WAM-RL: World-Action Model Reinforcement Learning with Reconstruction Rewards and Online Video SFT
1. Paper overview
In one sentence: WAM-RL improves a video world model and its action translator together, using reconstruction rewards and constrained video adaptation to balance execution fidelity against latent-feature drift. e-frameworke-videoe-kle-rewarde-maine-reward-ablatione-reporting
| At a glance | What to know |
|---|---|
| Research problem | Author claim Demonstration-trained World-Action policies can struggle beyond their training distribution. Updating only the actor leaves its predictive input unchanged; updating the world model can instead shift the latent features on which the actor depends. The authors frame online learning as a coordination problem between better future prediction and faithful execution of those predictions. e-probleme-framework |
| Core mechanism | Source description A two-component post-training scheme combines successful-rollout video SFT, Gaussian-feature KL regularization and actor reinforcement learning driven by imagined-versus-executed reconstruction. e-frameworke-videoe-kle-reward |
| A key reported result | LIBERO-Object manipulation: 82% Task success rate (%). Table 1 benchmark evaluation; pretrained Base, actor-only πRL and WAM-RL. Evaluation split, episode count and seeds are not specified. Base 68%; πRL 78%. WAM-RL gains 14 and 4 percentage points, respectively. The strongest reported absolute improvement. No uncertainty estimates establish statistical significance or generalization beyond this evaluation. e-implementatione-main |
| Reading caution | Author claim The authors identify a stability–adaptation tradeoff: KL preserves actor compatibility but can restrict capability expansion. They also acknowledge limited success/failure contrast in manually designed or pretrained-feature rewards. e-limitations |
Core contributions
- Source description
A two-component post-training scheme combines successful-rollout video SFT, Gaussian-feature KL regularization and actor reinforcement learning driven by imagined-versus-executed reconstruction. e-frameworke-videoe-kle-reward
- Reader analysis
The reward comparison distinguishes success/failure separation from usefulness for policy optimization. Pixel reconstruction performs best among the tested rewards, despite optical flow's larger separation. Alignment with video learning is the authors' hypothesis, not an isolated causal finding. e-reward-ablatione-reward-diagnostic
Figure 1. Two feedback routes adapt prediction and execution while keeping their latent interface compatible. Original paper, p. 4 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Start at the blue World Model box on the right. Its downward arrow reaches the orange Actor, then Action and environmental execution. The separate Imagination branch supplies predicted imagery, while Observation returns executed imagery. Their comparison feeds Recon Reward, whose feedback arrow updates the actor. Follow the outer branch through If Success to Video SFT With KL Regularization: this updates the world model using successful trajectories. Section 3.3 and Eqs. (7)–(10) confirm that the gate selects video-training data and that KL anchors current features to a frozen pretrained reference. The left panel provides the authors' actor-only, sparse-reward comparison. e-frameworke-videoe-kle-rewarde-implementatione-reporting
What it supports. The method changes both the actor's learning signal and the predictive representation it consumes. The success gate and reconstruction branch perform different jobs: selected experience trains prediction, while imagined-versus-executed consistency trains action translation. This explains why the method needs explicit protection against feature drift.
Where the evidence stops. The original diagram says Real World Execution and abbreviates the method as WA-RL. Section 4 reports benchmark experiments, not a physical-robot study. The overview also supplies no numerical KL-only ablation or detailed update schedule.
2. Motivation
2.1 The problem and the proposed response
Demonstration-trained World-Action policies can struggle beyond their training distribution. Updating only the actor leaves its predictive input unchanged; updating the world model can instead shift the latent features on which the actor depends. The authors frame online learning as a coordination problem between better future prediction and faithful execution of those predictions. e-probleme-framework
2.2 What this reading follows
A World-Action policy can imagine a useful future yet fail to execute it, or faithfully execute a poor prediction. WAM-RL addresses both failure points: its actor learns from agreement between predicted and observed trajectories, while successful experience updates the video world model. A KL penalty anchors the changing features to a pretrained reference so that the actor can continue using them. Read the architecture alongside the two small success-rate tables, then compare the reward chart with the recovery example. The evidence supports improvements on the reported benchmarks, while leaving reward calibration, statistical uncertainty and the causal contribution of each update incompletely resolved. e-frameworke-videoe-kle-rewarde-maine-reward-ablatione-reporting
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
The recorded WAM and Dual-system categories are supported by the video generator and separately optimized latent-to-action actor. IDM is compatible with the actor's translator role, but the supplied paper gives neither an explicit inverse-dynamics target nor a dedicated IDM training formulation. Thus the exact Q4 assignment remains only partly verified. Coordinated training does not establish a single unified model. e-frameworke-implementatione-reward
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 Separate prediction error from action-translation error
Begin with the implementation's interface: a DiT video world model provides intermediate latent features, and an actor converts them into actions. Improving the actor can make execution better match the imagined trajectory while leaving that trajectory itself unchanged. WAM-RL therefore also trains the world model on successful online observation sequences. Figure 1's two feedback routes encode this division of work. The actor receives reconstruction-derived advantages; the video generator receives a prediction loss. My interpretation is that the scheme targets two different error sources, but the main result table does not isolate them experimentally because the full method changes both routes. At deployment the useful information flow is prediction to actor to execution; the paper leaves the replanning schedule unspecified. e-frameworke-videoe-rewarde-implementatione-maine-reporting
5.2 Understand what the latent KL actually constrains
The regularizer compares feature distributions, not merely final generated images. Section 3.3 starts with deterministic current features and corresponding features from a frozen pretrained world model. It treats these as Gaussian means, assigns diagonal covariances and evaluates KL from the current approximation to the reference. Current feature covariance follows an exponential moving average; the reference covariance stays fixed. The weighted KL term is then added to video prediction loss. The authors' rationale is that changing the predictor too abruptly makes its actor's learned feature interface ineffective. This creates a direct tradeoff: a stronger anchor can preserve compatibility but restrict adaptation. The paper acknowledges that limitation, yet provides no coefficient sweep or numeric KL ablation to locate a useful operating range. e-videoe-kle-limitationse-reporting
5.3 Interpret reward quality through the policy it trains
Equation (11) compares imagined future observations with observations obtained after executing the actor. The tested metrics emphasize different properties: pixels measure appearance discrepancy, optical flow captures motion, and DINOv2 or V-JEPA2 supplies feature comparisons. Figure 2 asks how those signals differ between successful and failed trajectories after normalization. Table 2 asks a different empirical question: which reward produces a more successful policy? Optical flow leads in the first comparison, while pixel MSE leads in the second. The authors attribute this to alignment with video prediction and penalties for departures from imagined behavior. My reading is narrower: the ranking mismatch is established by the supplied results, while exact scaling, variance, update controls and missing uncertainty prevent a confident causal explanation. e-rewarde-reward-diagnostice-reward-ablatione-reporting
5.4 Training and inference
During training
Successful observation sequences train video prediction. The actor receives policy-gradient updates with advantages computed from reconstruction rewards. Flow-SDE introduces stochastic denoising transitions and expresses the action likelihood as a sum of transition log likelihoods, providing the flow-policy training foundation. e-flowe-videoe-reward
The reported mixed online-RL/video-fine-tuning run uses eight NVIDIA A800 GPUs for eight hours. This describes the stated training setting; the paper does not itemize rollout cost, reward-encoder cost or baseline compute matching. e-implementation
During inference
Prediction supplies intermediate features to the actor, whose actions produce subsequent environmental observations. Video SFT, KL regularization and reconstruction rewards are training mechanisms. The paper does not specify deployment update frequency, replanning cadence, denoising steps or action-chunk length; its recovery illustration concerns one open-loop chunk. e-frameworke-implementatione-recoverye-reporting
5.5 Implementation flow
- Predict and translate
The implementation builds on Genie Envisioner-ACT. A DiT video generator predicts future observations; its actor consumes intermediate latent features to produce actions. Section 3.2 describes a world-model-to-actor information flow, without specifying the feature layer, action parameterization or observation interface in detail. e-frameworke-implementation
- Execute and compare
Actions execute in the environment. The ensuing observations are compared with the imagined trajectory to provide a dense actor reward. Figure 1 separately routes successful experience into video SFT: success filtering applies to the world-model update, whereas reconstruction supplies actor feedback. e-frameworke-reward
- Keep the actor's feature interface stable
The current deterministic feature and the corresponding feature of a frozen pretrained world model become Gaussian means. Both covariances are diagonal; current covariance statistics use an exponential moving average, while the reference covariance is estimated from pretrained features and then fixed. KL regularization penalizes drift toward an incompatible latent representation. e-kl
6. Experiments & results
WAM-RL post-trains a video world model and a separate actor through environment interaction. Successful rollouts refine video prediction under a latent KL constraint; reconstruction rewards teach the actor to realize imagined futures. Reported success rises from 68% to 82% on LIBERO-Object and from 19% to 22% on RLBench Water Plants. These modestly documented benchmark results motivate coordinated adaptation, but do not establish a general long-horizon remedy.
6.1 Read the original evidence
Table 1. WAM-RL improves both reported benchmark columns, with substantially lower absolute success on Water Plants. Original paper, p. 5 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read down each benchmark column before comparing across tasks. Base is the pretrained World-Action policy; πRL applies actor-only reinforcement learning; the bottom row is WAM-RL. LIBERO-Object rises from 68% through 78% to 82%. Water Plants moves from 19% to 18% under actor-only RL, then to 22% with WAM-RL. These are success percentages, so the full method's gains over Base are 14 and 3 percentage points. Section 4.1 names the implementation and training hardware, but does not specify evaluation episode counts, seeds or exact train/test separation. The two columns should therefore remain separate reported settings. e-maine-implementatione-frameworke-reporting
What it supports. The full method exceeds actor-only RL by four percentage points in each column. Water Plants supplies the clearest contrast with the authors' actor-only baseline because that baseline does not improve over pretraining there. The table supports the direction of the reported gains, while the stronger long-horizon explanation requires additional controls.
Where the evidence stops. No confidence intervals or trial counts accompany these rates. World-model adaptation and actor reward design change together in the full method, so this table alone cannot identify which component causes the improvement.
6.2 Results and evaluation conditions
| Task & protocol | Reported result | Comparison & interpretation |
|---|---|---|
| LIBERO-Object manipulation Table 1 benchmark evaluation; pretrained Base, actor-only πRL and WAM-RL. Evaluation split, episode count and seeds are not specified. | 82% Task success rate (%) | Base 68%; πRL 78%. WAM-RL gains 14 and 4 percentage points, respectively. The strongest reported absolute improvement. No uncertainty estimates establish statistical significance or generalization beyond this evaluation. e-implementatione-main |
| RLBench Water Plants Table 1 evaluation on the Water Plants simulation task; rollout count, seeds and detailed success protocol are unreported. | 22% Task success rate (%) | Base 19%; actor-only πRL 18%. Gains are 3 and 4 percentage points. Consistent with the authors' argument for world-model adaptation on a multi-step task, but absolute success remains low and a single task does not isolate horizon as the cause. e-maine-implementatione-limitations |
| Reconstruction-reward choice on RLBench Water Plants Table 2 reward ablation; update controls and evaluation counts are not fully specified. | Pixel MSE 21%; optical-flow MSE 19%; DINO MSE 16%; V-JEPA2 17% Task success rate (%) | Base 19%; πRL 18%. Pixel MSE is best among these reward choices. Figure 2 assigns the strongest reward separation to optical flow. Table 2's 21% pixel result and Table 1's 22% full-method result are distinct reported settings; the paper does not supply a factorial comparison that attributes their difference to one component. e-maine-reward-ablatione-reward-diagnostic |
6.3 Ablations and diagnostic examples
Read component removals and qualitative examples within their stated evaluation conditions.
Table 2. Pixel MSE yields the highest reported policy success among the four reconstruction-reward choices. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Use the first two rows as reference points, then compare the four reconstruction choices beneath them. Section 3.4 describes comparisons between imagined and executed future observations using appearance, motion or pretrained-feature information. Pixel MSE reaches 21%, optical-flow MSE 19%, DINO MSE 16% and V-JEPA2 17% on Water Plants. The table's DINO label refers to the DINOv2 feature family discussed in the text and Figure 2. Keep this ablation separate from Table 1's 22% full-method score: the source does not completely specify the shared update controls or present a factorial decomposition of the one-point difference. e-rewarde-maine-reward-ablatione-reward-diagnostic
What it supports. The most successful tested reward is the direct pixel comparison, two percentage points above Base in this table. More semantic feature comparisons do not improve the reported policy success. Together with Figure 2, this motivates examining how a reward interacts with optimization rather than ranking it solely by success/failure separation.
Where the evidence stops. The source does not report uncertainty, exact reward scaling or a complete matched-control protocol. The authors' explanation that pixel reconstruction aligns better with world-model learning remains a hypothesis rather than a separately measured mechanism.
Figure 2. Reward separation and downstream control performance rank the tested metrics differently. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Read the legend first: blue means Fail and orange means Success. The caption says each metric is normalized so that its success reward is 1, with the failure reward scaled accordingly; error bars represent twice the standard deviation. Compare the within-group separation rather than interpreting this as a shared raw reward scale. Flow has the largest visible separation, whereas the other pairs are closer. Pixel MSE's failure bar is slightly above its success bar, so do not assume every successful trajectory has a higher plotted statistic. Table 2 is needed to connect this diagnostic to actual policy success. e-reward-diagnostice-reward-ablatione-reward
What it supports. Optical flow's larger success/failure separation does not deliver the highest reported Water Plants success: its table score is 19%, versus 21% for pixel MSE. This is evidence against using separation alone to choose the reward; it does not prove the paper's proposed optimization-alignment explanation.
Where the evidence stops. The original has no numerical y-axis ticks, and the paper leaves the exact MSE-to-reward mapping unspecified. Bar heights cannot recover raw reward magnitudes or calibration. The plotted error bars describe reward dispersion, not confidence intervals for task success.
Figure 3. The qualitative example depicts recovery inside a single predicted action chunk after video fine-tuning. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.
How to read it. Follow each row from left to right using the printed stage labels. The upper row, without video SFT, begins at Fail and proceeds through No Recovery to OOD, meaning out-of-distribution behavior. The lower row, with SFT, begins at Fail and proceeds through Recovery to Success. Section 4.4 describes predicted trajectories that include gripper adjustment and another grasp attempt after online video training. Its continuation on page 7 explicitly places this behavior within a single open-loop chunk. Read the images as a qualitative illustration of the prediction behavior being claimed, and use the benchmark tables separately for measured policy success. e-recoverye-videoe-maine-reporting
What it supports. The illustrated mechanism is richer future prediction: a plan can include a correction after an unsuccessful grasp. Successful training trajectories can contain an intermediate failure followed by recovery, so filtering by final success is compatible with learning such patterns. That compatibility is a reader interpretation of the training scheme and example.
Where the evidence stops. These selected sequences provide no recovery frequency, matched-initial-state protocol or quantitative video-SFT-only effect. A depicted successful future is not evidence of physical execution or of feedback-based correction within the open-loop chunk.
7. Analysis & limitations
7.1 What the evidence leaves open
The authors identify a stability–adaptation tradeoff: KL preserves actor compatibility but can restrict capability expansion. They also acknowledge limited success/failure contrast in manually designed or pretrained-feature rewards. e-limitations
There is no numerical KL ablation, broad horizon sweep or physical-robot evaluation. Figure 3 illustrates predicted recovery in one chunk; it does not measure recovery frequency or establish closed-loop robustness. Missing trial counts and seed variability limit interpretation of small success-rate differences. e-maine-reward-ablatione-recoverye-reporting
7.2 Questions for discussion
- Does matched-budget video SFT improve measured recovery after controlled failed grasps, independently of reward choice?
- Would optical flow remain worse than pixel MSE after controlling reward scale, advantage normalization and random seeds?
8. Reproducibility audit
8.1 Requirements and known gaps
Reproduction needs the Genie Envisioner-ACT initialization, observation/latent interface, benchmark environments, success filtering and reward computation. Section 4.1 supplies hardware and duration, but no precise checkpoint, optimizer, learning rates, batch sizes, software versions, interaction budget or update schedule. e-implementatione-reporting
Resolve covariance estimation, EMA decay, λ_KL, reward sign/scaling, frame alignment, horizon, Flow-SDE discretization and advantage estimation before interpreting a rerun. These are consequential gaps in otherwise useful high-level objectives. e-flowe-videoe-kle-reward
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: Isolate successful-rollout video adaptation and its KL constraint
Reader-proposed, not run: start from one Genie Envisioner-ACT checkpoint and compare actor RL with a frozen world model against the same actor objective plus video SFT, with and without KL. Match environment steps, actor updates, success-selection rules and evaluation states on Water Plants. Sweep the KL weight in the SFT arm; record latent drift, success and recovery after a controlled failed grasp across multiple seeds. Report fixed-budget and compute costs separately. The proposed mechanism predicts that constrained SFT improves recovery while reducing instability relative to unconstrained SFT; unchanged recovery or equal stability would weaken that explanation. Missing implementation settings must first be resolved and disclosed. e-implementatione-videoe-kle-recoverye-maine-reporting
Check 2: Test whether pixel MSE wins after reward calibration is controlled
Reader-proposed, not run: compare pixel-MSE and optical-flow rewards on Water Plants from identical initial checkpoints with the world model frozen in both arms. Fix the interaction budget, frame alignment, comparison horizon, actor optimizer and advantage estimator. Specify the error-to-reward transform, then compare both original-scale and matched-scale rewards using the same held-out success/failure trajectories for calibration. Measure seeded policy success, reward dispersion and prediction/execution discrepancy. If flow closes the success gap after calibration, scale or optimization conditions offer an alternative explanation to intrinsic pixel alignment. If pixel remains better despite matched conditions, the authors' alignment hypothesis earns more focused support, though it is still not proved. e-rewarde-reward-diagnostice-reward-ablatione-flowe-implementation
8.3 Reading coverage
Visual audit: All eight original PDF pages were rendered and visually inspected, including the title/version/affiliations on page 1; flow and framework formulations on pages 2–3; Figure 1 and Gaussian KL details on page 4; actor reward, hardware and Table 1 on page 5; Table 2 and Figures 2–3 on page 6; recovery continuation and limitations on page 7; and terminal references on pages 7–8. All five final original crops were individually viewed. Figure 1's arrows and success gate were cross-checked with Sections 3.2–3.4 and Eqs. (7)–(12). Figure 2's normalization and error bars were checked against its caption; its absent numerical y-axis and unresolved reward transformation are preserved. Figure 3's row labels and open-loop interpretation were checked against Section 4.4 on both pages. No appendix is present. Separate supplements remain unverified; code and experiments are outside this reading.
PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8. 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
- Title, authors, affiliations and version stamp (p. 1)
- Abstract (p. 1)
- 1. Introduction (pp. 1–2)
- 2. Related Work (p. 2)
- 3. Method: 3.1–3.4, including Eqs. (1)–(12) (pp. 2–5)
- 4. Experiment: 4.1–4.4 (pp. 5–7)
- 5. Conclusion and limitations (p. 7)
- References (pp. 7–8)
Outside the original text pass
- Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout.
- The inspected title and six authors agree with the catalog after normalizing name order. The title page identifies arXiv:2606.17906v1 [cs.RO], 16 Jun 2026. No revision or edition difference was observed; no other version was supplied or compared.
- Acquisition omission preserved: Text extraction does not reconstruct figure images; inspect the retained PDF for figures and equation/table layout. This edition resolves that visual gap by inspecting all eight pages, all three figures, both tables and every final crop.
- Separate supplemental material availability has not been fully verified.
- No appendix is present in the supplied eight-page PDF. Code was not inspected and experiments were not reproduced. Cited works were not independently read.
The visual audit above records the subsequent illustrated pass.
8.4 Traceable evidence
e-identityPDF p. 1, title block, author affiliations and arXiv margin stamp
Observed title matches the supplied title. Authors are Zezhong Qian, Xiaowei Chi, Yu Qi, Haozhan Li, Zhi Yang Chen and Shanghang Zhang. Affiliations name Peking University, Northeastern University and Tsinghua University. The stamp is arXiv:2606.17906v1 [cs.RO], 16 Jun 2026.
Go to primary source ↓e-problemPDF pp. 1–2, Abstract and Section 1
The authors motivate online adaptation beyond demonstrations and warn that changing world-model latents can undermine the actor. They claim actor-only learning is insufficient for long-horizon improvement.
Go to primary source ↓e-flowPDF pp. 2–3, Section 3.1, Eqs. (1)–(6)
Flow matching trains a velocity field; Flow-SDE adds Brownian noise, models Gaussian denoising transitions, sums their log likelihoods and applies an advantage-weighted policy gradient. Discretization and advantage-estimator settings are not given.
Go to primary source ↓e-frameworkPDF p. 3, Section 3.2; p. 4, Figure 1 and caption
The world model supplies imagined futures/latents to the actor. Executed observations feed reconstruction reward for actor learning and, when successful, video SFT with KL regularization for the world model. The graphic labels execution as Real World Execution, while reported experiments are benchmark evaluations.
Go to primary source ↓e-videoPDF p. 3, Section 3.3, Eq. (7)
Successful rollout observation sequences train a video prediction loss, described generically as flow matching or reconstruction; unconstrained feature drift is said to destabilize the actor.
Go to primary source ↓e-klPDF p. 4, Section 3.3, Eqs. (8)–(10) and covariance paragraphs
Deterministic features are Gaussian means with diagonal covariances. Current covariance uses EMA feature statistics; the pretrained reference and its covariance are fixed. KL from current to reference is added to video loss. No coefficient or EMA decay value is supplied.
Go to primary source ↓e-rewardPDF p. 5, Section 3.4, Eqs. (11)–(12)
Reward compares predicted and executed future observations. Candidate similarities include pixel MSE, optical flow, DINOv2 and V-JEPA2. Actor policy gradients use reward-derived advantages; exact reward transforms, horizon value and estimator details are unspecified.
Go to primary source ↓e-implementationPDF p. 5, Section 4.1, Implementation Details, Datasets, Training Setup and Baselines
Genie Envisioner-ACT uses a DiT video generator and an actor consuming intermediate latent features. Eight NVIDIA A800 GPUs and eight training hours are reported. Evaluation names LIBERO-Object and RLBench Water Plants; baselines are pretrained Base and actor-only πRL.
Go to primary source ↓e-mainPDF p. 5, Section 4.2, Table 1, all rows and both benchmark columns
LIBERO-Object success: Base 68%, πRL 78%, WAM-RL 82%. RLBench Water Plants: Base 19%, πRL 18%, WAM-RL 22%. Table 1 reports no trial counts, seeds or uncertainty.
Go to primary source ↓e-reward-ablationPDF p. 6, Section 4.3, Table 2, all rows
Water Plants success rates are Base 19%, πRL 18%, Pixel MSE 21%, Optical Flow MSE 19%, DINO MSE 16% and V-JEPA2 17%. No uncertainty or full update-control specification accompanies the table.
Go to primary source ↓e-reward-diagnosticPDF p. 6, Figure 2, caption and Section 4.3 discussion
Blue bars denote failure and orange success. Each metric's success reward is normalized to 1, with failure scaled accordingly; error bars denote twice the standard deviation. Optical flow has the largest separation. The original chart has no numerical y-axis ticks. The authors hypothesize pixel-reward alignment and regularization explain its better policy result.
Go to primary source ↓e-recoveryPDF p. 6, Figure 3 and Section 4.4; p. 7, Section 4.4 continuation
The upper row without video SFT is labeled Fail, No Recovery, OOD; the lower row with SFT is labeled Fail, Recovery, Success. The text describes predicted correction within one open-loop chunk. No recovery-rate statistic is provided.
Go to primary source ↓e-limitationsPDF p. 7, Section 5, Conclusion and limitations paragraphs
The authors argue for improving both components, while acknowledging KL can constrain adaptation beyond pretraining and reconstruction rewards can have limited success/failure contrast.
Go to primary source ↓e-reportingPDF pp. 5–7, Sections 4.1–4.4 and 5; pp. 7–8, terminal References
The supplied paper ends with references and contains no appendix. Experiment reporting comprises two benchmark columns, one reward table, a normalized reward chart and a qualitative video-SFT comparison. No numerical KL ablation, physical-robot experiment, evaluation seed/count specification or detailed optimizer/software configuration is supplied.
Go to primary source ↓8.5 Primary sources
WAM-RL: World-Action Model Reinforcement Learning with Reconstruction Rewards and Online Video SFT ↗
PDF · 5,089 extracted words
Source fingerprint
301d8c6c804bececf605b4a8ac7d4848d16a02db8e32bd662ba3dd293ee3043f