Keep the Future, Drop the Rollout: RIFT for World Action Models
1. Paper overview
In one sentence: RIFT replaces iterative future-video generation with one learned cache prefill, retaining explicit future-conditioned actions at low deployment latency within the tested simulation settings. probleminterventioncache-methodliberolimits
| At a glance | What to know |
|---|---|
| Research problem | Source description Rollout-based WAMs generate video although their action expert consumes internal future-position keys and values. Removing both rollout and future attention confounds their contributions. The paper asks whether an effective future read can survive when its producer becomes a single backbone pass. problemcache-method |
| Core mechanism | Source description A record-and-replay protocol tests future-cache use through closed-loop simulator trajectories, with matched initial states and policy seeds. Structured final-clean and frozen-present controls complement destructive value edits. intervention |
| A key reported result | LIBERO manipulation success and latency: RIFT: 98.8±0.17%; 247.9 ms. Success mean±evaluation-seed std (%); ms/action chunk. 40 tasks; one checkpoint per method; three evaluation seeds, 2,000 trials each. One A800, original denoising configurations. Fast-WAM: 96.8±0.27%, 235.7 ms; PFD: 97.3±0.12%, 257.0 ms; Joint: 98.4±0.26%, 780.2 ms; IDM: 98.6±0.34%, 1081.2 ms; LingBot-VA: 98.5±0.08%, 2270.3 ms. Reported success remains near rollout methods, with 68.2–89.1% lower latency. Error bars measure evaluation noise, not training-run variation; LingBot-VA has different embodied pretraining. libero |
| Reading caution |
Core contributions
- Source description
A record-and-replay protocol tests future-cache use through closed-loop simulator trajectories, with matched initial states and policy seeds. Structured final-clean and frozen-present controls complement destructive value edits. intervention
- Source description
RIFT learns position-aligned anticipation tokens under action, native video and conditional future-flow supervision, preserving the original future-read interface at deployment. cache-methodtraining
Figure 2. Test the representation the action expert actually reads. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each color against its own Original row. Bar length is end-effector average displacement from that model’s unmodified simulator trajectory; the compact right block gives success, a separate quantity. Original drift is zero by definition. Spatial shuffle and temporal swap reassign values while retaining the recorded keys. Final clean K/V is a different control: the Figure 2 caption and Section 3.5 say both keys and values are fixed from the clean rollout. N/A is structural: IDM and LingBot-VA already consume a fixed clean cache, and Cosmos-2 has no second future timestep to swap. Section 3.2 on page 4 defines common-prefix scoring; Table 3 on page 10 supplies task-cluster intervals. interventioncache-resultssource-discrepancy
What it supports. For Joint, masking leaves only 9.7% success, while final-clean replay retains 97.9% against 98.4% Original success. Table 3 refines the replay drift to 1.908 cm. The contrast supports meaningful future conditioning and a reusable complete cache; it does not show that this cache can be constructed without a rollout.
Where the evidence stops. Page 2 instead describes final values under original keys. That conflicts with the explicit full-cache protocol on pages 3–4 and 10. This guide follows the latter; separate key/value contributions remain unresolved, and destructive edits may be out of distribution.
2. Motivation
2.1 The problem and the proposed response
Rollout-based WAMs generate video although their action expert consumes internal future-position keys and values. Removing both rollout and future attention confounds their contributions. The paper asks whether an effective future read can survive when its producer becomes a single backbone pass. problemcache-method
2.2 What this reading follows
The useful output of a world action model’s video branch may be the representation read by its action expert, rather than the rendered movie. RIFT turns that observation into two separate experiments. First, it edits existing models’ future caches to ask what action generation depends on. Second, it learns anticipation tokens that produce a complete cache in one pass, then evaluates the resulting policy. Read the intervention figure before the architecture: successful replay of a rollout-produced cache motivates the interface, but does not prove that a learned producer will work. The deployment tables supply that separate empirical test. probleminterventioncache-methodliberolimits
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
Insufficient evidence to decide
WAM and efficient-inference labels are supported. Distinct video/cache and action experts support a dual-system reading. However, RIFT inherits Joint’s architecture and trains through the cache; fixed-cache consumption alone does not establish a separately defined inverse-dynamics model. The recorded IDM/Q4 assignment and JEPA subcategory therefore remain unverified by the architectural evidence. cache-methodtrainingconfiguration
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 |
|---|---|
|
|
4.2 Equations and their role
5. Method in detail
5.1 1. Establish use before proposing a cheaper producer
The intervention study begins with an unusually convenient boundary: video tokens cannot attend to action tokens, so the video cache can be recorded independently of action denoising. At a matched state and random seed, an edited action rollout can therefore be compared with the original. The two readouts answer different questions. EE-ADE measures how the executed trajectory changes, while success asks whether the task still completes. Spatial shuffling and temporal swapping test whether values matter at their assigned positions, although these edits may be out of distribution. Final-clean replay asks a different question: can the complete cache remain fixed throughout denoising? Near-original execution supports that consumption pattern. Because the clean cache was still produced by video rollout, the experiment does not yet establish an efficient cache producer. interventioncache-results
Figure 3. One cache prefill supplies every action-denoising step. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow panel (a) from the current observation through future placeholders E and the single VideoStack pass. Its output is a stack of layer-specific keys and values, not a decoded future movie. The action expert repeatedly reads that fixed stack while retaining observation and instruction conditioning. The 'no action read' label agrees with Section 4.2 on page 4: video tokens cannot attend to actions. In panel (b), native video supervision is a separate forward through shared weights. Prefill states S_phi also feed a conditional-FM head and a diagnostic probe. The probe’s stop-gradient label matters: Section 4.3 and Appendix A on page 10 restrict that loss to probe parameters. cache-methodtrainingconfigurationlimits
What it supports. The deployment change is concentrated in cache production. A single learned prefill replaces iterative future-video generation, while the action expert still receives explicit future-position conditioning. The training branches can improve what the placeholders encode without becoming extra computations in the policy-only action path.
Where the evidence stops. 'One pass' refers to the video/cache producer, not the complete policy: the action expert still runs ten flow steps. Auxiliary FM sampling and diagnostic decoding are absent from reported policy-only latency, and the diagram does not establish physical-robot performance.
5.2 2. Replace the producer while preserving its interface
RIFT fills the producer gap with learned placeholders at the original future spatiotemporal positions. One forward through the video expert turns current observation tokens and these placeholders into per-layer keys and values. The fixed cache is then reused across the action expert’s ten flow steps; a new prefill is made for the next chunk. The information route is explicit even though future pixels are never rendered during policy-only deployment. Training keeps two uses of the shared video expert separate: native video supervision preserves dynamics training, while the observation-plus-placeholder forward matches deployment. Action loss can backpropagate through the attended cache into its producer. Thus keeping the cache interface does not mean freezing the video expert. The detached diagnostic probe is the branch whose loss cannot alter the deployed representation. cache-methodtrainingconfiguration
5.3 3. Distinguish future supervision from a sampled future policy
Direct L2 prediction targets the conditional mean and can average distinct valid futures. The paper instead conditions a training-only flow head on the anticipation states, using noised future patches and a velocity target to shape those states. Reader interpretation: this encourages a useful predictive representation, but does not imply that deployment samples a future movie or that the learned cache equals a rollout cache. The token sweep supports retaining future-interface capacity; the full-recipe comparison gives weaker evidence about the precise source of the FM gain because its curriculum also changes. Likewise, diagnostic decodes establish what can be read out, not what controls execution. The optional L2–FM monitor compares future estimators outside the controller, and its averaged failure curve should not be treated as an individual-episode detection guarantee. trainingablationdecodesmonitor
5.4 Training and inference
During training
Each optimization step has two forwards through the shared video expert: ordinary clean-first-frame video flow training, and a deployment-matched observation-plus-anticipation forward. Action loss updates the action expert and cache producer; conditional FM updates the producer and its auxiliary head. The RMS-normalized linear probe receives stopped-gradient states and cannot shape the deployed representation. training
After 70% of the curriculum, first-frame perturbation probability and noise scale rise to 0.3 and 0.06 times latent standard deviation. Perturbed rows retain future losses but are excluded before action-loss reduction. Both auxiliary weights stay at 1 initially, then cosine-decay to 0.2. Thus action supervision always uses clean inputs. training
During inference
Deployment retains one prefill and 10 action-flow steps, with guidance 1.0. Native video co-training, ground-truth futures and auxiliary heads are removed. Neither video denoising nor VAE decoding is needed; optional shadow monitoring is outside the reported policy-only latency. configurationmonitor
5.5 Implementation flow
- Separate cache consumption from production
Record each model’s action-independent video cache, then replay action denoising. Masking removes future attention; spatial shuffle, temporal swap, noise and frozen-present edit values under recorded keys. Final-clean replay instead fixes both K and V from a rollout-produced clean future. Comparisons are within-model; structural N/A entries are not zeros. interventionsource-discrepancy
- Write a position-aligned cache
Concatenate observed-frame tokens with learned future placeholders. First-frame tokens read only the observed frame, anticipation tokens read that frame and one another, and video tokens cannot read actions. One video-stack pass writes every layer’s future K/V. Full alignment uses 196 tokens on LIBERO and 240 on RoboTwin. cache-method
- Generate actions from the fixed interface
The action expert reads the same cache throughout action denoising. Observation and instruction retain their original conditioning paths. The policy constructs a fresh cache per chunk; there is no evolving video-cache trajectory within that chunk. cache-methodconfiguration
6. Experiments & results
RIFT retains explicit future conditioning for robot actions while replacing iterative video generation with one learned cache prefill. Its central distinction is between consuming a future representation and producing it. Paired interventions motivate a fixed future interface; separate training and simulation experiments test whether anticipation tokens can supply that interface efficiently.
6.1 Read the original evidence
Table 1. RIFT retains the future read near current-only deployment cost. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First compare the Fut. read and Roll-out columns: RIFT is the row with future attention enabled and iterative rollout disabled. SR is success percentage; Lat. is milliseconds per action chunk, and Rel. normalizes against Fast-WAM. The retained footnotes distinguish released checkpoints evaluated by the authors from their matched reproductions; Emb. PT. marks embodied pretraining. Section 5.1 reports one checkpoint per method and three evaluation seeds, each covering 2,000 trials across all forty tasks. Timing uses one A800 under each method’s original denoising configuration. Appendix A on page 10 specifies RIFT’s ten action-flow steps; its latency includes prefill and action denoising. liberoconfiguration
What it supports. RIFT reports 98.8±0.17% success at 247.9 ms/chunk, compared with Fast-WAM’s 96.8±0.27% at 235.7 ms. Rollout-based Joint and IDM report 98.4% and 98.6% at 780.2 and 1081.2 ms. The main gain is retaining the high-success tier while removing much of the deployment delay.
Where the evidence stops. The standard deviations describe evaluation seeds for one checkpoint, not independent training runs. LingBot-VA has different embodied pretraining, and original denoising settings differ. These timings therefore compare deployed configurations rather than a matched-compute scaling experiment.
Table 2. A second simulated embodiment tests the same deployment interface. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read Clean and Rand. as separate scene settings; each entry is a success percentage. Emb. PT. denotes embodied pretraining, while the adjacent checks identify future attention and iterative rollout. The Table 2 caption specifies one checkpoint and a single evaluation seed. Section 5.1 on page 6 states fifty tasks with one hundred trials per task in each setting. The matched Fast-WAM/Joint/PFD/RIFT family trains on 2,500 clean-scene and 25,000 randomized demonstrations for 30k steps; LingBot-VA is an externally pretrained comparison. Read the aggregate alongside the per-task Table 6 on page 12, which exposes failures hidden by an overall average. robotwinper-tasklimits
What it supports. RIFT has the highest reported aggregate values here: 92.9% clean and 92.6% randomized, compared with PFD’s 92.5% and 92.1%. Its future-read column stays enabled without a rollout. This extends the empirical result to bimanual simulation, while the margin over PFD remains small and is reported without uncertainty.
Where the evidence stops. The table does not show a RoboTwin latency measurement or physical-robot test. Aggregate ranking also does not hold for every task: Table 6 reports Open Microwave at 64/45% for RIFT versus 82/86% for LingBot-VA.
Table 5(a–c). Inspect where OOD gains hold and where failures remain. Original paper, p. 11 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. The three panels regroup the same 10,030 variants rather than introducing three independent test sets. Parenthesized header values are variant counts. The caption specifies one rollout per variant, environment/policy seed 42, initial-state index zero, benchmark task-language fields and no additional training. Overall is a variant-micro average, so equally averaging the seven category percentages would use the wrong weights. 'Uncl.' means 121 variants with null official difficulty labels; it is not a sixth difficulty level. Read across RIFT’s row to locate residual weaknesses, then vertically within a column to compare methods under the same perturbation or difficulty group. ood
What it supports. RIFT reaches 81.07% overall versus IDM’s 71.36%, leading all seven perturbation categories, all five labeled difficulty levels and all four source suites. Nevertheless, camera perturbations yield only 59.97% and Level 5 only 57.99%. The detailed table shows both broad comparative gains and substantial remaining errors.
Where the evidence stops. These are point estimates from a pinned single-rollout protocol, without repeated-seed intervals. RIFT does not lead the unclassified group: Joint reports 100.00% there. Robustness on these variants does not establish robustness to arbitrary deployment shifts.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| LIBERO manipulation success and latency 40 tasks; one checkpoint per method; three evaluation seeds, 2,000 trials each. One A800, original denoising configurations. | RIFT: 98.8±0.17%; 247.9 ms. Success mean±evaluation-seed std (%); ms/action chunk | Fast-WAM: 96.8±0.27%, 235.7 ms; PFD: 97.3±0.12%, 257.0 ms; Joint: 98.4±0.26%, 780.2 ms; IDM: 98.6±0.34%, 1081.2 ms; LingBot-VA: 98.5±0.08%, 2270.3 ms. Reported success remains near rollout methods, with 68.2–89.1% lower latency. Error bars measure evaluation noise, not training-run variation; LingBot-VA has different embodied pretraining. libero |
| Final-clean cache replay 2,000 paired trials across 40 LIBERO tasks per supported intervention; task-cluster bootstrap. | Joint: 1.908 [1.579,2.260], 97.9%; Cosmos-2: 1.690 [0.260,3.620], 98.2%. EE-ADE (cm), 95% CI; success (%) | Both Originals: 98.4% success. Joint masking: 18.725 cm and 9.7%. A fixed rollout-produced cache nearly preserves execution. This establishes consumption-side sufficiency, not one-pass production or separate key/value contributions. interventioncache-results |
| LIBERO-Plus OOD robustness No further training; all 10,030 variants; one rollout each, seed 42, initial-state index 0. | RIFT: 81.07%. Variant-micro success (%) | Fast-WAM 49.73%; Joint 68.06%; IDM 71.36%. The overall gain over IDM is 9.71 percentage points. RIFT leads across all seven perturbation categories, five labeled difficulty levels and four source suites. These are single-protocol point estimates. ood |
| RoboTwin 2.0 bimanual manipulation 50 tasks, clean/randomized scenes; 100 trials/task/setting; one checkpoint, single evaluation seed. | RIFT: 92.9/92.6%. Clean/randomized success (%) | PFD 92.5/92.1%; Fast-WAM 91.9/91.6%; Joint 91.0/91.1%; LingBot-VA 92.4/91.4%. Highest observed aggregate success among evaluated methods; no uncertainty is reported. Individual tasks can reverse the ranking. robotwinper-task |
| Anticipation supervision and token capacity LIBERO; three evaluation seeds; L2 versus full conditional-FM recipe; token sweep 2–196. | L2: 98.37±0.12%; full FM: 98.8±0.17%; both 247.9 ms. Success (%); full-alignment chunk latency | Full alignment is best for both; no-cache Fast-WAM reference is 96.75%. Figure 5’s FM peak is 98.78%, rounded in Table 1. More interface capacity helps overall, but the small recipe gap does not isolate FM from its conditioning curriculum. ablation |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 5. Capacity helps, while the supervision comparison remains a recipe comparison. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read token count m along the horizontal axis, whose successive labels mostly double; the final point is full alignment at 196 tokens. Blue circles use direct L2 future supervision and orange squares use conditional FM. The dashed 96.75% line is current-only Fast-WAM, not another point trained with a small cache. Both curves finish at their highest mean, though neither rises strictly at every intermediate count. Section 5.3 gives the full-alignment values, and Table 4 on page 11 gives suite-level results and evaluation-seed variation. The plot itself shows three-seed means without uncertainty bars; its vertical axis begins at 96.5%. ablationtraining
What it supports. Even small anticipation interfaces exceed the current-only reference in this sweep. At full alignment, the text reports 98.37±0.12% for L2 and 98.8±0.17% for the full recipe, with the same 247.9 ms deployment cost. More tokens improve the best achieved mean, while auxiliary supervision changes training rather than the deployed graph.
Where the evidence stops. The full FM recipe includes its conditioning curriculum, so the roughly 0.4-point difference is not an isolated test of loss type. The truncated vertical axis magnifies small differences, and the plot does not provide token-count-specific latency or training-run uncertainty.
7. Analysis & limitations
7.1 What the evidence leaves open
All evaluation is simulated. Qualitative future decodes diagnose representations and do not establish cache equivalence or physical-robot success. limitsdecodes
The introduction’s final-values-under-original-keys description conflicts with the methods, Figure 2 and Table 3, which replace both K and V. This report follows the repeated explicit full-cache protocol while preserving the discrepancy. Destructive value edits can also be out of distribution. source-discrepancyintervention
Figure 7’s caption describes average alarm lead, while Appendix E reports threshold crossing of the mean failure curve. These are different statistics. The latter does not establish individual-episode detection quality; both heads may agree on a wrong future. monitor
7.2 Questions for discussion
- Would RIFT retain success when its learned cache is subjected to the same paired interventions?
- How much of the full-recipe gain survives a factorial control for late perturbations and auxiliary-weight decay?
8. Reproducibility audit
8.1 Requirements and known gaps
Reconstruction needs Wan2.2-5B/Fast-WAM, the 1B action expert, aligned future latents and original masks. Appendix A specifies AdamW at 1e-4, weight decay 0.01, cosine annealing, mixed precision and clipping 1.0; LIBERO/RoboTwin budgets are 20k/30k steps. RoboTwin uses 2,500 clean and 25,000 randomized demonstrations. configurationrobotwin
Batch size, training GPU count, software versions and detailed FM-head architecture are not specified here. The optional monitor does not give numerical choices for sample count K, denominator stabilizer epsilon or the conformal calibration level. Proposed checks should isolate cache substitutions and supervision/curriculum effects. configurationtrainingmonitor
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: Resolve the cache intervention and test RIFT’s own future read
Reader-proposed check, not performed: first reproduce exact unedited replay for Joint using matched observations, video randomness, policy seeds and initial simulator states. Compare three separately named controls: final values under recorded keys, a fixed final-clean K/V cache, and a masked future read. Score common-prefix EE-ADE and task success with whole-task bootstrap intervals. Then apply matched masking and value reassignment to a trained RIFT checkpoint. Full-cache replay should remain close to Original if the methods’ claim holds; a materially different value-only result would expose the introduction’s ambiguity. If RIFT is insensitive to removing its own future read, the proposed mechanism would need reconsideration despite good aggregate success. interventioncache-resultssource-discrepancycache-method
Check 2: Separate FM supervision from the conditioning curriculum
Reader-proposed check, not performed: at full alignment with 196 tokens, compare L2 and conditional-FM supervision under identical Wan2.2-5B initialization, training data, 20k-step budget, action loss, and deployment graph. Cross loss type with late perturbations on/off and auxiliary-weight decay on/off; retain the detached probe identically so it cannot explain representation changes. Repeat independent training seeds, then the paper’s three evaluation seeds and 2,000 trials per seed. Report success, per-suite effects and prefill-plus-action latency. An FM benefit that persists across matched curriculum settings supports the supervision explanation; a benefit only in the combined recipe attributes the gain to an interaction instead. trainingconfigurationablationlibero
8.3 Reading coverage
Visual audit: Inspected the title/version and author block on page 1; all original Figures 1–9 and Tables 1–6 on their declared pages; method equations and attention/gradient routes on pages 3–6; training settings and intervention intervals on page 10; per-suite/OOD results, optional monitoring and limitations on page 11; per-task RoboTwin results on page 12; and diagnostic decodes on pages 13–14. All six final original crops were inspected. Figure 3’s arrows, fixed-cache route and stop-gradient label agree with Sections 4.2–4.3 and Appendix A. The final-clean protocol discrepancy on page 2 and the Figure 7/Appendix E alarm-statistic discrepancy are disclosed. All five text chunks, including reference pages 8–9, were read; those reference pages were not visually inspected. Separate supplemental availability remains unverified; code and external resources were not inspected and no experiments were run.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 10, 11, 12, 13, 14. 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
- 3 What does the action expert read? (3.1–3.5)
- 4 RIFT: One-pass Future-Token Imagination (4.1–4.3)
- 5 Experiments (5.1–5.4)
- 6 Conclusion
- References
- A Training implementation details
- B Uncertainty for the intervention study
- C Per-suite LIBERO success rates
- D LIBERO-Plus OOD robustness
- E Optional L2–FM uncertainty warning
- F Limitations and future work
- G RoboTwin results and future decodes
Outside the original text pass
- 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 extraction’s image omission was addressed by inspecting PDF pages 1–7 and 10–14, including every figure and table. Reference pages 8–9 were read as text.
- Reviewed arXiv:2608.11521v2 dated 13 August 2026. The catalog submission date is 12 August 2026; the earlier version was not supplied, so revision changes cannot be compared.
- All five author names match. The title page lists Australian National University (1) and Beijing Normal University (2), but Yikai Wang carries marker 3; his institutional mapping remains unresolved.
- Code and external resources were not inspected; no experiments were reproduced.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
identityPDF p. 1, title block and arXiv margin
The title and five authors match the catalog. The artifact is arXiv:2608.11521v2, 13 August 2026. Yikai Wang has superscript 3, although only affiliations 1 and 2 are listed.
Go to primary source ↓problemPDF pp. 1–2, Abstract and Section 1
The paper separates test-time future conditioning from the iterative video generation used to produce that conditioning.
Go to primary source ↓interventionPDF p. 3, Figure 2 and Section 3.1; p. 4, Sections 3.2–3.5, Eq. (1)
Each supported intervention uses 2,000 paired trials across 40 LIBERO tasks. Value edits retain recorded keys; final-clean replay fixes both K and V. EE-ADE uses recorded simulator positions over the common executed prefix, then task-macro averaging.
Go to primary source ↓cache-resultsPDF p. 3, Figure 2, Joint and Cosmos-2 rows; p. 10, Table 3, corresponding final-clean rows
Joint masking yields 18.725 cm EE-ADE and 9.7% success. Joint final-clean replay yields 1.908 cm (95% task-cluster CI 1.579–2.260) and 97.9%; Cosmos-2 yields 1.690 cm (0.260–3.620) and 98.2%. Both Originals have 98.4% success.
Go to primary source ↓source-discrepancyPDF p. 2, Section 1, final-clean paragraphs and contribution 2; p. 3, Figure 2 caption; p. 4, Section 3.5; p. 10, Table 3 caption
The introduction describes final values under original keys and calls the result value-side. The figure caption, methods and appendix instead explicitly replace both keys and values with one fixed final-clean cache.
Go to primary source ↓cache-methodPDF p. 4, Sections 4.1–4.2, Eq. (2); p. 5, Figure 3(a), Eq. (3)
RIFT retains Fast-WAM-Joint architecture, uses position-aligned anticipation tokens and creates complete per-layer future K/V in one prefill. First-frame tokens read only the observed frame; anticipation tokens read that frame and each other; video cannot read actions. Full alignment uses 196/240 tokens on LIBERO/RoboTwin, with action horizon 32.
Go to primary source ↓trainingPDF p. 5, Section 4.3, Figure 3(b), Eqs. (4)–(5); p. 6, Eq. (6) and gradient routes; p. 10, Appendix A
Two forwards share the video expert. Native video supervision uses no anticipation tokens or action supervision. Action and conditional-FM losses train the cache producer; stopped-gradient probe loss trains only the probe. Late perturbations exclude rows from action loss, and auxiliary weights decay.
Go to primary source ↓configurationPDF p. 10, Appendix A, Architecture and training configuration; p. 6, Section 5.1
In-house models use Wan2.2-5B and a 1B action expert (6B total), action width 1024, horizon 32, 10 action-flow steps and guidance 1.0. AdamW uses learning rate 1e-4, weight decay 0.01, cosine annealing, mixed precision and clipping 1.0. Training budgets are 20k/30k steps. Batch size, training GPU count, software versions and detailed FM-head architecture are not specified.
Go to primary source ↓liberoPDF p. 6, Section 5.1 and Table 1, all rows, SR/Lat./Rel. columns
LIBERO uses one checkpoint per method, three evaluation seeds and 2,000 trials per seed (50 per task). RIFT reports 98.8±0.17% and 247.9 ms/chunk; Fast-WAM 96.8±0.27%, 235.7 ms; PFD 97.3±0.12%, 257.0 ms; Joint 98.4±0.26%, 780.2 ms; IDM 98.6±0.34%, 1081.2 ms; LingBot-VA 98.5±0.08%, 2270.3 ms. Timing uses one A800 and each method’s original denoising configuration.
Go to primary source ↓robotwinPDF p. 6, Sections 5.1–5.2; p. 7, Table 2, all rows
RoboTwin 2.0 has 50 tasks and 100 trials per task per setting, with one checkpoint and single-seed evaluation. The matched family trains on 2,500 clean and 25,000 randomized demonstrations for 30k steps. Clean/randomized success is RIFT 92.9/92.6%, PFD 92.5/92.1%, Fast-WAM 91.9/91.6%, Joint 91.0/91.1%, LingBot-VA 92.4/91.4%.
Go to primary source ↓ablationPDF pp. 6–7, Section 5.3 and Figure 5; p. 11, Table 4, RIFT-L2 and RIFT (full) rows
Full alignment gives the best mean for both recipes. The text gives RIFT-L2 98.37±0.12%, full recipe 98.8±0.17%, and FM peak 98.78%; Table 4 rounds L2 to 98.4%. Both full-alignment recipes cost 247.9 ms/chunk. The full recipe includes a conditioning curriculum.
Go to primary source ↓oodPDF p. 6, Figure 4; p. 10, Appendix D; p. 11, Table 5(a–c) and caption
Without further training, four checkpoints are evaluated on 10,030 variants, one rollout each, environment/policy seed 42 and initial-state index 0. Overall variant-micro success is 49.73/68.06/71.36/81.07% for Fast-WAM/Joint/IDM/RIFT. RIFT leads every perturbation category, labeled difficulty level and source suite; unclassified difficulty is a separate 121-variant group.
Go to primary source ↓monitorPDF p. 7, Section 5.4 and Figure 7 with caption; pp. 10–11, Appendix E, Eqs. (7)–(8)
The optional monitor compares the detached L2 probe with FM samples using normalized disagreement and CUSUM. Calibration uses 1,967 successes; the mean curve over 33 failures crosses threshold 210 steps before endpoint 420. The figure caption instead phrases this as average alarm lead. Heads do not control actions and are excluded from policy latency; agreement can be confidently wrong.
Go to primary source ↓limitsPDF p. 11, Appendix F
Evaluation is simulation-only; physical robots and non-WAM fusion backbones are future work.
Go to primary source ↓per-taskPDF p. 12, Table 6, Open Microwave row and Average row
RIFT reports Open Microwave success of 64/45% clean/randomized, compared with LingBot-VA 82/86%, despite RIFT’s higher overall averages.
Go to primary source ↓decodesPDF p. 7, Figure 6; p. 13, Figure 8; p. 14, Figure 9 and captions
Matched future decodes are diagnostic, not proof of cache equivalence or executed behavior. Figure 9 compares Joint’s 20-step video rollout with one-pass RIFT states decoded by a trained linear diagnostic probe and frozen VAE, outside the deployed action path.
Go to primary source ↓8.5 Primary sources
Keep the Future, Drop the Rollout: RIFT for World Action Models ↗
PDF · 7,812 extracted words
Source fingerprint
2ddd27d9ab30be3e6253c5ec61024dd0942a7c965cf23ba24a5a15ef4c98c372