The Unexpired Plan: A Free Monitor for Accelerated Diffusion Policies

Yi Zhao; Sebastian Scherer

cs.ROOct 5, 2026

The Unexpired Plan: A Free Monitor for Accelerated Diffusion Policies

Authors: Yi Zhao, Sebastian Scherer

Organizations: AirLab, The Robotics Institute Carnegie Mellon University, Pittsburgh, PA, USA

Abstract

Training-free acceleration of a diffusion policy is accepted when an internal similarity signal reports that the shortcut changed nothing. We price every monitor a control loop can afford in closed-loop success rather than feature distance, over $114$ accelerator configurations and four policy families: two accelerators each clear their own gate's bar and log every reuse as certified, while on one task one finishes every episode and the other none. A monitor is two designs, not one --- the statistic it reads, and what it does when that statistic fires. Published gates re-arm after every rejection, and under that response even an oracle handed every forward pass and the exact local action error loses twenty points; absorbing the same statistic costs one point, and most of its speed. What makes a response that never forgets affordable is a statistic that rarely fires, and what it must measure is deviation from the policy the accelerator replaced --- which a chunked policy has already paid for, its last plan not yet expired and free to read. Guarding every call this way cuts per-call compute by $1.55$ -- $3.09\times$ , where any reference-requiring check at the same coverage would have to stop accelerating altogether. On three of our four families the schedule alone already holds the pre-stated $\pm2$ -point margin. What the monitor is measurably worth shows in three places: on the fourth family, where it rescues the candidate selection landed on; on four configurations it did not select; and on a contact-rich fifth family, chosen where the schedule was expected to fail and run after every design choice was frozen, where no unmonitored arm at its speed holds the margin and the monitored one does.

Figures & tables

Fig. 1: A gate cannot be read from its own log. One LIBERO-10 task, same initial states: each accelerator scores its own cached quantity against its own bar, so both certify every reuse, while one finishes every episode and the other none. The last rows read the quantity § IV monitors against one bar shared by all.

Free, but not the deviation the accelerator introduced
					$\Delta$ SR by response
What the statistic measures	Monitor	Permission	accel. (%)	base	re-arming	absorbing
agreement of cached features	cross-call gate, deployed	admits	$53.8\%$	$97.0$	$-91.0$	$-88.1$
agreement of cached features	same statistic, veto-only	vetoes only	$61.4\%$	$97.0$	$-92.3$	$-89.4$
the call’s own inputs	observation gate, safe settings	admits	$1.3\%$	$98.2$	$-6.4$	$-5.9$
the call’s own inputs	observation gate, at the coin’s rate	admits	$46.2\%$	$98.2$	$-91.1$	$-88.4$

TABLE I: Every monitor a control loop can afford, priced in closed-loop success.

\Delta

SR is paired against base on identical initial states over the episode count of § V ’s protocol; the half-widths

h=\pm 1.1\,\mathrm{pp}

at base

98.2

and

h=\pm 0.9\,\mathrm{pp}

97.0

hold near base only. accel. is the share that ran accelerated under re-arming ; bold marks our rows.

Fig. 2: The method. (a) A reference plan spans

H

indices while only

H_{\mathrm{exec}}

execute, so the next

m-1

calls land on indices it has predicted and are read against them at no forward pass. (b)

\delta

is Eq. ( 2 ),

\lambda

the repetition check of § IV-B ; re-emitting the plan it is compared against reads

\delta\approx 0

, because

e\approx-d

. (c) A veto costs a rung before it costs the accelerator; a rung is a slower configuration.

unmonitored — no guarantee at any operating point
		Cosmos $m{=}2$ , $98.2$		$\pi_{0.5}$ $m{=}5$ , $97.0$		PushT $m{=}2$ , $58.0$		MetaWorld $m{=}4$ , $94.2$
Method	guarded	$\Delta$ SR	$S_{\mathrm{eff}}$	$\Delta$ SR	$S_{\mathrm{eff}}$	$\Delta$ SR	$S_{\mathrm{eff}}$	$\Delta$ SR	$S_{\mathrm{eff}}$
paired half-width		$\pm 1.1\,\mathrm{pp}$		$\pm 0.9\,\mathrm{pp}$		$\pm 2.6\,\mathrm{pp}$		$\pm 1.4\,\mathrm{pp}$
Feature cache [ 3 ] , selected	$0\%$	$+0.1$ †	$1.40\times$	$-0.3$ †	$1.20\times$	$-0.9$	$1.70\times$	$-0.2$ †	$1.50\times$
Chunk retrieval [ 7 ] , selected	$0\%$	$-0.6$ †	$1.70\times$	$-1.4$	$1.60\times$	$-1.8$	$2.40\times$	$-0.9$	$1.80\times$
Learned tolerance [ 14 ]	$0\%$	$-0.8$ †	$2.10\times$	$-1.1$	$1.90\times$	$-3.1$	$5.20\times$	$-0.6$	$3.40\times$

TABLE II: Closed-loop comparison. Column heads: family,

m

, base success rate; first row: paired half-width. guarded is § IV-B ’s guarded share; rows charging no reference call report raw

S

; grey means unranked. † marks

|\Delta\mathrm{SR}|+h<2

strictly ; bold and underline rank

S_{\mathrm{eff}}

among fully guarded arms whose point estimate is inside the margin on every family. The

\pi_{0.5}

pair (

97.0

-91.0

) is the collapse of Fig. 1 .

		$\Delta$ SR, monitor:			monitored
Rejected arm (family)	raw $S$	none	schedule	no $\lambda$	$\Delta$ SR	$S_{\mathrm{eff}}$
Cross-call cache ( $\pi_{0.5}$ )	$1.70\times$	$-91.0$	$-79.2$	$-0.3$ †	$-0.3$ †	$1.05\times$
Step reduction (PushT)	$88.9\times$	$-58.0$	$-37.7$	$-1.4$	$-1.4$	$1.30\times$
Token reuse (Cosmos)	$2.60\times$	$-18.4$	$-12.0$	$-12.0$	$-0.6$ †	$1.10\times$
Feature cache (MetaWorld)	$2.20\times$	$-9.0$	$-7.5$	$-0.3$ †	$-0.3$ †	$1.18\times$

TABLE III: Four configurations the selection rejected, one per family, monitored unchanged.

\Delta

SR against each family’s base, half-widths as in Table II ; schedule runs the arm on the reference period with every check off; no

\lambda

keeps only Eq. ( 2 ); the

\pi_{0.5}

arm is Fig. 1 ’s collapse. The

-91.0

is measured.

ManiSkill3 insertion, $m{=}3$ , base $71.3$ , $6000$ episodes
Method	guarded	$\Delta$ SR	$S_{\mathrm{eff}}$
paired half-width		$\pm 1.2\,\mathrm{pp}$
Feature cache, selected	$0\%$	$-8.7$	$4.60\times$
Step reduction, matched to our speed	$0\%$	$-3.9$	$1.96\times$
Reference every call	$100\%$	$-0.3$ †	$1.00\times$
Paid deviation at every call	$100\%$	$-0.1$ †	$0.87\times$

TABLE IV: A fifth family, run after the design was frozen: the Diffusion Policy [ 26 ] on ManiSkill3’s tight-tolerance peg insertion [ 29 ] . Columns are Table II ’s, and the first body row gives the paired half-width, here inside the margin.

	no monitor		$+$ monitor
Selection	PT $\Delta$ SR	$\bar{S}_{\mathrm{eff}}$	PT $\Delta$ SR	$\bar{S}_{\mathrm{eff}}$
off-policy	$-31.2$	$7.56\times$	$-1.9$	$1.31\times$
on-policy	$-9.4$	$5.99\times$	$+1.2$	$1.93\times$
mean instead of tail	$-38.4$	$9.12\times$	$-22.8$	$1.93\times$
no $\lambda$	$-9.4$	$5.99\times$	$+1.2$	$1.93\times$
latch instead of ladder	$-9.4$	$5.99\times$	$+1.2$	$1.85\times$

TABLE V: Ablation: monitor crossed with selection above the rule, one design choice knocked out below it.

\Delta

SR is PushT’s, the family that separates them;

\bar{S}_{\mathrm{eff}}

is the geometric mean over all four, which Eq. ( 3 ) does not commute with; no monitor charges no reference call, so it reports raw

S

. The no

\lambda

row repeats the deployed row exactly.

Fig. 3: (a) The tail deviation (Eq. ( 1 )’s 95th percentile over an episode) of each of the

114

configurations of § III , catastrophic when it loses ten points or more; values below

0.001

pool at the left edge with their count; values within a marker width overprint. The dashed line is

\theta

, the dotted lines the range over which every verdict in Tables I and II holds. (b) Stepping the reference period

k

past full coverage: the share of the tolerance

n_{c}

that Eq. ( 4 ) spends at

p_{d}{=}1

, at the composite miss rate (solid) and across its interval (dotted), for Cosmos,

\pi_{0.5}

, PushT and MetaWorld at tolerances of

41

38

34

47

chunks. (c) What the step buys, Eq. ( 3 ) at raw

S

3.4

2.0

18.6

10.2

before the ladder’s vetoes; PushT deploys at Table II ’s

1.75\times

Explore similar work

Aug 3, 2026cs.LG

Disagree to Accelerate: Closing the Loop on Diffusion Feature Forecasts

Training-free feature forecasting accelerates diffusion sampling by predicting features at skipped denoising steps. Recent work has mainly focused on designing stronger forecasters. Yet forecast error varies sharply across steps, and open-loop caches trust the forecast in full at every skipped step. This fixed trust is what breaks as acceleration turns aggressive. The missing question is not only how to forecast better, but when and how much to trust a forecast. We show that reliability can be observed from the cache itself. Two forecasts agree where the feature trajectory is smooth, and they diverge where prediction turns hard. Their disagreement is a cheap runtime signal, and it costs no extra denoiser evaluation. Based on this signal, we introduce RACER, a training-free closed-loop controller with two responses. It continuously shrinks uncertain forecasts toward the last computed feature. At the riskiest steps, RACER refreshes the feature and repays the added evaluation by skipping a later scheduled one. We derive a deterministic error bound for the shrinkage and empirically evaluate its validity and tightness across acceleration regimes. At the same number of denoiser evaluations, RACER improves the strongest open-loop baseline across SD3.5-Large, FLUX.1-dev, Wan2.1-14B, and HunyuanVideo on DrawBench, VBench, and COCO. On SD3.5, we further show that RACER samples faster at equal quality. RACER generalizes across forecasting designs as well. For example, it recovers much of the quality lost on a Taylor base. These results show that reliable diffusion acceleration also depends on how forecasts are used. Code is available at https://github.com/LiZaiyuan0619/RACER

Yanchao Li, Jiaqing Xie, Ben Gao +6

1Nanjing University · 2Shanghai Artificial Intelligence Laboratory · 3North University of China +2

Iterative Denoising Methods Forecast Autonomous Racing Disagreement

Ask PDF

Jul 22, 2026cs.CV

Evolving Cache Schedules for Fast Diffusion Policy Inference

Diffusion policies achieve strong visuomotor control by iteratively denoising action chunks, but repeated denoising makes real-time deployment computationally demanding. Cache-based methods reduce inference cost by reusing intermediate activations, but existing training-free schedules typically allocate computation uniformly across blocks, ignoring heterogeneous redundancy across blocks and leading to a suboptimal performance-efficiency trade-off. To bridge this gap, we introduce Evolving Cache Schedules (EVO), a training-free acceleration framework that globally schedules cache refreshes via evolutionary search. EVO represents each candidate as a complete schedule over the block-timestep lattice. Thus, redundant transformer computations during iterative denoising can be skipped through cache reuse while preserving closed-loop rollout performance. To make the search practical, EVO introduces redundancy-aware initialization, which seeds the population with promising schedules, and target-conditioned early stopping, which verifies and terminates once a desired performance target is reached. The offline-optimized schedule can be directly plugged into pretrained diffusion policies without retraining. Extensive manipulation benchmarks show that EVO preserves near-full performance while substantially reducing computation, achieving up to 8.05x action-generation speedup and reducing FLOPs from 15.77G to as low as 1.96G. Source code is available at https://github.com/pillom/EVO.

Siying Wang, Kangye Ji, Di Wang +1

School of Telecommunications Engineering, Xidian University, Xi’an, China · Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · School of Computer Science and Technology, Xidian University, Xi’an, China +1

Diffusion Policies Iterative Denoising Methods Cache Visuomotor Control

Ask PDF

Jun 25, 2026cs.AI

ResilPhase: Plug-and-Play Phase Mapping and Noise-Resilient Macro-Trajectory Extrapolation for Diffusion Acceleration

The adoption of powerful diffusion models is hindered by their significant inference latency. Recent ``cache-then-forecast'' schemes alleviate this issue by accelerating DiTs using derivative-based polynomials, but they suffer from severe quality degradation at high acceleration ratios. Our analysis reveals its root cause: the discrete extrapolation performed on representations that are misaligned with the continuous diffusion trajectory and are numerically unstable. Thus, accelerated DiTs suffer from accumulated spatial errors, noisy derivative amplification, and high-order instability. We therefore reformulate accelerated inference as stable macro-trajectory extrapolation in ordinary differential equation (ODE) space. Instead of predicting intermediate features, we align forecasting with the model's Global Drift (GD), i.e., the end-to-end state evolution, thereby eliminating feature inconsistency and memory overhead. However, even this smooth macro-trajectory remains vulnerable to the derivative fallacy: its higher-order temporal derivatives are intrinsically noisy. Thus, we introduce a derivative-free barycentric Lagrange extrapolator to effectively bypass derivative instability and approximation error. We further propose a bounded Phase Mapping that regularizes the extrapolation domain, suppressing oscillatory error growth. These elements collectively constitute ResilPhase, a noise-resilient acceleration framework. Experiments on FLUX.1-dev and HunyuanVideo demonstrate state-of-the-art fidelity under aggressive acceleration ratios.

Qicheng Zhao, Yu Li, Qi Sun +1

Zhejiang University, Hangzhou, China

Residual Conditional Diffusion Model Diffusion Models Implicit Differentiation Accelerometer Data Resilience 3D Mapping

Ask PDF