ZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training

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

Authors: Joy Future Academy

Source: 2609.00188 ↗ · Project page ↗ · Catalog record

Reading: 4 / 558 · 5 original figures & tables · ~21 min ·

1. Paper overview

In one sentence: ZimaBlue grounds video priors in robot actions, then reuses a slow world model’s cached features for fast corrective control, trading fresher feedback against stale guidance and distillation loss. e03 e04 e05 e06 e07 e10 e12 e17

At a glance	What to know
Research problem	Author claim Robot demonstrations are costly and limited in diversity, while abundant embodied videos lack motor labels. The paper asks how video-derived dynamics can transfer to executable manipulation, and how an iterative generative world model can provide that knowledge without delaying corrective control. e02
Core mechanism	Source description A three-stage curriculum separates learning visual dynamics, grounding those dynamics in heterogeneous robot actions, and specializing control to a deployment domain. e05 e06 e07
A key reported result	Deployment acceleration: Slow: 449.6 ms, 63.9%; undistilled Dual-System: 145.6 ms, 77.8%; distilled and compiled Dual-System: 33.0 ms, 75.0%. Reported end-to-end inference latency (ms) and overall task-macro success (%). Real-robot evaluation; RTX 4090 deployment. The accelerated variant combines two-step distillation and compilation. The paper reports 13.6× overall acceleration and a 2.8-percentage-point success reduction relative to undistilled Dual-System. The approximately 30 Hz configuration has 75.0% success. Pairing 33.0 ms with the separate 77.8% result would conflate configurations. e12 e13
Reading caution	Reader analysis Physical evaluation covers one Franka platform and a small task suite. Ten rollouts per task provide limited precision, and Toys uses placement-level rather than whole-task success. The appendix establishes exclusion from DROID post-training but supplies no corpus-wide overlap audit for the large video mixtures. e05 e10 e11

Core contributions

Source description
A three-stage curriculum separates learning visual dynamics, grounding those dynamics in heterogeneous robot actions, and specializing control to a deployment domain. e05 e06 e07
Source description
A semantic state–action interface and asynchronous Slow–Fast architecture connect a shared world model to different embodiments and to frequently refreshed control observations. e03 e04 e08
Author claim
The authors present real-robot scaling experiments and simulation benchmarks as evidence that large-scale embodied video improves generalization, especially under visual shifts and on unseen task compositions. e10 e14 e15 e16

A blue 30-layer Slow video transformer sends video key/value features from its first 12 layers to a green 12-layer Fast action transformer; each branch has separate observation and state inputs. — **Figure 2.** Cached visual dynamics meet an updated observation in the Fast action tower. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

2. Motivation

2.1 The problem and the proposed response

Author claim

Robot demonstrations are costly and limited in diversity, while abundant embodied videos lack motor labels. The paper asks how video-derived dynamics can transfer to executable manipulation, and how an iterative generative world model can provide that knowledge without delaying corrective control. e02

2.2 What this reading follows

A predictive robot model faces two distinct problems: learning useful dynamics from abundant video and producing corrections before the scene changes again. ZimaBlue addresses the first through video pretraining, action-grounded mid-training, and target-domain adaptation. It addresses the second with a 5B Slow transformer and a 0.5B Fast controller that reads both cached Slow features and fresh observations. These illustrations follow that transfer of information, then separate the evidence for data scaling, asynchronous control, and acceleration. The real-robot results support the combined recipe, while task-level regressions and simulation diagnostics show why a higher aggregate score should not be read as uniformly better control. e03 e04 e05 e06 e07 e10 e12 e17

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 dimension	Recorded classification
Major category	WAMs
Architecture	Dual-system
Prediction paradigm	IDM
Quadrant	Q4 · Dual-system × IDM

3.1 Evidence-based assessment

Conflicts with the recorded classification

Reader analysis

WAM and Dual-system are directly supported: separate Slow and Fast transformers exchange cached video features and run asynchronously. The recorded IDM designation is not established by the described mechanism. Slow jointly predicts video and actions with mutual attention; its action loss is auxiliary for deployed dual-system control. Fast is an action flow model conditioned on fresh observations and predictive internal features, without an explicit generated-frame transition-to-action inverse-dynamics stage. Its behavior may be inverse-dynamics-like in a broad sense, but that does not verify IDM or Q4 as recorded. Preserve the snapshot for editorial review. e04 e06 e07 e08 e20

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

Inputs	Outputs
RGB observations packed into a three-view canvas, with missing views masked [e05, e06] Proprioceptive state and language instruction [e03, e04] Training targets comprising future video latents and, in action-supervised stages, synchronized action chunks [e05, e06] Fresh observations, a committed action prefix, and cached Slow video features for Fast inference [e07, e08]	Slow: future video latents, auxiliary action predictions, and layer-wise video key/value features [e04, e06, e09] Fast: an action chunk decoded into physical robot commands; this branch supplies deployed real-robot actions [e04, e08]

4.2 Equations and their role

\Delta p_h = R_0^{-1}(p_h-p_0),\qquad \Delta R_h = R_0^{-1}R_h,\qquad \Delta q_h = q_h-q_0.

The chunk anchor has position p_0, rotation R_0, and joint configuration q_0. Future targets are expressed relative to it, reducing dependence on absolute scene coordinates. ΔR_h is represented by its first two columns. e03

\begin{aligned} z_t &= (1-\sigma_t^{\mathrm{vid}})z+\sigma_t^{\mathrm{vid}}\epsilon, & v^{\mathrm{vid}} &= \epsilon-z,\\ \mathcal L_{\mathrm{sup}}^S &= \lambda_{\mathrm{vid}}\mathcal L_{\mathrm{vid}}+\lambda_{\mathrm{act}}\mathcal L_{\mathrm{act}}. \end{aligned}

Flow matching learns the velocity between clean data and Gaussian noise. L_vid and L_act are squared velocity errors masked to valid visual tokens and action coordinates. Stage I uses only video loss; Stage II adds analogous action corruption and action loss, with mutually conditioning noisy video and action streams. e05 e06

\begin{aligned} \tilde a &= \operatorname{Pad}(a_{\delta:H}),\\ a^F_{t,i} &= \begin{cases}\tilde a_i,&i<p,\\(1-\sigma_t^{\mathrm{act}})\tilde a_i+\sigma_t^{\mathrm{act}}\eta_i,&i\ge p,\end{cases}\\ \mathcal L_{\mathrm{sup}}^F &= \mathbb E\!\left[\left\|m_{\mathrm{act},F}\odot\left(\hat v_{\theta}^{\mathrm{act},F}(a_t^F,c_\delta,s_\delta,K_{\mathrm{slow}},\ell,\delta)-(\eta-\tilde a)\right)\right\|_2^2\right].\end{aligned}

δ selects an updated observation within the original horizon, and p is the committed prefix length. Only the suffix is noised. The combined mask removes invalid coordinates, padded positions, and prefix supervision, teaching Fast to revise remaining actions with fresh feedback and frozen Slow guidance. e07

g_{\mathrm{DMD}}^b=\frac{\hat x_{0,f}^b-\hat x_{0,r}^b}{\operatorname{mean}_{\mathrm{valid}}\!\left(\left|x_g^b-\hat x_{0,r}^b\right|\right)+\epsilon_{\mathrm{num}}},\qquad b\in\{S,F\}.

The normalized difference between fake-score and real-score clean-sample estimates supplies a stopped-gradient distribution-matching direction. Gradients pass through the student sample. Slow samples contain video and actions; Fast samples contain actions only. e09

5. Method in detail

5.1 First give video dynamics a motor vocabulary

Reader analysis

Action-free video can teach visual transitions without specifying how a particular robot realizes them. ZimaBlue therefore trains only Slow during video pretraining, then adds synchronized robot actions during mid-training. The shared 100-dimensional interface assigns semantic slots to end effectors, grippers, joints, torso, and base channels; masks remove undefined coordinates from supervision. Chunk-relative geometry further reduces dependence on a robot’s global coordinate frame: a target translation becomes Δp = R₀⁻¹(pₕ − p₀). Video and action flow losses then align predicted transitions with motor consequences. Only afterward does target-domain training introduce Fast with Slow frozen. My interpretation is that the curriculum separates representational coverage from deployment specialization: video supplies breadth, robot trajectories supply grounding, and the target controller learns how to use that representation under its own observation and action interface. e03 e05 e06 e07

Blue Slow and purple Fast timelines show repeated Fast requests reusing a Slow video cache, receiving new observations, and preserving committed action prefixes through RTC. — **Figure 4.** A cache can remain useful after its originating observation is old. Original paper, p. 10 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

5.2 Train Fast for the information it will actually receive

Source description

The cache consumed by Fast was computed before the newest observation, so ordinary synchronized training would present an easier problem than asynchronous execution. ZimaBlue samples a temporal offset within the action horizon, supplies the corresponding updated observation and state, and shifts the remaining action targets. It also keeps a short prefix clean to represent actions already committed by the previous request. Padding, inactive coordinates, and that prefix are excluded from the action loss. At runtime, Fast conditions on the fresh observation, the available Slow cache, and the committed prefix, replacing only the unexecuted suffix. The important causal distinction is that feedback can change the next action without waiting for another full world-model prediction. The cache provides guidance, while local sensing anchors the command to the robot’s current situation. e07 e08

5.3 Read acceleration as a success–latency tradeoff

Reader analysis

After adaptation, distribution matching distillation compresses each branch from eight to two transformer evaluations. Slow is distilled first; its frozen features then condition Fast distillation. Each stage uses a deployable student, a fixed teacher-derived score model, and a learned score model tracking the student distribution, together with supervised anchors. Those training-time score networks are not extra runtime modules. Table 4 measures the resulting system alongside the eight-step variants, with compilation included in the fastest configuration. My reading is that the useful outcome is the joint operating point: 33.0 ms and 75.0% success, compared with 145.6 ms and 77.8% for the undistilled dual system. The paper establishes that tradeoff on its hardware and task suite; neither the individual task rows nor the simulation transfer diagnostic supports a blanket claim that acceleration or adaptation improves every behavior. e09 e12 e13 e17

5.4 Training and inference

During training

Source description

Stage I adapts Wan2.2-TI2V-5B using video-only flow matching. A broad mixture of human, simulated, and robot videos is followed by a curated manipulation-focused mixture that includes proprietary human video. The default sample has 33 frames across four temporal chunks. Blockwise noise and teacher forcing permit parallel training with causal context; state/action placeholders carry no supervision. e05

Source description

Stage II trains only Slow on synchronized video, state, instruction, and 24-step action chunks from DROID, AgiBot, Galaxea, and RoboMIND2-Franka. Video and action noise levels are coupled, and their masked flow losses train a shared backbone. Action prediction grounds the visual representation but is not the deployed Slow branch's control output. e04 e06

Source description

Stage III first specializes Slow, then freezes it while training Fast. Fast weights are transferred from Slow's first 12 layers and interpolated to a narrower width. Random observation offsets expose Fast to partially executed trajectories; clean, loss-masked action prefixes teach continuity with previously committed commands. e07

Source description

Sequential distribution matching distillation reduces Slow and then Fast from eight to two DiT evaluations each. Each stage uses a student, frozen teacher-derived real-score network, and online fake-score network, with five fake-score updates per generator update. Supervised flow losses remain as anchors. Fast distillation uses frozen distilled Slow features matched by denoising-step index. e09

During inference

Source description

Slow and Fast run asynchronously. Fast reuses the available Slow cache while receiving newer observations and states. Real-Time Chunking preserves committed prefix actions and replaces only the unexecuted suffix. Decoding restores physical units and composes relative targets with robot state before execution. e03 e08

Source description

Distilled deployment retains only the two student generators; score networks are discarded. Slow still predicts latent video and auxiliary actions, whereas Fast supplies robot commands. The simulation benchmarks instead use Slow alone, so their scores do not establish the performance of the accelerated dual-system controller. e09 e13

5.5 Implementation flow

Standardize observations and robot coordinates
State and action share 100 semantic slots covering end-effectors, grippers, joints, torso, base, and hands. Undefined coordinates are zero-filled and loss-masked. End-effector targets are relative to the chunk's initial pose; translations use robust-percentile normalization, while continuous 6D rotations remain unscaled. Camera padding is excluded from attention and video supervision. e03 e05
Predict dynamics and expose internal context
The 5B Slow DiT jointly denoises future video and action tokens after action grounding. Its block-causal mask allows mutual interaction within the current noisy video–action block while blocking access to clean future outcomes. It exports video key/value features that encode predictive context. e04 e06 e20
Generate actions from predictive context and fresh feedback
The 0.5B Fast DiT receives updated observation and state tokens, language, and a noisy action chunk. Action queries also attend to cached features from the first 12 Slow layers. Fast has no future-video query stream; it generates executable actions without waiting for a newly completed Slow rollout or decoded future RGB sequence. e04 e08 e20

6. Experiments & results

ZimaBlue converts action-free embodied video into robot control through video pre-training, cross-embodiment video–action alignment, and target-domain adaptation. Its deployed controller combines a predictive Slow transformer with a smaller Fast action generator that consumes cached video features and fresh observations. The central evidence is executed Franka manipulation: expanding training raises reported task-macro success from 36.1% to 77.8%. A separately accelerated configuration reaches 33.0 ms inference latency with 75.0% success. The work supports video scaling and asynchronous feedback as useful ingredients, while leaving the specific necessity of predictive cached features insufficiently isolated.

6.1 Read the original evidence

Zero-shot Franka success table compares two external policies and four cumulative ZimaBlue training configurations across eight standard tasks, four perturbed tasks, and their combined task average. — **Table 3.** Additional egocentric video improves both suites, with a larger final gain under visual perturbations. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

6.2 Results and evaluation conditions

Task & protocol	Reported result	Comparison & interpretation
Zero-shot physical manipulation under increasing pre-training DROID-adapted Franka; eight Standard and four Perturbed tasks; ten rollouts per task, with three separately scored placements per Toys rollout. All ZimaBlue variants share target-domain adaptation and eight-step inference.	Overall: 36.1 → 46.1 → 66.9 → 77.8 for the Wan-initialized baseline, +6K hours multi-embodiment data, +60K hours video, and +120K hours video. Full model: Standard 87.9; Perturbed 57.5. Task-macro success rate (%)	Official DROID-adapted π0.5: 54.4 overall; DreamZero: 53.6. These are reported executed robot outcomes. The cumulative comparison supports the training recipe, with matched downstream schedules, but does not equalize total pre-training compute or external baselines' upstream data. e10 e11
Contribution of asynchronous Slow–Fast control Same real-robot task suites; eight-step Slow versus eight-step dual-system deployment.	Slow → Dual-System: Standard 80.8 → 87.9; Perturbed 30.0 → 57.5; Overall 63.9 → 77.8. Task-macro success rate (%)	Reported gains are 7.1 percentage points on Standard and 27.5 on Perturbed. The larger perturbation gain is consistent with useful fresh feedback. Because the intervention changes architecture and control latency together, it does not isolate the benefit of predictive K/V content. e12
Deployment acceleration Real-robot evaluation; RTX 4090 deployment. The accelerated variant combines two-step distillation and compilation.	Slow: 449.6 ms, 63.9%; undistilled Dual-System: 145.6 ms, 77.8%; distilled and compiled Dual-System: 33.0 ms, 75.0%. Reported end-to-end inference latency (ms) and overall task-macro success (%)	The paper reports 13.6× overall acceleration and a 2.8-percentage-point success reduction relative to undistilled Dual-System. The approximately 30 Hz configuration has 75.0% success. Pairing 33.0 ms with the separate 77.8% result would conflate configurations. e12 e13
LIBERO-Plus robustness Slow-only simulation evaluation on 10,030 episodes. Zero-shot transfer trains on standard LIBERO; supervised fine-tuning additionally uses LIBERO-Plus training data.	Zero-shot transfer: 86.7; supervised fine-tuning: 92.0. Unweighted mean success across seven perturbation categories (%)	Within the corresponding protocols: InternVLA-A1.5 scores 85.8 and CAC-VLA scores 90.1. Episode-weighted scores are separately 86.0 and 91.5. Camera success rises from 58.1 to 95.4, but robot-initial-state success falls from 88.9 to 81.1; aggregate improvement does not imply improvement in every category. e13 e14 e17
RoboTwin 2.0 bimanual manipulation One Slow policy post-trained on clean and randomized demonstrations across all 50 tasks; 100 trials per task under each condition.	Clean 94.7; Randomized 94.3; their mean 94.5. Mean task success (%)	ABot-M0.5: Clean 94.0; Randomized 94.2; mean 94.1. This demonstrates simulated robustness on trained tasks under the benchmark's randomization. The average improvement is small, and the table supplies no uncertainty for assessing its significance. e13 e15
RoboCasa365 household manipulation and unseen compositions Slow-only policy trained on Human300; 18 Atomic-Seen, 16 Composite-Seen, and 16 Composite-Unseen evaluation tasks, each with 50 rollouts.	Atomic-Seen 78.1; Composite-Seen 50.4; Composite-Unseen 16.5; Overall 49.5. Split success and overall success (%)	ABot-M0.6: 46.6 overall and 7.9 Composite-Unseen. Xiaomi-Robotics-1: 57.4 overall and 32.1 Composite-Unseen. Overall weights the splits by their task counts, rather than averaging three split percentages equally. Unseen-composition performance exceeds the compared WAMs, but remains low in absolute terms and below the strongest compared VLA. e13 e16 e19

6.3 Ablations and diagnostic examples

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

Per-task success counts and suite means for Slow, Dual-System, and DMD-Distilled control, ending with overall success of 63.9, 77.8, and 75.0 percent and latency of 449.6, 145.6, and 33.0 milliseconds. — **Table 4.** Faster control improves the average, but the per-task rows expose its tradeoffs. Original paper, p. 16 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

Four LIBERO task-suite rows report fine-tuning minus zero-shot success in percentage points across seven perturbations; camera changes are all positive and robot-state changes are all negative. — **Table 12.** Fine-tuning changes the shape of robustness, not just its average. Original paper, p. 28 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

7. Analysis & limitations

7.1 What the evidence leaves open

Reader analysis

Physical evaluation covers one Franka platform and a small task suite. Ten rollouts per task provide limited precision, and Toys uses placement-level rather than whole-task success. The appendix establishes exclusion from DROID post-training but supplies no corpus-wide overlap audit for the large video mixtures. e05 e10 e11

Reader analysis

The scaling experiments control DROID adaptation but add both experience and pre-training work. The baseline already inherits Wan video pre-training despite the appendix's 'From scratch' label. Cross-model RoboCasa comparisons vary architectures and datasets, so they cannot independently establish a causal video-volume scaling law. e05 e10 e16 e18

Reader analysis

No reported ablation separates future-predictive K/V guidance from ordinary visual features, cache reuse, or increased feedback frequency. Causal attention establishes permissible temporal information flow; the experiments do not directly measure interventional dynamics accuracy. e06 e12 e20

Author claim

The authors identify failures to advance from correct intermediate states, lost contact, target displacement, and interaction with the tablecloth. They call for progress-aware replanning and broader evaluation. These remaining errors qualify the claim of generalizable physical knowledge. e21

Reader analysis

Some presentation details require reconciliation: Table 4 calls a Standard task 'Basket', whereas Appendix A defines cup selection. Table 7's vague Average caption is clarified by the task-weighted protocol and Table 14. The supplied Figure 8 label removes both pre-training and mid-training, limiting attribution of that qualitative example specifically to video pre-training. e11 e12 e16 e19 e22

7.2 Questions for discussion

Does predictive Slow guidance outperform observation-only guidance when Fast capacity and feedback frequency are matched? [e04, e12]
How much of the video-scaling gain survives fixed-compute training and an explicit overlap audit across video and evaluation tasks? [e05, e10]
What cache-age threshold causes Fast to lose the benefits of its predictive context under abrupt scene changes? [e07, e08]
Why does LIBERO-Plus fine-tuning improve camera robustness while reducing robot-initial-state robustness? [e14, e17]

8. Reproducibility audit

8.1 Requirements and known gaps

Source description

The most concrete training specification is DROID adaptation: Slow runs for 60K steps with global batch 256, learning rate 1×10^-4, and 2% warmup; frozen-Slow Fast training runs for 160K steps with batch 64, the same learning rate, and 1% warmup. Both use action horizon 24. e11

Reader analysis

Rebuilding the full recipe requires the Wan initialization, public and proprietary video mixtures, embodiment mappings and normalization statistics, causal/view masks, target-domain data, and sequential distillation. The supplied text does not fully specify Stage I/II mixture weights, optimization budgets, total training compute, or exact normalization percentiles, preventing an exact reconstruction. e03 e05 e06 e09

Reader analysis

A focused mechanism test would train matched Fast branches using predictive Slow caches, observation-only caches, and no Slow cache, holding action horizon, target data, and feedback rate fixed. Perturbed-suite success, latency, and cache age would test whether predicted dynamics add value beyond frequent visual correction. e07 e08 e12

Reader analysis

The paper lists project and code links, but no repository contents were supplied or inspected and no experiments were reproduced. The release's completeness, proprietary-data accessibility, and reproducibility of the reported latency remain unverified. e01 e05 e13

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 predictive cache content matter after controlling feedback rate?

Proposed check: hold the Fast architecture, DROID adaptation, fresh observation cadence, RTC prefix, and inference budget fixed. Compare the normal Slow cache with a same-shaped cache from current observations only, a cache shuffled across matched task episodes, and no Slow guidance. Sweep cache age under the same controlled perturbations. Measure task success, recovery delay, and action discontinuities. A consistent advantage for the predictive cache at matched timing would support its dynamics contribution; equal performance would suggest that faster local feedback explains much of the reported gain. e04 e07 e08 e12

Check 2: Does adaptation preserve robustness when perturbations are balanced?

Proposed check: reproduce the zero-shot and SFT comparison on the identical LIBERO-Plus episodes, retaining the official robot-state protocol. Add an SFT variant with balanced camera and robot-initial-state perturbations while matching training steps and total demonstrations. Report each category, each suite, and both category-macro and episode-weighted averages. The falsifiable prediction is that balancing reduces the negative Robot column without erasing the camera gains. Include paired confidence intervals to distinguish retention from task-sampling noise. e14 e17

8.3 Reading coverage

Visual audit: Original PDF pages 6, 10, 14, 16, and 28 and all five final crops were visually inspected. The pass covers Figure 2 architecture, Figure 4 asynchronous timeline, Table 3 training comparison, Table 4 latency/task ablation, and Table 12 robustness deltas; axes, legends, row labels, and denominators were checked against source text. The PDF SHA-256 was recomputed and matches the base record. Other paper figures, rollout videos, released code, and actual robot behavior were not visually or experimentally audited in this illustrated pass. Base evidence additions record this later visual inspection without implying a new full-paper reproduction.

PDF pages inspected for this edition: 6, 10, 14, 16, 28. Appendix coverage: reviewed.

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
1 Introduction
2 Related Work, including Sections 2.1–2.3
3 Model Architecture, including Sections 3.1–3.2
4 Training Pipeline, including Sections 4.1–4.4.2
5 Acceleration Schemes, including Sections 5.1–5.3
6 Experiments, including Sections 6.1–6.3.3
7 Conclusion and Future Work
8 Authors
References
A Real-Robot Evaluation Details, including Sections A.1–A.6
B Additional Simulation Experiment Details, including Sections B.1–B.3
C Attention Masks

Outside the original text pass

Reading covered the supplied text extraction. PDF figure images and graphical attention masks were not visually inspected; only their captions and surrounding explanations were available.
Separate supplemental material availability was not verified.
The supplied artifact hash was retained as provenance; the PDF binary was not independently rehashed.
Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
Separate supplemental material availability has not been fully verified.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e01PDF p. 1, title, arXiv version line, and Abstract links; p. 20, Section 8 AuthorsInspect

The supplied title and identifier match the catalog. The text identifies arXiv:2609.00188v1 dated 31 August 2026, and Section 8 lists the catalog's authors. Project and code URLs are printed on page 1.

1. Paper overview

Core contributions

2. Motivation

2.1 The problem and the proposed response

2.2 What this reading follows

3. Research context

3.1 Evidence-based assessment

4. Problem formulation

4.1 Inputs and outputs

4.2 Equations and their role

5. Method in detail

5.1 First give video dynamics a motor vocabulary

5.2 Train Fast for the information it will actually receive

5.3 Read acceleration as a success–latency tradeoff

5.4 Training and inference

During training

During inference

5.5 Implementation flow

6. Experiments & results

6.1 Read the original evidence

6.2 Results and evaluation conditions

6.3 Ablations and diagnostic examples

7. Analysis & limitations

7.1 What the evidence leaves open

7.2 Questions for discussion

8. Reproducibility audit

8.1 Requirements and known gaps

8.2 Proposed reproduction checks

Check 1: Does predictive cache content matter after controlling feedback rate?

Check 2: Does adaptation preserve robustness when perturbations are balanced?

8.3 Reading coverage

Outside the original text pass

8.4 Traceable evidence

8.5 Primary sources

ZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training ↗