PAPER REPORTENAll readings ↗

C3^3ache: Accelerating World Action Models with Cross Inference Chunk Cache

English reading report: Method, equations, original figures, experiments and reproducibility.

Authors: Weisen Zhao; Lam Nguyen; Zhicong Lu; Yuzhang Shang

Affiliations: George Mason University; University of Central Florida

Source: 2606.08962 ↗ · Catalog record

Reading: 178 / 558 · 5 original figures & tables · ~18 min ·

1. Paper overview

In one sentence: C³ache reuses an action transformer's residual across control cycles, reducing inference cost when the cache remains useful but risking large success losses when it becomes stale or replaces late denoising steps. e-flowe-backbonee-residuale-cachee-similaritye-liberoe-robotwine-ablatione-conditioning-gap

At a glanceWhat to know
Research problem
Source description

Fast-WAM already removes test-time future-video synthesis, yet repeatedly denoises each action chunk. C³ache asks whether temporally adjacent control cycles repeat enough internal computation to reuse it across chunks. The target is action-inference latency, not a new learned dynamics model or video-quality objective. e-probleme-backbonee-flow

Core mechanism
Source description

Characterizes adjacent-chunk similarity of residuals, final-block outputs and predicted velocities at matched denoising steps across LIBERO. e-similarity

A key reported resultLIBERO: four-suite action success and inference speed: 97.10% success; 2.51× speedup.

Average task success (%) and total wall-clock inference speedup. All four suites, 2,000 evaluation episodes; Fast-WAM; 10 denoising steps, horizon 32; cache [0,7], τ=0.

Uncached Fast-WAM: 96.90%, 1.00×; difference +0.20 percentage points. This is the strongest reported speedup, not evidence of a statistically established accuracy improvement; no uncertainty is reported. e-evaluatione-computee-timinge-liberoe-reporting

Reading caution
Source description

Figure 4 shows severe degradation when caching extends into the final steps. RoboTwin falls at nine and ten cached steps; LIBERO collapses at ten across refresh settings. Exact nine-/ten-step values are not tabulated. Manual refresh tuning remains necessary. e-ablatione-limitations

Core contributions

  • Source description

    Characterizes adjacent-chunk similarity of residuals, final-block outputs and predicted velocities at matched denoising steps across LIBERO. e-similarity

  • Author claim

    Reuses the aggregate DiT residual while retaining the current action embedding and output head. Compatibility with within-chunk caching is claimed, but no combined-method comparison is reported. e-probleme-cachee-reporting

Figure 3. One residual per denoising step, shared across successive action chunks. Original paper, p. 6 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Follow the upper green path from h_0 through the L DiT blocks to h_L. The cache receives ΔR, defined in the caption and Equation (6) as h_L minus h_0. On the orange path, the bypass arrow and cache arrow meet at h_L=h_0+R̃; no DiT block is evaluated at that cached step. Below, horizontal arrows connect matching step positions across chunks. Green chunks initialize or refresh, and orange chunks reuse. The schematic draws all K steps, but the experiments select a prefix of zero-indexed steps and still compute the tail. For positive τ, the next full refresh occurs after τ chunks. e-residuale-cachee-conditioning-gape-computee-timing

What it supports. The saved work is the action transformer's aggregate residual computation. Current action embeddings, the output head and scheduler still run, while selected steps reuse a stored additive contribution. This explains how the method can reduce computation without directly replaying a previously generated action chunk.

Where the evidence stops. The 'new obs' label is under-specified: Section 3.1 defines h_0 as noisy-action embeddings and locates video conditioning inside the skipped attention. Figure 3 and the Equation (7) prose assert fresh observation content in h_0 without specifying that pathway. Also, τ=0 is a separate no-refresh case.

2. Motivation

2.1 The problem and the proposed response

Source description

Fast-WAM already removes test-time future-video synthesis, yet repeatedly denoises each action chunk. C³ache asks whether temporally adjacent control cycles repeat enough internal computation to reuse it across chunks. The target is action-inference latency, not a new learned dynamics model or video-quality objective. e-probleme-backbonee-flow

2.2 What this reading follows

A robot policy often predicts several actions, executes only part of them, observes the scene again, and repeats. Fast-WAM already avoids generating future video during that loop, but its action expert still runs a transformer repeatedly to denoise every new chunk. C³ache asks which internal computation can survive from one chunk to the next. Its answer is the aggregate residual added by the transformer stack, stored separately for each denoising step. The five visuals below connect that choice to measured similarity, contrasting benchmark results, and a sharp late-step failure boundary. The useful result is a configuration-dependent speed–success tradeoff; the supplied source also leaves the cached observation-conditioning path unclear. e-flowe-backbonee-residuale-cachee-similaritye-liberoe-robotwine-ablatione-conditioning-gap

3. Research context

We place the paper in the collection through its world–action interface. The catalog labels and the reading’s assessment are shown separately.

Catalog dimensionRecorded classification
Major categoryWAMs
ArchitectureDual-system
Prediction paradigmOther mechanisms
QuadrantOutside quadrants

3.1 Evidence-based assessment

Supports the recorded classification

Reader analysis

The efficiency classification is supported. Fast-WAM has distinct video and action branches linked through joint attention, supporting Dual-system rather than inferring one model from co-training alone. C³ache is an inference cache on direct action generation; it neither jointly predicts future video at inference nor derives actions through inverse dynamics. Other mechanisms and Outside quadrants therefore fit this acceleration method. e-backbonee-residuale-cache

This is the collection’s architectural analysis, not a new related-work survey. Benchmark comparisons and their protocols appear in Section 6.

4. Problem formulation

4.1 Inputs and outputs

InputsOutputs
  • Fresh observation encoded as video-token context for each control cycle; noisy action-token embeddings and denoising time.
  • A separate cached residual for each selected denoising step, initialized by full computation.
  • A generated action chunk with horizon 32; the control loop executes a prefix before taking another observation.
  • Updated per-step residual caches on refresh chunks.

4.2 Equations and their role

x(k)c=x(k1)c+(tktk1)vkc,vkc=vθ ⁣(x(k1)c,tk1,ctxc)x^{c}_{(k)}=x^{c}_{(k-1)}+(t_k-t_{k-1})v^{c}_{k},\qquad v^{c}_{k}=v_{\theta}\!\left(x^{c}_{(k-1)},t_{k-1},\mathrm{ctx}^{c}\right)
Equation (2): c indexes chunks, k denoising steps, x the current action iterate, t the flow time, ctx the chunk context, and vθ the trained velocity predictor. The process starts from Gaussian noise and integrates toward actions. e-flow
R=hLh0,vkc=Head ⁣(h0c,k+Rkc)R=h_L-h_0,\qquad v^{c}_{k}=\operatorname{Head}\!\left(h^{c,k}_0+R^{c}_{k}\right)
Equations (3)–(4): h_0 is the action-token embedding, h_L the output after L DiT blocks, R their aggregate residual, and Head the linear velocity projection. e-residual
hLc,k=h0c,k+R~c,kh^{c',k'}_L=h^{c',k'}_0+\widetilde R_{c,k}
Equation (7) reconstructs the cached stack output using stored residual R̃ from an earlier full chunk c. The method and Figure 3 specify matching denoising steps, so k′=k for cross-chunk reuse; c′ denotes the later chunk. This is an approximation to fresh stack computation. e-cache

5. Method in detail

5.1 1. Locate the remaining repeated computation

Source description

Fast-WAM supplies the starting point: video co-training informs its representations, but action inference does not synthesize a future video. A video branch encodes the current observation once, and the action branch integrates a learned flow from Gaussian noise toward an action chunk. In the supplied formulation, the velocity predictor is trained on the difference between clean actions and noise. At inference, each integration step embeds the current noisy action iterate, runs the action transformer, and projects its output to velocity. Only part of the resulting action chunk is executed before the next observation arrives. C³ache leaves that control loop and the learned weights intact. It targets the repeated transformer stack within the action expert, while observation encoding, the velocity head and scheduler remain in the pipeline. e-backbonee-flowe-residuale-probleme-timing

Figure 2. Similarity motivates reuse, while late-step variability warns against caching the full trajectory. Original paper, p. 5 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read the left and right panels together: they summarize the same adjacent-chunk comparisons at each matched denoising step. Blue circles represent residual R, orange squares final-block output h_L, and green triangles predicted velocity. The left panel reports average cosine similarity; the right reports its standard deviation across the reported chunk sample, not uncertainty on task success. Section 3.2 describes 2,000 LIBERO episodes and 29,924 inference chunks with ten denoising steps. The green curve is most similar on average, so residual selection is not justified by the largest cosine score. Its rationale instead depends on retaining a changing current input. e-similaritye-cachee-consistency

What it supports. Residuals are correlated across neighboring chunks, and their similarity becomes less consistent toward the denoising tail. That pattern motivates testing reuse early while recomputing later steps. It supports a caching hypothesis, but does not by itself establish acceptable action error or closed-loop success.

Where the evidence stops. Section 4.1 incorrectly says standard deviation drops past step 6; the plot and Section 3.2 show it rising. The aggregation also leaves first-chunk pair handling unspecified. Cosine similarity ignores magnitude and is not a guarantee of interchangeable predictions.

5.2 2. Distinguish a cached correction from a frozen prediction

Reader analysis

The residual identity h_L=h_0+R separates the current action embedding from the contribution of the transformer stack. Directly caching h_L would hold the head's input fixed at a matched step; directly caching velocity would hold its output fixed. Reusing only R lets the current h_0 change, which is the source's reason for preferring it even though velocity has higher cosine similarity. A reader interpretation is that this preserves some flexibility while approximating the costly correction. It does not prove responsiveness to fresh visual input. Section 3.1 places observation tokens in joint attention, while Figure 3 labels h_0 as carrying the new observation even when that attention is skipped. This distinction makes a controlled observation-sensitivity check more informative than another similarity plot alone. e-residuale-cachee-similaritye-conditioning-gap

5.3 3. Read cache age and denoising position separately

Reader analysis

Two choices determine the approximation: which step-specific residuals are reused, and when those entries are refreshed across chunks. The equations index steps from one to K, while the experiments label ten steps from zero to nine; the inclusive range [0,7] therefore caches eight steps and leaves two freshly computed. For positive τ, a full chunk refreshes the cache every τ chunks. The special τ=0 case instead freezes the initialized entries; it cannot be implemented by evaluating modulo zero in Equation (8). LIBERO's strongest speed result uses that special case, but RoboTwin's corresponding setting loses 15.55 percentage points of average success. My interpretation is that cache age and step position need separate validation: a favorable average correlation does not identify a universally safe refresh schedule or reusable denoising tail. e-flowe-cachee-computee-liberoe-robotwine-ablation

5.4 Training and inference

During training

Source description

C³ache adds no training, losses or weight updates. Its inherited flow expert learns clean-action-minus-noise velocity, and Fast-WAM retains video co-training. The supplied paper does not provide the backbone's full data mixture, optimizer, loss weighting or training schedule. e-probleme-flowe-backbonee-reporting

During inference

Source description

The first chunk is fully computed. For positive τ, refresh every τ chunks; τ=0 specially means no further refresh, not refresh every step. Main configurations reuse zero-indexed steps [0,6] or [0,7] out of ten; remaining tail steps stay fully computed. e-cachee-compute

Source description

VAE encoding, video prefill, action embedding, velocity projection and scheduler updates remain in the measured inference pipeline. Predicted action generation and execution are separate: only part of a generated chunk is executed before observation feedback. e-timinge-flow

5.5 Implementation flow

  1. Encode and denoise

    Fast-WAM encodes observations once per chunk. Embedded noisy actions enter an action expert whose DiT joint attention reads action and video tokens. The linear output head predicts flow velocity; iterative integration turns fresh Gaussian noise into actions. e-backbonee-flowe-residual

  2. Store the stack's additive contribution

    A full pass computes all L blocks, then stores their aggregate output-minus-input residual. This is neither an action cache nor the video branch's KV cache: it approximates what the whole action-transformer stack adds. e-residuale-cachee-timing

  3. Reuse at the matching step

    A cached pass skips all L blocks, adds the stored residual to the current action embedding, and applies the velocity head. Each selected step has its own residual; refreshing updates these entries together on a full chunk. e-cache

6. Experiments & results

C³ache accelerates Fast-WAM by carrying a DiT residual cache across successive action chunks at matching denoising steps. It changes inference without retraining. LIBERO reaches 2.51× total inference speedup at 97.10% success, but the same aggressive configuration sharply reduces RoboTwin success; refresh and late-step computation are therefore central to the result.

6.1 Read the original evidence

Table 1. LIBERO tolerates an eight-step cache with no refresh after initialization. Original paper, p. 7 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Start at the Fast-WAM row, then compare each cache range against that shared baseline. The inclusive range [0,6] covers seven steps; [0,7] covers eight. Refresh 0 means that the first chunk populates the cache and later chunks do not refresh it. Read the four suite columns before the average, because an average near baseline need not preserve every suite. ΔSR is a difference in percentage points. The bold [0,7], refresh 0 row pairs 97.10% average success with 2.51× speedup. Evaluation covers 2,000 episodes across the four suites, and speedup includes the full inference pipeline. e-liberoe-evaluatione-cachee-computee-timinge-reporting

What it supports. At the fastest listed setting, average success is 97.10% versus 96.90% for uncached Fast-WAM, a +0.20-point difference at 2.51× speedup. The result demonstrates a favorable operating point on LIBERO. It does not show that eliminating refresh is universally appropriate across tasks or benchmarks.

Where the evidence stops. No confidence intervals accompany these small success differences. The row includes Goal 96.00% and Long 95.20%, below their baseline 96.40% and 95.40%. Total inference speedup is not a demonstrated physical control frequency or a per-call latency distribution.

Table 2. RoboTwin exposes the cost of extending the cache and delaying refresh. Original paper, p. 8 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. Read the clean and random columns as different evaluation conditions. Section 4.1 uses 50 tasks and 20 seeded episodes per task per setting. Clean randomizes object placement; random additionally changes visual and scene conditions. Within the seven-step [0,6] block, refresh 0 gives the largest speedup, whereas refresh 4 gives the highest average success. Then move to [0,7]: the configuration that worked well on LIBERO now suffers a large drop without refresh. Compare rows at the same cache range when isolating refresh effects, and compare equal refresh settings when studying the extra cached step. All speedups use uncached Fast-WAM as baseline. e-robotwine-liberoe-evaluatione-timinge-limitationse-reporting

What it supports. Seven cached steps yield 90.35% success at 1.84× speedup without refresh, or 91.85% at 1.56× with refresh every four chunks, against a 91.40% baseline. Eight steps without refresh instead fall to 75.85%, a 15.55-point loss. The headline's negligible-degradation claim applies only to selected configurations.

Where the evidence stops. These are benchmark averages without uncertainty estimates or a reported validation split for choosing cache settings. Section 4.2 links larger losses to longer tasks, but that grouping does not independently control difficulty, visual change or the number of inference calls.

6.2 Results and evaluation conditions

Task & protocolReported resultComparison & interpretation
LIBERO: four-suite action success and inference speed

All four suites, 2,000 evaluation episodes; Fast-WAM; 10 denoising steps, horizon 32; cache [0,7], τ=0.

97.10% success; 2.51× speedup.

Average task success (%) and total wall-clock inference speedup

Uncached Fast-WAM: 96.90%, 1.00×; difference +0.20 percentage points.

This is the strongest reported speedup, not evidence of a statistically established accuracy improvement; no uncertainty is reported. e-evaluatione-computee-timinge-liberoe-reporting

RoboTwin 2.0: fastest reported configuration

50 tasks; 20 episodes per task in each clean/randomized setting; cache [0,6], τ=0.

91.00% / 89.70% / 90.35%; 1.84× speedup.

Clean/randomized/average success (%) and total inference speedup

Baseline 92.20% / 90.60% / 91.40%; average −1.05 percentage points.

The approximately 1.8× headline permits a measurable average success loss. e-evaluatione-computee-timinge-robotwin

RoboTwin 2.0: refreshed seven-step cache

Same RoboTwin protocol; cache [0,6], τ=4.

91.85%; 1.56× speedup.

Average success (%) and total inference speedup

Baseline 91.40%; +0.45 percentage points.

A more conservative speed–success operating point; small improvements lack uncertainty estimates. e-evaluatione-robotwine-timinge-reporting

RoboTwin 2.0: failure of aggressive eight-step caching

Same RoboTwin protocol; cache [0,7], τ=0.

75.85%; 1.68× speedup.

Average success (%) and total inference speedup

Baseline 91.40%; −15.55 percentage points. Refreshing every four chunks gives 88.90% at 1.55×.

The LIBERO-winning setting transfers poorly. Claims of negligible degradation must be restricted to particular configurations. e-evaluatione-robotwine-libero

6.3 Ablations and diagnostic examples

Read component removals and qualitative examples within their stated evaluation conditions.

Figure 4. Recomputing the denoising tail is a central part of the successful configurations. Original paper, p. 8 ↗

Excerpt from the authors’ paper; cropped without altering the figure or table.

How to read it. The horizontal axis counts cached steps, not denoising iterations performed: the sampler still takes ten steps. Moving from seven to ten replaces progressively more fresh transformer evaluations with residual reuse on cached chunks. Red circles mean no refresh after initialization; orange squares and blue triangles refresh every four and eight chunks. Dashed lines mark uncached Fast-WAM success. Compare the two panels independently. RoboTwin degrades markedly when nine steps are cached, even with refresh. LIBERO is more tolerant at nine but deteriorates sharply at ten across all shown schedules. Section 4.3 states these trends; exact nine- and ten-step results are not tabulated. e-ablatione-similaritye-computee-reporting

What it supports. The useful regime preserves freshly computed tail steps. More frequent chunk refresh helps in parts of the sweep, but does not make caching all ten steps reliable. Together with Figure 2, this supports sensitivity to denoising position as well as to how long a residual has been reused.

Where the evidence stops. The graph does not isolate why tail steps matter: it also reduces the number of fresh evaluations. It provides no speedup curve for these ablations and no error bars. Exact values should not be inferred from marker positions as if they were tabulated measurements.

7. Analysis & limitations

7.1 What the evidence leaves open

Source description

Figure 4 shows severe degradation when caching extends into the final steps. RoboTwin falls at nine and ten cached steps; LIBERO collapses at ten across refresh settings. Exact nine-/ten-step values are not tabulated. Manual refresh tuning remains necessary. e-ablatione-limitations

Reader analysis

Observation conditioning is under-specified: Figure 3 calls h_0 'new obs', while Section 3.1 defines it as noisy-action embedding and places video conditioning inside skipped DiT attention. The source does not establish a separate fresh-observation path into h_0. e-conditioning-gap

Reader analysis

The implementation paragraph says late-step mean and standard deviation both fall; Figure 2 and Section 3.2 instead show standard deviation increasing. Equation (1) contains stray punctuation; Eq. (8)'s modulo rule requires a separate τ=0 branch. e-consistencye-flowe-cache

Reader analysis

Evidence covers one backbone and two manipulation benchmarks. No physical deployment, latency distribution, absolute latency, memory overhead, statistical intervals or controlled comparison of residual versus velocity caching is provided. Task-length associations do not isolate cache staleness causally. e-reportinge-limitations

7.2 Questions for discussion

  1. Does residual reuse outperform velocity reuse once cache schedule and noise are controlled?
  2. Where does fresh visual information enter a cached step, given the h_0/attention ambiguity?
  3. Can adaptive refreshing retain gains on long tasks without requiring task-specific tuning?

8. Reproducibility audit

8.1 Requirements and known gaps

Source description

Match Fast-WAM's Wan2.2-5B backbone, 1B action expert, hidden dimension 1024, ten steps and horizon 32. Reported hardware is one NVIDIA A100 80GB, a 64-core AMD CPU and 500 GB memory, using PyTorch. e-backbonee-compute

Reader analysis

A faithful reproduction still needs checkpoint identifiers, software versions, seed lists, executed prefix length and the cache-selection validation protocol. Time the entire inference pipeline and report inference-call counts, since aggregate episode timing may depend on trajectory length. e-flowe-reportinge-timing

Reader analysis

Proposed checks: compare residual, final-output and velocity reuse under identical schedules; then test refresh intervals and tail-step recomputation with paired seeds and controlled observation changes. Measure action error, task success and full-pipeline timing. e-cachee-conditioning-gape-ablatione-limitations

8.2 Proposed reproduction checks

The following checks are proposals motivated by the paper. They have not been run as part of this reading.

Check 1: Does residual reuse preserve useful responsiveness?

Reader-proposed, not run: begin with one verified Fast-WAM checkpoint and a held-out set of matched observation/action-noise states. Compare full computation, residual reuse, final-output reuse and velocity reuse with identical ten-step schedules, cache ages and cached-step masks. First measure predicted-velocity and final-action differences from full computation on the same states; then compare closed-loop task success with paired seeds. Repeat selected states with a controlled observation change while holding the action iterate fixed, and check whether h_0 and cached-step output respond. The residual-choice rationale predicts better retained behavior than frozen outputs under some changes. Equal or worse behavior, or no identifiable visual-input pathway, would weaken that rationale without invalidating the reported speed measurements. e-flowe-residuale-cachee-conditioning-gape-similaritye-compute

Check 2: Separate stale caches from the value of fresh tail steps

Reader-proposed, not run: on RoboTwin, repeat a factorial sweep of cache ranges [0,6], [0,7] and [0,8] with refresh settings 0, 4 and 8, plus an uncached baseline, using the same task/seed list. Stratify results by baseline episode length and record cache age, success, action error, inference-call count and full-pipeline time. Add a matched-compute control that moves an equal number of fresh evaluations from the tail to earlier denoising positions. If late variability is the key mechanism, preserving tail computation should outperform that control. If staleness drives long-task failures, refreshing should help most at greater cache ages within comparable tasks. Failure of either pattern would narrow the corresponding causal explanation. Report paired uncertainty rather than interpreting small mean differences alone. e-ablatione-limitationse-similaritye-evaluatione-computee-timinge-reporting

8.3 Reading coverage

Visual audit: Visually inspected the title/author/version page, all method and equation pages, Figures 1–4, Tables 1–2, experimental and hardware descriptions, limitations, and reference pages. All five final original crops were viewed, including the corrected table crops. Figure 3 arrows were cross-checked against Eqs. (6)–(8); the h_0 observation-conditioning ambiguity, special τ=0 handling and schematic-versus-experimental step selection are disclosed. Figure 2 was checked against Sections 3.2 and 4.1; their standard-deviation inconsistency is preserved. All pages supporting report, tutorial and proposed-check details are included. No appendix is present; separate supplemental availability remains unverified. Code, checkpoints and cited works were outside this reading, and no reproduction was run.

PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11. Appendix coverage: not present.

Original figures and tables remain the work of the source’s authors. Extractions preserve their scientific content; any HTML wrapper layout is disclosed with each figure. The surrounding reading notes are our own.

Text reading scope & known omissions
  • Abstract (p. 1)
  • 1 Introduction (pp. 1–2)
  • 2 Related Work: 2.1–2.3 (p. 3)
  • 3 Method: 3.1 Preliminaries, 3.2 Key Observation, 3.3 C³ache (pp. 3–6)
  • 4 Experiment: 4.1 Settings, 4.2 Main Results, 4.3 Ablation (pp. 6–8)
  • 5 Limitation and 6 Conclusion (p. 8)
  • References (pp. 9–11)

Outside the original text pass

  • Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
  • Identity notes: The inspected artifact is arXiv:2606.08962v1, dated 8 June 2026. Its title is rendered C³ache, represented by the supplied observedTitle as C^3ache: Accelerating World Action Models with Cross Inference Chunk Cache. The catalog title starts C3^3ache, a formatting discrepancy; all four authors agree. The catalog submittedDate is 7 June 2026, one day before the PDF's displayed date. No other revision was supplied or compared.
  • Supplied acquisition omission: Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This reading additionally inspected every PDF page and all final crops.
  • Separate supplemental material availability has not been fully verified.
  • No appendix appears in the supplied 11-page PDF. Code, checkpoints and cited works were not inspected; no experiments were reproduced.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e-identityPDF p. 1, title, author/affiliation block and arXiv marginInspect

The title is C³ache: Accelerating World Action Models with Cross Inference Chunk Cache. Authors, in order, are Weisen Zhao, Lam Nguyen, Zhicong Lu and Yuzhang Shang. Zhao and Lu are affiliated with George Mason University; Nguyen and Shang with University of Central Florida. The margin identifies arXiv:2606.08962v1 [cs.LG], 8 Jun 2026.

Go to primary source ↓
e-problemPDF pp. 1–2, Abstract, Section 1 and Figure 1Inspect

The proposed training-free acceleration targets residual redundancy across inference chunks at matching denoising steps. Model weights and training are unchanged. Figure 1 illustrates full-stack residual storage and later reuse; the introduction claims compatibility with within-chunk caching.

Go to primary source ↓
e-backbonePDF p. 3, Section 2.1; p. 7, Section 4.1, Model and DatasetsInspect

Fast-WAM retains video co-training but skips future-frame synthesis at action inference. A video backbone encodes the observation once and a separate flow-matching action expert generates the action chunk. The experiments use a Wan2.2-5B backbone with a 1B action expert, hidden dimension 1024, and 6B total parameters.

Go to primary source ↓
e-flowPDF p. 4, Section 3.1, multi-chunk generation and Eqs. (1)–(2)Inspect

Each control cycle generates H actions, executes part of the chunk, and receives a fresh observation. The flow model starts from Gaussian noise and predicts the clean-action-minus-noise velocity along a straight interpolation path; Eq. (2) gives the discrete integration update. Equation (1) visibly contains an extraneous comma between t and the clean action term.

Go to primary source ↓
e-residualPDF pp. 4–5, Section 3.1, DiT description and Eqs. (3)–(4)Inspect

Noisy actions are embedded as h_0; denoising time enters adaptive normalization. L DiT blocks use joint attention to action and fixed within-chunk video tokens. Their aggregate residual is R=h_L−h_0, and a linear head maps h_L=h_0+R to the flow velocity.

Go to primary source ↓
e-similarityPDF p. 5, Figure 2, Section 3.2 and Eq. (5)Inspect

The analysis uses all four LIBERO suites, each with 10 tasks and 500 episodes: 2,000 episodes and 29,924 reported inference chunks, with 10 denoising steps. At matching steps it measures adjacent-chunk cosine similarity of h_L, R and velocity. Residual mean similarity falls toward the tail while its standard deviation rises; velocity has the highest mean similarity in the plot.

Go to primary source ↓
e-cachePDF pp. 5–6, Section 3.3, Eqs. (6)–(8) and Figure 3Inspect

The method rejects directly reusing h_L or velocity because doing so freezes those tensors. Full computation stores h_L−h_0 for each step; cached computation adds the stored residual to the current h_0. For positive refresh interval τ, the first chunk and each multiple of τ refresh the cache. The prose separately defines τ=0 as never refreshing after initialization; the modulo expression in Eq. (8) cannot itself be evaluated at zero.

Go to primary source ↓
e-conditioning-gapPDF p. 4, Section 3.1, action-token and joint-attention description; p. 6, Figure 3/caption and Eq. (7) discussionInspect

Section 3.1 defines h_0 as embedded noisy action tokens and places video conditioning in DiT joint attention. Figure 3 labels h_0 as new obs, and its caption and Eq. (7) discussion assert that h_0 carries the new observation despite bypassing all DiT blocks. The supplied formulation does not specify an additional observation-to-h_0 pathway.

Go to primary source ↓
e-evaluationPDF p. 5, Section 3.2; pp. 6–7, Section 4.1, Model and DatasetsInspect

LIBERO evaluation covers four suites and 2,000 episodes. RoboTwin 2.0 evaluation covers 50 bimanual tasks with 20 differently seeded episodes per task per clean/randomized setting. Clean varies object placement; randomized additionally varies background, lighting, table height, camera distance and clutter.

Go to primary source ↓
e-computePDF p. 7, Section 4.1, model, Hardware and Implementation Detail paragraphsInspect

Both benchmarks use 10 denoising steps and action horizon 32. Experiments use PyTorch on one NVIDIA A100 80GB GPU, a 64-core AMD CPU and 500 GB memory. Main cache ranges are [0,6] and [0,7], with refresh settings 0, 4 and 8; ablations extend to [0,8] and [0,9].

Go to primary source ↓
e-timingPDF p. 7, Section 4.1, MetricsInspect

The speed metric is total wall-clock inference speedup across all episodes, including VAE encoding, video prefill for KV-cache construction, action token embedding, output projection and scheduler steps, not merely the skipped DiT stack. Success rate is the other reported metric.

Go to primary source ↓
e-liberoPDF p. 7, Table 1, Fast-WAM and all [0,6]/[0,7] rowsInspect

Fast-WAM averages 96.90% success. For [0,7], refresh 0 yields 97.10% and 2.51× speedup: Spatial 97.60, Object 99.60, Goal 96.00, Long 95.20. Refresh 4 gives 96.70%/1.82×; refresh 8 gives 97.00%/2.13×. For [0,6], refresh 0/4/8 gives 97.45/96.55/96.75% and 2.10/1.69/1.86×, respectively.

Go to primary source ↓
e-robotwinPDF p. 8, Table 2, Fast-WAM and all [0,6]/[0,7] rowsInspect

Baseline clean/random/average success is 92.20/90.60/91.40%. For [0,6], refresh 0 gives 91.00/89.70/90.35% at 1.84×; refresh 4 gives 92.60/91.10/91.85% at 1.56×; refresh 8 averages 90.40% at 1.71×. For [0,7], refresh 0/4/8 averages 75.85/88.90/85.25% at 1.68/1.55/1.72×, respectively.

Go to primary source ↓
e-ablationPDF p. 8, Figure 4 and Section 4.3; p. 7, Implementation DetailInspect

Figure 4 varies the number of cached steps from 7 through 10 at refresh settings 0, 4 and 8. Extending caching to nine or ten steps substantially degrades RoboTwin. LIBERO degrades at nine with refresh 0 and at ten for all shown settings. The graph supplies curves, not a numerical table for the nine- and ten-step cases.

Go to primary source ↓
e-limitationsPDF p. 7, Section 4.2, task-length analysis; p. 8, Section 5Inspect

The authors associate greater loss under [0,7], refresh 0 on RoboTwin with longer episodes, using task-length groups rather than a controlled intervention. They identify manual tuning of refresh interval as a limitation and propose adapting it to task difficulty and scene change in future work.

Go to primary source ↓
e-consistencyPDF p. 5, Figure 2 and Section 3.2; p. 7, Section 4.1, Implementation DetailInspect

Figure 2 and Section 3.2 show increasing late-step standard deviation of cross-chunk cosine similarity. The implementation paragraph instead says both mean and standard deviation drop past step 6; that statement conflicts with the inspected plot and earlier description.

Go to primary source ↓
e-reportingPDF pp. 6–8, Sections 4–5, Tables 1–2 and Figure 4Inspect

The supplied experiments compare cache settings with uncached Fast-WAM on LIBERO and RoboTwin 2.0. They report success percentages and aggregate inference speedup, without confidence intervals, absolute latency or cache-memory measurements. The experimental description does not specify checkpoint identifiers, training recipe, software versions, exact seed list, executed chunk prefix length, or a validation protocol for selecting cache settings.

Go to primary source ↓

8.5 Primary sources

Scroll across the image to inspect details. Press Esc to close.