Making Latent Evolution Explicit: Operator-Structured Transitions for World Action Models
1. Paper overview
In one sentence: LEON makes latent transitions explicit through shared evolution operators and additive forcing, improving VLA-JEPA while preserving near-baseline performance in LaWAM's inference-time prediction pathway. e02e04e05e06e09e10e11e13e14
| At a glance | What to know |
|---|---|
| Research problem | Source description Predictive representation specifies what future information a WAM predicts; prediction–policy coupling specifies how that prediction influences actions. Neither fixes the transition architecture. The paper asks whether explicitly organizing latent evolution improves prediction and control while preserving the host model's targets and policy interface. e02 |
| Core mechanism | |
| A key reported result | LIBERO closed-loop manipulation: 99.0 / 99.8 / 99.4 / 98.0; average 99.05%. Success rate (%) and four-suite average. VLA-JEPA + LEON; Spatial, Object, Goal and LIBERO-10 suites; published baseline comparison. VLA-JEPA: 96.2 / 99.6 / 97.2 / 95.8; average 97.2%. Reported average gain: 1.85 percentage points. Highest average among listed methods; no WAM confidence intervals are supplied. e09e16 |
| Reading caution |
Core contributions
Figure 1. Context chooses how shared operator components act on the current latent state. Original paper, p. 1 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the upper panel as the interface that must remain intact: current latent X and action-related context Z produce a future latent. In the lower panel, follow X through the observable map phi to H. A second route combines H and Z into transition context xi. That context controls the coefficients alpha over the shared basis B_k and the additive term b. The central box applies a residual evolution step, after which the learned readout psi restores the host prediction representation. The bottom boxes separate the context-modulated operator from forcing; they describe complementary components of one update, not two competing predictions. e02e03e04e15
What it supports. The architectural commitment is that context changes an operator's composition within a common basis, while forcing supplies additional change. This makes the evolution rule an explicit part of the predictor. The shared structure concerns transition realization; the surrounding WAM still determines what the future target means and how it affects actions.
Where the evidence stops. Learned observables and context are not measured physical states or controls. The paper motivates LEON with Koopman generators but does not demonstrate an invariant Koopman subspace for WAM features or guarantee exact physical evolution.
2. Motivation
2.1 The problem and the proposed response
Predictive representation specifies what future information a WAM predicts; prediction–policy coupling specifies how that prediction influences actions. Neither fixes the transition architecture. The paper asks whether explicitly organizing latent evolution improves prediction and control while preserving the host model's targets and policy interface. e02
2.2 What this reading follows
A robot policy can learn from a predicted future without consulting that prediction when it acts. LEON studies the transition architecture across both this training-mediated use and a second use in which future features directly condition actions. Its intervention replaces a Transformer predictor with learned observables, context-modulated operators and additive forcing, while retaining the host model's targets and policy interface. Read the architecture first, then trace where each host consumes the prediction. The benchmark tables show a strong LIBERO result, a close RoboTwin comparison and uneven robustness gains. Controlled dynamics explain why the operator structure may help, while also exposing limits to extrapolation. e02e04e05e06e09e10e11e13e14
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 | Not applicable |
| Prediction paradigm | Other mechanisms |
| Quadrant | Outside quadrants |
3.1 Evidence-based assessment
Supports the recorded classification
The catalog's WAM, latent-prediction and generalization tags fit the targets and evaluations. Outside quadrants / Other mechanisms is appropriate for a transition module spanning two different policy couplings. Architecture Not applicable describes the module's taxonomy position, not absence of an architecture. Joint losses do not establish a single unified future/action generator; LEON is not presented as inverse dynamics. e02e05e06e11
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 Why a shared operator basis still permits nonlinear transitions
LEON borrows a specific idea from controlled Koopman generators: conditions can change the coefficients of common evolution components. In the exact motivating theory, a finite observable representation requires an invariant span. LEON does not establish that condition for robot latents. Instead, it learns the observable map, context construction, operator coefficients, forcing and readout. Reader interpretation: because context depends on the current observable as well as action-related information, the complete transition need not be a single globally linear map. The structural restriction is narrower: within a layer, context-dependent operator variation is assembled from a fixed collection of basis operators. Factor width 96 limits those independent operator directions, while the diagonal baseline and additive branch remain separate. This explains both the inductive bias and why an exact physical-law interpretation would exceed the evidence. e03e04e07e15
Figure 2. Locate the future-to-action connection before interpreting a policy gain. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Begin with the legend: dashed connections indicate training-time signals, while solid connections denote inference computation. Both panels contain a current state, action context, transition model and predicted future. The decisive difference is downstream of that prediction. In panel a, the predictive signal shapes policy-relevant representations during training; the future feature is not an action-expert input at inference. In panel b, a solid purple route connects the predicted future to the action expert. VLA-JEPA instantiates the first pattern and LaWAM the second. Read these as different functional uses of the same conditional transition, rather than assuming every WAM executes an imagined rollout. e02e05e06e10
What it supports. The two integrations test different consequences of replacing a predictor. VLA-JEPA can benefit through better training signals even without consuming future features during action generation. LaWAM requires the replacement's predicted horizon feature inside its inference pathway, so its matched RoboTwin result tests compatibility with that direct policy dependency.
Where the evidence stops. The diagram explains information flow, not the magnitude of prediction's causal contribution. Neither joint training nor a future-feature arrow establishes a unified future/action generator, model-predictive search, or that every action improves because the prediction is more accurate.
5.2 Follow supervision and action generation through each host
In VLA-JEPA, latent-action tokens condition the prediction of future V-JEPA features. LEON uses same-timestep visual-to-action cross-attention to construct transition context, without adding visual self-attention or cross-timestep mixing. Predictive supervision helps shape the latent actions, but the future prediction itself is not consumed by the action generator at inference. LaWAM supplies a different context: one global latent action conditions each patch through an MLP. Its frozen DINOv3 encoder has already contextualized the patches, and the predicted horizon feature remains an input to the action expert alongside the current feature. The original action objectives and interfaces are retained in both cases. Consequently, the LIBERO improvement and RoboTwin near-parity result concern different paths from prediction to executed benchmark actions; neither should be paraphrased as proof of online rollout planning. e05e06e07e09e10
5.3 Separate policy outcomes from evidence about the mechanism
The benchmark results answer whether complete transition replacement remains useful inside a WAM. They do not isolate every part of LEON. Reader analysis: the matched four-task LaWAM comparison offers clearer training control than a comparison with its fifty-task reference, yet it still changes a complete architecture. The controlled systems ask more specific questions. Spring amplitude extrapolation keeps the evolution law unchanged, separating unseen state magnitude from new dynamics. Pendulum tests then reveal a weaker advantage after a qualitative regime change. Finally, oscillator condition swaps test whether conditioning matters, and post-hoc branch removals show that the trained predictor uses both operator propagation and forcing. This sequence supports the architectural motivation while leaving open whether the same branch dependence explains robot success, or whether separately retrained reduced models could recover performance. e08e10e12e13e14e15
5.4 Training and inference
During training
VLA-JEPA retains joint future-state prediction and flow-matching action objectives. LaWAM retains latent-action distillation, horizon-feature supervision and conditional action learning. No extra predictive target or separate replacement-model pretraining is introduced. LaWAM's DINOv3 encoder is frozen. e06e07
The LaWAM comparison retrains both variants under the same four-task protocol. LIBERO and LIBERO-Plus baselines are reproduced from published tables; these are not described as a complete common retraining of all listed methods. e08e09e11
During inference
VLA-JEPA's future prediction shapes latent-action representations during training; its predicted future is not an action-generator input at inference. The retained flow-matching head produces actions. Thus its success gain does not establish a benefit from executing imagined rollouts online. e05
LaWAM predicts a global latent action, evolves the encoded current observation to a horizon DINO feature, then conditions its action expert on current and predicted features. LEON adds no patch–patch attention; upstream frozen DINOv3 features already have global context. Neither integration specifies a new planning or action-execution algorithm. e06
5.5 Implementation flow
- Lift the state into observables
A learned map converts each transition-state token from 1024 dimensions to a 768-dimensional observable. The readout returns updated observables to the host prediction space. These are learned features, not identified physical state variables. e04e05e06e07
- Construct transition context
VLA-JEPA uses each visual observable to query same-timestep latent-action tokens, then fuses that feature with their summary. LaWAM uses a per-patch MLP conditioned on the projected global latent action. e05e06
- Propagate, force and compose
Context sets coefficients over shared rank-one operators and an additive change. A scaled residual update evolves each observable. Eight layers replace VLA-JEPA's transition stack; twelve replace LaWAM's released runtime decoder. The respective residual scales are 0.35 and 0.10; both use factor width 96. e04e07
6. Experiments & results
LEON replaces a latent future predictor's Transformer transition with context-modulated operators and additive forcing. It improves VLA-JEPA's LIBERO success and retains near-baseline LaWAM performance on four RoboTwin tasks. These are two different uses of prediction: training-time representation shaping and inference-time action conditioning, respectively (e02, e05, e06, e09, e10).
6.1 Read the original evidence
Table 2. LEON raises the reported VLA-JEPA average across all four LIBERO suites. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start with the VLA-JEPA row in the predictive/latent-action group and compare it with the final VLA-JEPA + LEON row. Keep each suite aligned: Spatial tests spatial variation, Object object variation, Goal different task objectives, and LIBERO-10 longer multi-step tasks. All values are success percentages; the final column averages the four suites. Then scan the other rows to place the result within the paper's reported comparison. The table's family labels organize the listed systems, but the most relevant transition comparison is the pair of VLA-JEPA variants. The published baselines are copied from prior benchmark reports, as the original caption states. e05e08e09e16
What it supports. VLA-JEPA + LEON reports 99.0, 99.8, 99.4 and 98.0%, averaging 99.05%, against VLA-JEPA's reported 97.2% average. The paper reports a 1.85 percentage-point gain, with larger gains in Spatial, Goal and LIBERO-10 than the almost saturated Object suite. This is the highest average among the methods listed here.
Where the evidence stops. The table supplies no confidence intervals or evaluation counts. Published baseline rows are not all retrained under one documented common compute budget, and a representation-mediated gain does not establish a need for future prediction at inference.
Table 3. Matched training supports near-baseline policy-facing performance, with task-dependent changes. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. First read the Training scope column. The top LaWAM pair comes from a model trained on fifty tasks; the next two pairs are the controlled four-task comparison. Within each pair, compare clean and randomized evaluation separately before looking at the average. The task columns are Lift Pot, Beat Block Hammer, Dump Bin Bigbin and Hanging Mug. Moving from matched LaWAM to LEON improves clean Hanging Mug from 51 to 55%, but clean Dump Bin Bigbin falls from 97 to 94%. Keeping these columns visible prevents a nearly unchanged aggregate from being mistaken for uniform task-level equivalence. e06e08e10e16
What it supports. LEON averages 85.00% clean and 83.25% randomized success, versus 84.50% in both conditions for matched LaWAM. The paper reports combined performance of 84.13 versus 84.50%. A complete transition replacement therefore remains close to the baseline even when future prediction directly conditions the action expert.
Where the evidence stops. The fifty-task reference has a different training mixture and cannot isolate transition effects. Only four tasks are evaluated here; the small aggregate difference has no reported uncertainty and should not be described as statistically established equivalence.
Table 4. An improved aggregate conceals two robustness regressions. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read across the last two rows rather than stopping at the Avg. column. Camera, light, background and noise alter visual observations; layout changes spatial configuration. The paper distinguishes these from robot initial-state changes and language conditioning shifts. Compare LEON's value with VLA-JEPA's within every column, because the aggregate combines categories with very different baseline difficulty. Background rises from 93.6 to 98.0%, whereas robot falls from 67.1 to 63.0%. The table uses success percentages, so subtracting these values gives percentage-point changes. Bold cells indicate the strongest listed value within a column, not a guarantee that all perturbation families improve. e08e11e15
What it supports. The reported aggregate rises from 79.5 to 80.6%. Camera, light, background, noise and layout improve by 2.9, 3.3, 4.4, 2.3 and 2.6 percentage points. Robot and language instead decline by 4.1 and 3.9 points. The evidence supports selective robustness to visual and environmental changes.
Where the evidence stops. This is a benchmark-level outcome, not a direct measurement of latent dynamics accuracy. The category pattern alone cannot determine why robot and language degrade, nor establish that LEON handles arbitrary distribution shifts or deployment conditions.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| LIBERO closed-loop manipulation VLA-JEPA + LEON; Spatial, Object, Goal and LIBERO-10 suites; published baseline comparison. | 99.0 / 99.8 / 99.4 / 98.0; average 99.05%. Success rate (%) and four-suite average | VLA-JEPA: 96.2 / 99.6 / 97.2 / 95.8; average 97.2%. Reported average gain: 1.85 percentage points. Highest average among listed methods; no WAM confidence intervals are supplied. e09e16 |
| RoboTwin 2.0 four-task manipulation Matched four-task LaWAM training: Lift Pot, Beat Block Hammer, Dump Bin Bigbin and Hanging Mug; clean and randomized evaluation. | LEON: 85.00 clean, 83.25 randomized; combined 84.13. Success rate (%) | Matched LaWAM: 84.50 in each condition and combined. Published 50-task-trained reference: combined 83.50. Near-baseline aggregate performance, not an overall matched-protocol improvement. The 50-task reference uses a different training mixture. e08e10 |
| LIBERO-Plus robustness VLA-JEPA variants under the same seven-category perturbation protocol. | LEON aggregate 80.6; camera 66.2, robot 63.0, language 81.5, light 98.9, background 98.0, noise 68.6, layout 87.7. Category and aggregate success rate (%) | VLA-JEPA aggregate 79.5; respective categories 63.3, 67.1, 85.4, 95.6, 93.6, 66.3, 85.1. Aggregate improves 1.1 points, while robot and language fall 4.1 and 3.9 points. Robustness is selective. e11 |
| Damped-spring amplitude extrapolation Bounded-amplitude training; unseen amplitudes under the same linear dynamics; Figure 3(a) averages five seeds. | Adaptive Koopman: 0.0554 and 0.0018. 20-second energy MAE and energy-balance residual; lower is better | Fixed Koopman: 3.0315 and 0.0223; Transformer: 5.3546 and 0.0387. Fixed Koopman's near-exact scale equivariance does not guarantee the best long-horizon fidelity. These controlled predictors are not robot policies. e12 |
| Nonlinear-pendulum parameter and regime shifts Eight individual/joint length and damping shifts; separate initial-energy regime evaluation. | At L=0.7, Adaptive Koopman terminal RMSE is 0.5707. 8-second terminal angle RMSE (rad); separate full-rollout cumulative angle RMSE | Fixed Koopman 0.6066; Transformer 0.6351. Each structured model wins four of eight terminal-error settings. Cumulative-error advantages weaken after crossing the separatrix into rotation. Terminal and trajectory metrics must remain separate. e13 |
| Driven-oscillator condition use and path intervention Paired condition swaps, symmetric averaging over ten seeds; parameter-matched Transformer/MLP; post-hoc knockouts in the trained structured model. | Full structured: matched 0.08434, swapped 0.24944, retrieval 81.4%; removing operator 10.95× error, removing forcing 3.03×. Trajectory RMSE, condition retrieval and relative OOD rollout RMSE | Transformer matched 0.68908, retrieval 54.5%; residual MLP 0.22717 and 50.3%; full-model knockout reference 1.00×. Both paths are used in this controlled system. Knockouts do not establish what separately retrained reduced models could achieve. e14 |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Figure 3(a–c). Structured evolution helps extrapolation, but regime changes and branch removal expose its limits. Original paper, p. 9 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read each panel with its own metric. Panel a plots amplitude-OOD spring state RMSE against rollout seconds on a logarithmic vertical axis; its curves average five seeds. Panel b compares full-trajectory, eight-second cumulative pendulum angle RMSE, not Table 6's terminal error. The dashed separatrix separates oscillation-related shifts from topology OOD rotation; the caption specifies three seeds. Panel c normalizes driven-oscillator OOD rollout error by the full-model mean across ten seeds. Its operator and forcing removals are post-hoc interventions in an already trained model. Lower is better throughout, but the three vertical axes have different units or normalization. e12e13e14e15
What it supports. Adaptive Koopman's annotated pendulum advantage grows from 20.8% near OOD to 24.2% far OOD, then contracts to 5.5% across the regime boundary. Removing the operator raises driven-oscillator error to 10.95 times the full-model level; removing forcing raises it to 3.03 times. Both branches contribute in this trained controlled model.
Where the evidence stops. These are controlled dynamics tests, not WAM branch ablations or robot successes. Post-hoc removal differs from retraining a reduced model. The caption describes pendulum mean ± SD, but numerical SD values and distinct error bars are not recoverable from this graphic.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors explicitly reject an inference that learned WAM representations follow these physical systems or admit an exact Koopman description. The invariant-observable assumption behind the motivating theory is not established for learned WAM latents. e03e15
Full module replacement also changes token mixing and architecture. Without WAM-level operator/forcing or factor-width ablations, robot gains cannot be assigned uniquely to the shared operator basis. Four RoboTwin tasks do not establish performance across the full 50-task mixture. e05e06e08e14
WAM tables omit trial counts, seed variability and confidence intervals. The supplied experiments contain no physical robot deployment or runtime/compute benchmark, so neither real-world transfer nor efficiency gains are established. e16
7.2 Questions for discussion
- Does operator sharing retain an advantage when transition parameter count and token-mixing access are controlled? (e05, e06, e14)
- Which training changes could recover the robot/language perturbation losses without erasing visual robustness gains? (e11)
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs the host encoders, latent-action pathways, action heads, benchmark data and original objectives, plus Table 1's depth, dimensions, rank and residual scales. Preserve the matched RoboTwin training mixture and report each perturbation category separately. e05e06e07e08e11
The note does not provide a self-contained optimizer, learning-rate, batch-size, training-duration or loss-weight specification. It also leaves toy-system data-generation ranges, full dynamics settings and model configurations insufficiently specified. Recovering those details is necessary before claiming exact reproduction. e16
Proposed checks: compare capacity-controlled transition variants on matched RoboTwin tasks, and repeat driven-oscillator condition swaps with both post-hoc knockouts and separately retrained branch removals. These distinguish architecture effects from reliance induced by joint training. e08e14
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: Control capacity and conditioning in the policy-facing comparison
Reader-proposed experiment: recover the original training recipe, then train the Transformer transition, full LEON and a LEON variant with context-independent operator coefficients on the same four RoboTwin tasks. Keep encoder checkpoints, demonstrations, action expert, objectives, training steps and evaluation initial states fixed. Give the LEON variants identical conditioning access and comparable parameter budgets; report the Transformer budget separately. Repeat seeds and record clean/randomized success plus horizon-feature error for each task. If full LEON consistently outperforms the fixed-coefficient variant while using the same conditioning and compute, context modulation gains more direct support. If only model size or conditioning changes explain the difference, the shared-operator interpretation weakens. This check has not been run. e04e06e07e08e10e16
Check 2: Distinguish branch reliance from necessity in the oscillator
Reader-proposed experiment: first recover the missing driven-oscillator generator and training settings. Repeat the paper's paired condition swaps across ten seeds, then compare the same trained model with operator or forcing removed against reduced models retrained from scratch. Keep train/OOD splits, initial states, condition pairs, trajectory horizons and optimization budgets fixed. Report matched and swapped RMSE, retrieval accuracy and OOD error relative to the full model. Reproducing large post-hoc penalties would confirm branch reliance. If retrained reduced models close those gaps, the knockout ratios demonstrate dependence of the trained solution rather than necessity of both branches for the task. Failure to reproduce condition retrieval would challenge the proposed conditioning mechanism. This check has not been run. e14e16
8.3 Reading coverage
Visual audit: Inspected the original title/author/version page and Figure 1; Figure 2; Table 1; Tables 2–3; Table 4, Figure 3 and Table 5; and Tables 6–7. All six final original-PDF crops were separately viewed at their helper-produced dimensions, with complete diagram labels, table headers, legends and axes. Long captions and surrounding prose are excluded from crops; their protocol details are retained in the accompanying explanations. No separate supplement was supplied. Figure 3(b)'s caption states mean ± SD, but the inspected graphic does not permit recovery of numerical SD values or distinct uncertainty bars.
PDF pages inspected for this edition: 1, 5, 7, 8, 9, 10. 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
- PDF p. 1: title, authors, version, abstract and Figure 1
- PDF pp. 2–3: Section 1, Introduction
- PDF pp. 3–4: Sections 2.1–2.3, Background and Related Work
- PDF pp. 4–6: Sections 3.1–3.3, formulation, operator structure, both WAM instantiations and objectives
- PDF pp. 6–10: Sections 4.1–4.4, setup, policy results, distribution shift and controlled dynamics
- PDF p. 10: Section 5, Discussion and Conclusion
- PDF pp. 10–11: complete References
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- Identity: the inspected title page establishes arXiv:2608.27259v1 [cs.LG], 27 August 2026, Technical Note, August 2026. The title and all four authors agree with the catalog after normalizing title/subtitle layout and name order. No revision or edition difference was observed; other versions were not supplied.
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This acquisition limitation was addressed by inspecting original rendered pages and all six final crops.
- Separate supplemental material availability has not been fully verified.
- No supplement or appendix accompanies the supplied 11-page PDF. All six supplied text chunks were read individually, including the complete references.
- Code and external baseline reports were not inspected; no experiments were reproduced.
- Figure 3(b)'s caption describes mean ± SD over three seeds, but numerical SD values and distinct uncertainty bars are not recoverable from the inspected graphic; uncertainty is not estimated here.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e01PDF p. 1, title/author block and arXiv margin
Title and subtitle match the supplied observedTitle. Authors are Xiaoxiao Lu, Yunlong Dong, Jiahao Shi and Ye Yuan. Lu, Shi and Yuan are affiliated with the School of Artificial Intelligence and Automation, Huazhong University of Science and Technology; Dong with Principia AI. The margin states arXiv:2608.27259v1 [cs.LG], 27 Aug 2026.
Go to primary source ↓e02PDF pp. 2–4, Sections 1, 2.2 and 3.1; PDF p. 1, Figure 1; PDF p. 5, Figure 2
The paper distinguishes predictive representation, prediction–policy coupling and transition realization. LEON replaces the conditional latent transition while preserving host targets and interfaces. Figure 2 separates representation-mediated predictive training from policy-facing future-feature conditioning.
Go to primary source ↓e03PDF p. 4, Section 3.2, Equations (4–6)
The motivating system is control-affine. Its generator decomposes into drift and control components; a finite matrix representation requires an invariant observable span. LEON substitutes learned context modulation for prescribed affine physical-control coefficients.
Go to primary source ↓e04PDF pp. 4–5, Section 3.2, Equations (7–12) and Transition composition; Figure 1
State tokens are lifted to observables and combined with conditioning. The residual update uses multiplicative operator propagation plus additive forcing. D is diagonal, B_k are rank-one shared basis elements, and context-dependent variation lies in a subspace of dimension at most r. A learned readout and composition across depth form the complete transition.
Go to primary source ↓e05PDF pp. 5–6, Section 3.3, Representation-mediated VLA-JEPA; PDF p. 5, Figure 2(a)
Eight LEON layers preserve V-JEPA targets, VLM-produced latent-action tokens and the flow-matching action head. Same-timestep visual-to-action cross-attention and an action-token summary form context. There is no visual–visual or cross-timestep attention in the replacement stack. Predicted futures are not action-generator inputs at inference.
Go to primary source ↓e06PDF p. 6, Section 3.3, Policy-facing LaWAM; PDF p. 5, Figure 2(b)
All twelve runtime Transformer decoder blocks are replaced. Frozen DINOv3 produces 256 globally contextualized 768-wide patches; projections connect a 1024-wide transition state to 768-wide observables. Per-patch MLP context uses a global latent action. The predicted horizon feature and current feature condition the retained action expert.
Go to primary source ↓e07PDF p. 6, Section 3.3, Training objectives and intervention scope; PDF p. 7, Table 1
Original VLA-JEPA predictive/flow-matching and LaWAM distillation/horizon/action objectives are retained, without extra target or separate pretraining. Depth is 8/12, state/observable width 1024/768, factor width 96/96, residual scale 0.35/0.10 for VLA-JEPA/LaWAM.
Go to primary source ↓e08PDF pp. 6–7, Section 4.1, both setup paragraphs; PDF p. 8, Table 3 caption
LIBERO covers four suites and LIBERO-Plus seven perturbation families. LaWAM and LEON are retrained on the same four RoboTwin tasks with matched protocols. The published reference comes from full 50-task training and is explicitly contextual rather than matched.
Go to primary source ↓e09PDF p. 8, Table 2, VLA-JEPA and VLA-JEPA + LEON rows, all suite/Avg. columns; PDF p. 7, Section 4.2
LEON reports 99.0, 99.8, 99.4, 98.0 and average 99.05%; VLA-JEPA reports 96.2, 99.6, 97.2, 95.8 and 97.2%. The text reports +1.85 percentage points. Baseline table values are taken from prior reports; LEON has the highest listed average.
Go to primary source ↓e10PDF p. 8, Table 3, all six rows; PDF p. 7, Section 4.2, Policy-facing LaWAM
Matched LaWAM clean/randomized averages are 84.50/84.50%; LEON is 85.00/83.25%, with reported combined 84.13 versus 84.50. LEON task scores are clean 100/91/94/55 and randomized 100/93/93/47. Published 50-task LaWAM averages are 84.50/82.50, combined 83.50.
Go to primary source ↓e11PDF p. 9, Table 4, VLA-JEPA and VLA-JEPA + LEON rows; PDF pp. 7–8, Section 4.3
In camera/robot/language/light/background/noise/layout order, LEON reports 66.2/63.0/81.5/98.9/98.0/68.6/87.7, aggregate 80.6; VLA-JEPA reports 63.3/67.1/85.4/95.6/93.6/66.3/85.1, aggregate 79.5. Five categories improve; robot and language decline.
Go to primary source ↓e12PDF pp. 8–9, Section 4.4, Extrapolation under unchanged linear dynamics; PDF p. 9, Figure 3(a), caption and Table 5
Spring amplitude shifts preserve the linear law. Figure 3(a) plots log-scale state RMSE over rollout time, averaged over five seeds. Table 5 gives scale-equivariance error, 20-s energy MAE and energy-balance residual: Transformer 6.20e-2/5.3546/0.0387, Fixed Koopman 4.60e-8/3.0315/0.0223, Adaptive Koopman 8.66e-5/0.0554/0.0018.
Go to primary source ↓e13PDF p. 9, Section 4.4, Nonlinear dynamics and regime dependence, Figure 3(b) and caption; PDF p. 10, Table 6
Table 6 reports terminal 8-s angle RMSE under shifts in length L and damping c; at L=0.7, Transformer/Fixed/Adaptive values are 0.6351/0.6066/0.5707. Fixed and Adaptive each win four settings. Figure 3(b) instead measures cumulative 0–8-s angle RMSE; annotated improvements are 20.8% near OOD, 24.2% far OOD and 5.5% topology OOD. The separatrix is marked at energy approximately 19.62; the caption specifies three seeds and mean ± SD.
Go to primary source ↓e14PDF p. 10, Section 4.4, Condition dependence and transition-path intervention, Table 7; PDF p. 9, Figure 3(c) and caption
Transformer and residual MLP budgets match the full structured transition. Across ten seeds, full structured matched/cross-condition RMSE is 0.08434/0.24944, swap penalty 0.16510, retrieval 81.4%. Transformer matched error/retrieval is 0.68908/54.5%; MLP 0.22717/50.3%; fixed structured 0.84688/50.0%. Post-hoc removal of operator or forcing produces 10.95× or 3.03× OOD error relative to the trained full-model mean.
Go to primary source ↓e15PDF p. 10, Section 5, final two paragraphs
The authors limit the controlled-system interpretation: these experiments do not imply that learned WAM representations follow the corresponding physical dynamics or admit an exact Koopman description. Reported robustness gains are concentrated in visual/environmental shifts.
Go to primary source ↓e16PDF pp. 6–10, Sections 3.3 and 4.1–4.4; Tables 1–7 and Figure 3
The note supplies structural hyperparameters, retained objective names, benchmark categories and some controlled-system seed counts. It does not give a self-contained training recipe, WAM trial/seed uncertainty, measured hardware/runtime costs or a physical deployment experiment. Controlled studies describe shifts and metrics without complete numerical generation/model specifications.
Go to primary source ↓8.5 Primary sources
Making Latent Evolution Explicit: Operator-Structured Transitions for World Action Models ↗
PDF · 8,660 extracted words
Source fingerprint
9f21b25f0b884df9780bc9c380d011c0a962c88d45c619fe2876c0b193d88804