanghong Mei; Zuxing Lu; Xiaofan Zhu; Zihao Liang; Zhirui Gao

cs.ROSep 27, 2026

ReSync: Re-Aligning the Two Clocks of Asynchronous World-Action Models

Authors: Xi Lin, Feihong Zhang, Yulong Shi, Yanghong Mei, Zuxing Lu, Xiaofan Zhu, Zihao Liang, Zhirui Gao, +1 more

Organizations: Johns Hopkins University · Tsinghua University · Yinwang Intelligent Technology Co., Ltd. · University of Chinese Academy of Sciences

Abstract

Jointly generating future video and actions has become a standard recipe for world-action models, and the strongest systems denoise the two streams on separate schedules: actions are decoded in few steps so control stays fast, while the video stream runs longer to keep the predicted future sharp. The design is deliberate, but it leaves the two streams on different clocks, and an action can become executable while the future that should justify it is still largely unresolved. We formalize this as a two-clock view of asynchronous inference and introduce the commitment-evidence gap, a quantity read directly from a model's own sampling schedule rather than measured by search. The gap is predictive: as it widens, candidate utility becomes harder to identify and extra candidate sampling buys less, while advancing the world stream buys more, and the two cross. Spending more world computation is therefore not simply better. The useful interval is closed at both ends, and both ends can be read off the schedule before any rollout. ReSync places the computation inside it: hold the action state, advance only the world within the supported window, then resume native denoising. No parameters change and no candidates are compared. On a frozen paired RoboCasa panel this improves success by 4.48 points, while an equal-compute control that waits without advancing the world does not move, and the same rule transfers to a second benchmark and a second backbone without retuning.

Figures & tables

Figure 1: Action commitment can precede useful world evidence. Native X-WAM commits after the ten-step action schedule finishes; ReSync holds that action state, advances the world inside supported denoising time, and resumes native action denoising before commitment. Photographs are recorded execution frames; blur indicates evidence resolution schematically and is not model-decoded output.

Figure 2: The preferred compute axis reverses as the gap grows. Larger gaps coincide with lower

\kappa

and smaller selector gains, but larger evidence-advancement gains. The reversal lies between the measured

G=0.44

and

G=0.58

schedules; the marker interpolates.

Setting	$T_{O}$	$\hat{r}_{\mathrm{sched}}$	$G$	BASE	$\kappa$	Selector $\Delta$ pp	ReSync $\Delta$ pp
Sync-10	10	1.00	0.00	76.15	0.268	+4.06	+0.21 †
Mid-14	14	0.71	0.29	79.48	0.211	+3.54	+0.83
Mid-18	18	0.56	0.44	81.04	0.134	+3.02	+1.67
Mid-24	24	0.42	0.58	81.25	0.094	+1.98	+2.71
Mid-30	30	0.33	0.67	80.73	0.061	+1.04	+3.54
X-WAM (native)	50	0.20	0.80	80.31	0.0187	+0.41	+4.48

Table 1: Empirical crossover across six schedules of one checkpoint. Only

T_{O}

changes. The selector uses a fixed

N=32

bank; ReSync inserts

M=\{0,2,4,8,12,24\}

world updates. These budgets are not matched; Table 2 supplies the equal-budget comparison.

Figure 3: ReSync advances only the evidence clock inside the usable window. The action state is cached, the world advances, and native action denoising resumes while recovery budget remains.

Method	SR	$\Delta$ pp	Rescue/Break	95% CI	Compute
BASE	80.31	0.00	–	–	1.00 $\times$
allocation $r25,N4$	82.71	+2.40	64/41	[+0.18,+4.61]	1.75 $\times$
wait-only	80.52	+0.21	43/41	[-1.84,+2.19]	3.40 $\times$
matched resampling	82.08	+1.77	66/49	[-0.27,+3.83]	3.40 $\times$
Action-34	81.56	+1.25	50/38	[-0.84,+3.35]	3.40 $\times$
Sync-34	82.29	+1.98	57/38	[-0.10,+4.05]	3.40 $\times$

Table 2: Where the compute goes matters. Every

3.40\times

row uses 34 shared-DiT forwards; Sync-50 is shown as a higher-cost reference. Confidence intervals are clustered over physical-state clusters.

$M$	0	8	16	24	32	40
SR (%)	80.31	81.67	83.33	84.79	83.96	82.29
$\Delta$ pp	0.00	+1.35	+3.02	+4.48	+3.65	+1.98

Table 3: World-side compute has an interior optimum.

M=24

is the largest supported advancement at

j^{\star}=6

;

M=32

and 40 deliberately cross that boundary.

Figure 4: Schedule constraints carve out the high-return region. Shading marks support constraints rather than success. The schedule-selected

(6,24)

point lies on the same high-performing plateau as the best safe cells.

Figure 5: Late placement loses the recovery margin. Performance collapses when only one or two native action updates remain.

Figure 6: Evidence advancement improves until the supported boundary, then gives the gain back. Candidate allocation saturates earlier, while ReSync peaks at the interior

M=24

point.

Schedule $(J,T_{O})$	Rule $(j^{\star},M^{\star})$	BASE	$M^{\star}\!-\!8$ SR	Rule SR	$M^{\star}\!+\!8$ SR	Warm-start SR	Rule $\Delta$ pp	95% CI
$(8,40)$	$(4,16)$	79.58	82.29	83.65	83.02	82.19	+4.06	[+1.72,+6.35]
$(12,60)$	$(8,32)$	80.94	84.06	85.21	84.58	83.65	+4.27	[+1.96,+6.61]

Table 4: Frozen-rule validation on two unseen schedule pairs. For each pair we evaluate the rule-selected point, its

M\pm 8

neighbors, and an equal-budget world warm-start; the

M+8

point deliberately crosses the schedule-derived support boundary. CIs are cluster-bootstrap intervals for the selected point relative to that schedule’s BASE.

Backbone	Benchmark	$G$	$\kappa$	$(j^{\star},M)$	BASE	ReSync	$\Delta$ pp
X-WAM	RoboCasa	0.80	0.0187	$(6,24)$	80.31	84.79	+4.48
X-WAM	RoboTwin 2.0	0.80	$-0.079^{\ddagger}$	$(6,24)$	90.33	93.00	+2.67
Cosmos Policy	RoboCasa	0.60	0.082	$(10,10)$	66.25	70.63	+4.38

Table 5: The same timing rule transfers across benchmark and backbone. RoboTwin reuses the native X-WAM schedule; Cosmos uses the fixed constructed 14/35 interface, with BASE and ReSync executing the same 14 action updates.

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Method	Allocated object	Signal	Primary axis	Evidence at commit
RoboMonkey	candidates	VLM verifier	candidate	resolved (implicit)
ADV	drafts	self-verification	candidate	resolved (implicit)
GeoBoN	WAM trajectories	geometry	candidate	resolved (implicit)
CAPS	search	SNR	search depth	unmodeled
SANTS	video steps	scheduler	stopping time	explicit noise
NoiseGate	latent timesteps	learned gate	schedule	explicit noise

Appendix

Table 6: The neighboring methods act on different decision axes. The final column describes the evidence state assumed or modeled at commitment.

Method	SR	$\Delta$ pp	95% CI	Rescue/Break	Compute
BASE	80.31	0.00	–	–	1.00 $\times$
allocation $r25,N4$	82.71	+2.40	[+0.18,+4.61]	64/41	1.75 $\times$
ReSync + allocation	85.94	+5.63	[+3.17,+8.12]	95/41	4.15 $\times$

Appendix

Table 7: Allocation operating points on the updated paired panel.

Figure 7: Paired RoboCasa simulator executions. Each pair starts from the same reset state and shows successful and failed executions. Frames are aligned at the initial state, gripper closure, divergence, and stable outcome.

Figure 8: Paired RoboTwin 2.0 simulator executions. Each pair shares the initial state and shows successful and failed executions. Frames are aligned at gripper closure and subsequent task events. Source frames retain their native $640\times 480$ resolution.

Figure 9: Schematic schedule-progress ladder. A rendered future frame is degraded by controlled downsampling and blur to visualize the intuition behind increasing schedule progress. L0 is the reference frame; L1–L5 are annotated with schedule-progress values 0.20, 0.32, 0.44, 0.56, and 0.68. These panels are an illustration of the proxy and are not model-decoded world states.

Task	BASE	ReSync	$\Delta$
Cup placement	6/10	7/10	+10 pp
Block placement	6/10	8/10	+20 pp
Cluttered reaching	7/10	7/10	0 pp
Overall	19/30	22/30	+10.0 pp

Appendix

Table 8: Physical-robot policy rollouts. BASE and ReSync share the same checkpoint, resets, and action interface.

Figure 10: Paired physical-robot policy rollouts. BASE/ReSync pairs share task and reset conditions and are aligned at commitment and subsequent task events.

Figure 11: The selector screen is directional rather than a headline result. Mean paired deltas are shown for 14 SAR and 12 random-selector configurations. Random selection wins 9 of the 12 matched schedule/depth pairs.

Figure 12: Routing-map consistency check. Reachable and unvisited bins follow the executed timestep map across three seeds, including the dead- $k5$ /live- $k6$ pattern.

Metric	$t_{0}$	$t_{1}$	$t_{2}$	$t_{3}$	$t_{4}$	$t_{5}$	$t_{6}$	$t_{7}$	$t_{8}$	$t_{9}$
kNN density	.435	.424	.428	.433	.436	.426	.431	.431	.428	.432
Mahalanobis	.433	.483	.505	.479	.462	.460	.453	.493	.500	.471
KDE	.516	.516	.513	.514	.516	.516	.516	.516	.516	.516
Dist.-center	.489	.489	.486	.486	.486	.486	.489	.486	.486	.486
Local agree.	.482	.491	.485	.497	.488	.491	.488	.482	.483	.494
Full dist.-center	.449	.449	.449	.449	.449	.449	.449	.449	.449	.449

Appendix

Table 9: In-process geometry remains weak at every executed action position. Values are within-set AUC.

$j$	$t_{a}^{j}$	$\psi(j)$	$t_{O}^{\psi(j)}$	$M_{\max}$	recovery
0	999	0	999	0	9
1	978	1	995	4	8
2	952	2	991	8	7
3	920	3	987	12	6
4	882	4	982	16	5
5	833	5	978	20	4

Appendix

Table 10: The schedule makes the X-WAM operating point directly auditable.

R_{\min}=3

removes

j=7,8

and leaves

j^{\star}=6

as the latest safe maximizer.

$M$	Schedule progress proxy	SR	$\Delta$ pp	Compute	Latency / decision	Effective Hz
0	0.20	80.31	0.00	1.00 $\times$	2.68 s	0.37
8	0.36	81.67	+1.35	1.80 $\times$	4.74 s	0.21
16	0.52	83.33	+3.02	2.60 $\times$	6.83 s	0.15
24	0.68	84.79	+4.48	3.40 $\times$	8.97 s	0.11
32	0.84	83.96	+3.65	4.20 $\times$	11.14 s	0.09
40	1.00	82.29	+1.98	5.00 $\times$	13.32 s	0.08

Appendix

Table 11: World advancement peaks at the training-supported boundary. Wall-clock measurements use one NVIDIA H800 80GB GPU with identical batching and compilation state;

M>24

deliberately crosses joint training support.

Figure 13: Specialization is learned, not inherited. Cross-bin variance of the effective gain during the coupling-adapter support test. It is zero at initialization and rises monotonically; the unvisited bins remain at 0.50.

Contrast	$\Delta$ pp	cluster-bootstrap 95% CI	McNemar $p$
random placement vs BASE	+0.94	[-1.16,+3.07]	0.39
early vs random	+0.28	[-1.51,+2.11]	0.76
late vs random	-0.86	[-2.84,+1.13]	0.35
decay vs random	+1.31	[-0.58,+3.22]	0.17

Appendix

Table 12: Schedule shaping changes point estimates, not clustered inference. McNemar is a secondary diagnostic.

Explore similar work

Sep 8, 2026cs.CV

ActionSplice: In-Flight Action Editing for Interactive World Models

Chunk-autoregressive video world models typically condition each generated chunk on one action. An action received during sampling must therefore wait for the next chunk, condition future solver evaluations on a state produced under the previous action, or trigger rollback that repeats completed evaluations. We introduce ActionSplice, an inference framework that formulates this problem as Counterfactual State Transport (CST). A lightweight corrector transports the interrupted backbone-native representation toward the matched state induced by the revised action at the same solver step. The world model and sampler remain frozen, and sampling resumes without replaying completed evaluations. The retargeting variant

\mathrm{CST}*{R}

updates the entire active chunk, while the temporal-splicing variant

\mathrm{CST}*{T}

preserves a temporal prefix and updates only the suffix. Across minWM-Wan Action2V and HY-WM1.5,

\mathrm{CST}*{R}

reduces rollback-relative LPIPS by 61.5% and 75.9% relative to direct condition swapping.

\mathrm{CST}*{T}

reduces suffix LPIPS by 56.1% and 77.5%, respectively, while providing

2.73\times

and

1.69\times

pixel-ready speedups over waiting. Under the HY-WorldPlay protocol,

\mathrm{CST}_{R}

obtains a PSNR of 25.66 dB, an SSIM of 0.6902, and an LPIPS of 0.1337 against the original rollout.

Pardis Taghavi, Tingyu Guo, Jonas Lossner +2

Texas A&M University

Video World Models Camera-Conditioned Video Generation Action Expert Autoregressive Video Generation Functional Code Solvers

Ask PDF

Sep 28, 2026cs.CV

What Makes World Action Models Generalize? An Empirical Study of Test-Time Future Modeling

World action models (WAMs) predict the future alongside actions during training. Due to the heavy computation cost of video denoising, whether the future must still be generated during inference is disputed: Explicit WAMs denoise it into clean frames along with every action chunk, whereas Latent WAMs discard it entirely for acceleration. We find that latent WAMs, despite matching explicit ones on in-distribution tasks, fail to retain the generalization benefits that originally motivated WAMs. To demonstrate this, we evaluate generalization along three axes: environmental perturbation, data efficiency, and task generalization. Controlled comparisons with a matched backbone, training data, and budget reveal consistent degradation across all three axes when the action expert no longer conditions on future representations. Further analysis shows that the gap arises almost entirely from the first denoising step: the benefit comes from preparing the future, not generating it. We therefore propose Simple-WAM, which simplifies future modeling into a single forward pass of fully noised video tokens and adapts the training-time noise schedule to this inference behavior. Across simulation and real-world tasks, Simple-WAM achieves the best of both worlds, leading explicit WAMs in generalization performance with efficiency comparable to Latent WAMs. Project Page: https://zrporz.github.io/Simple-WAM-Web/

Renping Zhou, Zanlin Ni, Zihao Fan +8

Leap Lab, Tsinghua University · University of Science and Technology of China · Beijing Institute of Technology

Latent World Models World Models Iterative Denoising Methods Empirical Study

Ask PDF

Sep 24, 2026cs.RO

Streaming-WAM: Action-Conditioned World-Action Model for Asynchronous Robot Manipulation

World action models (WAMs) that use future visual prediction at inference time incur substantial generation costs. Asynchronous execution reduces waiting by overlapping inference with robot motion, but visual predictions used for subsequent action generation must anticipate the effects of actions already scheduled for execution during inference. We introduce Streaming-WAM, which couples action-conditioned world modeling with asynchronous robot control to account for committed actions in future visual prediction. At each streaming update, the model conditions future visual prediction on the latest observation and the committed actions, which form the fixed prefix of the next action chunk. The resulting action-conditioned visual features guide generation of the remaining actions within the same joint update, so the continuation is informed by the scene changes expected during execution of the fixed prefix. On LIBERO, Streaming-WAM achieves an average success rate of 98.35% and reduces mean episode time by a factor of 2.93 relative to Fast-WAM. On the real-world Stamp Paper task, mean episode time falls from 90 s with synchronous Joint-WAM to 38 s with Streaming-WAM. These results show that Streaming-WAM supports efficient asynchronous control while maintaining high task success rates.

Xuyao Huang, Yixuan Wang, Zengyao Ye +4

SJTU · SDU · ECUST +1

Action Chunks Robotic Manipulation Robot Systems

Ask PDF

$j$	$t_{a}^{j}$	$\psi(j)$	$t_{O}^{\psi(j)}$	$M_{\max}$	recovery
0	999	0	999	0	9
1	978	1	995	4	8
2	952	2	991	8	7
3	920	3	987	12	6
4	882	4	982	16	5
5	833	5	978	20	4

$j$	$t_{a}^{j}$	$\psi(j)$	$t_{O}^{\psi(j)}$	$M_{\max}$	recovery
0	999	0	999	0	9
1	978	1	995	4	8
2	952	2	991	8	7
3	920	3	987	12	6
4	882	4	982	16	5
5	833	5	978	20	4

$j$	$t_{a}^{j}$	$\psi(j)$	$t_{O}^{\psi(j)}$	$M_{\max}$	recovery
0	999	0	999	0	9
1	978	1	995	4	8
2	952	2	991	8	7
3	920	3	987	12	6
4	882	4	982	16	5
5	833	5	978	20	4