VLAW: Iterative Co-Improvement of Vision-Language-Action Policy and World Model

1. Paper overview

In one sentence: VLAW turns real policy failures into better visual dynamics, then uses filtered imagined successes to improve a separate robot policy, with reliability limited by both model fidelity and reward selection. e-problem e-loop e-filter-policy e-wm-metrics e-policy-results e-limits

At a glance	What to know
Research problem	Source description Real robot rollouts require resets and supervision. A simulator could multiply this experience, but demonstration-trained video models can depict success even when an action fails. VLAW targets this mismatch on contact-rich and deformable-object manipulation, where small physical errors can corrupt policy supervision. e-problem
Core mechanism	Source description An alternating procedure grounds a pretrained action-conditioned world model in successful and failed online rollouts, then uses its synthetic trajectories to post-train a separate VLA policy. e-loop e-world-training
A key reported result	Real-robot success across five manipulation categories: Ours-2: stacking 92%, wiping 86%, open book 86%, scooping 92%, drawing 78%; mean 86.8%. Ours-1 mean: 74.4%. Task success rate and unweighted mean. DROID Franka Panda with Robotiq gripper; multitask training for two iterations. Appendix B states 50 evaluations per task, except DSRL with 10. No separately specified held-out evaluation split or repeated-seed uncertainty. Base 46.0%; DSRL 50.0%; Filtered BC-1 64.8%; Filtered BC-2 75.2%. Ours-2 exceeds Filtered BC-2 by 11.6 percentage points. These are real execution results. Table 2 implies a 40.8-point gain over base; the abstract instead claims 39.2 points, an unresolved discrepancy. Section 5.3 claims matched online rollout budgets, but Appendix B explicitly gives fewer DSRL evaluations. e-setup e-policy-results e-identity
Reading caution	Source description Evaluation covers five task categories and world-model adaptation uses online data from those downstream tasks. Broader task coverage is explicitly future work. e-limits

Core contributions

Source description
An alternating procedure grounds a pretrained action-conditioned world model in successful and failed online rollouts, then uses its synthetic trajectories to post-train a separate VLA policy. e-loop e-world-training
Source description
Successful real and synthetic trajectories train the policy through flow matching; a fine-tuned video-language reward model filters imagined data. This avoids explicit action likelihoods and policy-gradient optimization. e-filter-policy e-rl
Reader analysis
The experiments connect improved interaction-outcome prediction to real-robot policy gains, with drawing-data and reward-threshold diagnostics. They do not establish general-purpose simulator accuracy. e-wm-metrics e-policy-results e-drawing e-reward

Two-stage pipeline: a VLA policy collects real data to fine-tune a world model and generate synthetic data; real and synthetic trajectories then feed policy training with a reward model. — **Figure 3.** Two separate learned models exchange training data through an iterative loop. Original paper, p. 3 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

2. Motivation

2.1 The problem and the proposed response

Source description

Real robot rollouts require resets and supervision. A simulator could multiply this experience, but demonstration-trained video models can depict success even when an action fails. VLAW targets this mismatch on contact-rich and deformable-object manipulation, where small physical errors can corrupt policy supervision. e-problem

2.2 What this reading follows

A robot can learn little from an imaginary success if its simulator has quietly removed the physical difficulty. VLAW addresses this by adapting a video world model to the policy’s own successful and failed attempts before generating more training experience. The policy then learns from successful real trajectories and imagined trajectories accepted by a video-language reward model. Read the evidence in three stages: whether the simulator predicts interaction outcomes more faithfully, whether the trained policy improves on the physical robot, and what the data and reward ablations reveal. The paper supports this loop on five manipulation categories, while leaving broader generalization and several evaluation details unresolved. e-problem e-loop e-filter-policy e-wm-metrics e-policy-results e-limits

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	Other mechanisms
Quadrant	Outside quadrants

3.1 Evidence-based assessment

Supports the recorded classification

Reader analysis

The separate VLA policy and action-conditioned video world model support Dual-system; the reward model is an additional evaluator. The mechanism is model-based policy post-training through filtered imagined experience. It is neither joint future/action prediction within one model nor inverse dynamics from a predicted future, supporting Other mechanisms and Outside quadrants. e-loop e-world-training e-filter-policy

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
Task instruction and current observation; the robot setup records two third-person cameras and one wrist camera. Training trajectories with recorded actions and binary task-success labels, plus original DROID data.	Robot actions from the updated VLA policy. Predicted future visual observations from Ctrl-World and trajectory-success labels from the separate reward model.

4.2 Equations and their role

\mathcal{L}_{\mathcal{D}_{\mathrm{real}}}=\mathbb{E}_{x_0,\epsilon,t^{\prime}}\left\|\hat{x}_0(x_{t^{\prime}},t^{\prime},c)-x_0\right\|^2,\qquad \mathcal{L}=\mathcal{L}_{\mathcal{D}_{\mathrm{real}}}+\lambda\mathcal{L}_{\mathcal{D}_{\mathrm{DROID}}}

Equations (1)–(2): x₀ is the future observation sequence; x at diffusion step t′ is its noised version, ε is the noise, and x̂₀ is the prediction. Conditioning c includes current observation and action chunk. λ controls DROID regularization. The paper gives no numerical λ. e-world-training

R(\tau^i)=\mathbf{1}\!\left[P(\text{yes}\mid\tau^i,I^i)>\alpha\right],\qquad \mathcal{L}=\mathbb{E}_{(o,a)\sim\mathcal{D}_{\mathrm{syn}}^{+}\cup\mathcal{D}_{\mathrm{real}}^{+}}\mathcal{L}_{\mathrm{FM}}(\theta;o,a)

Equations (3)–(4): R labels trajectory τⁱ for instruction Iⁱ; α is the acceptance threshold. A plus marks successful trajectories. The policy parameters θ minimize the flow-matching loss on observation–action pairs (o,a) from accepted real and synthetic data. e-filter-policy e-reward

5. Method in detail

5.1 Why failures belong in the simulator’s training set

Source description

Consider a scoop that never catches the intended contents. If a world model has mostly seen successful demonstrations, it may continue the action sequence with a successful transfer anyway. VLAW’s first intervention is to collect the current policy’s real attempts and preserve both outcomes in world-model training. Ctrl-World predicts future observations under the recorded action chunk, while DROID co-training regularizes adaptation to the smaller online dataset. Failure trajectories therefore teach the simulator what should remain unsuccessful; they are not discarded at this stage. Table 1 tests the resulting prediction change using identical recorded actions, and Figure 6 supplies selected visual comparisons. The method aims to make imagined supervision more faithful around the policy’s encountered states and actions. e-problem e-world-training e-wm-metrics e-replay

5.2 How imagined experience becomes a policy update

Reader analysis

After adaptation, the policy and world model alternate from a real initial observation: propose an action, predict its visual consequence, then propose the next action from that prediction. The resulting trajectory is a candidate training example, not an action sequence already validated on hardware. Qwen3-VL-4B-Instruct scores the video against the instruction, and a strict yes-probability threshold selects successes. The policy then learns through ordinary flow-matching regression on accepted synthetic and successful real trajectories. Appendix A relates this procedure to advantage-weighted policy improvement using an approximate projection, but the implemented selector is binary. Reader interpretation: the simplicity moves much of the burden from optimization into the quality and coverage of accepted trajectories. That is why both world-model fidelity and reward errors matter. e-loop e-filter-policy e-rl e-reward

5.3 Separate simulator evidence from physical improvement

Reader analysis

The paper’s strongest argument requires two different measurements. Five-second action replay establishes whether predicted observations and interaction outcomes agree with reality under recorded actions. Table 2 then measures the trained policy’s physical task success, reaching an 86.8% mean after two rounds compared with 75.2% for Filtered BC-2. Neither measurement alone answers the other question. The drawing ablation adds evidence that both real successes and a larger synthetic pool help in the first round, while the reward matrices expose the cost of conservative filtering. Reader interpretation: these results support the overall data-generation loop, but do not isolate which improvement in prediction causes which gain in control. Unequal DSRL evaluation counts, unspecified splits, and the abstract/table discrepancy further limit precision. e-wm-metrics e-policy-results e-drawing e-reward e-identity

5.4 Training and inference

During training

Source description

Warm-start π0.5 with 25 expert demonstrations per task category. Each of two iterations collects 50 real trajectories per category, fine-tunes the world model for 50,000 steps, generates 500 synthetic trajectories per task, and updates the policy for 2,000 steps at batch size 256. e-schedule

Source description

Qwen3-VL-4B-Instruct is fine-tuned on first-iteration labeled rollouts. Each reward input is a 16-frame video; training uses 200 steps and batch size 128. Synthetic success requires the probability of the yes token to exceed 0.8. e-filter-policy e-schedule e-reward

Reader analysis

The regularized-RL interpretation is approximate: Appendix A substitutes flow-matching regression for a likelihood-based projection, then motivates binary success weights using a discount approaching one and strongly penalized failures. The implementation does not estimate explicit per-transition advantages. e-rl e-filter-policy

During inference

Reader analysis

During synthetic collection, policy and world model interact autoregressively. During real rollouts, the policy’s actions execute on the robot and receive real observations. The described method uses imagination for post-training; it does not specify deployment-time world-model planning or inverse-dynamics action extraction. e-loop e-filter-policy

5.5 Implementation flow

Collect the policy’s own experience
Execute the current policy in the real environment. Label success when resetting the robot and append trajectories to a growing dataset. Failures remain useful world-model training examples, even though policy regression subsequently excludes them. e-loop e-world-training e-filter-policy
Fit action-conditioned visual dynamics
Fine-tune Ctrl-World to denoise future observations conditioned on the current observation and action chunk. Mix online experience with original DROID data to regularize the update; the source describes visual prediction rather than an explicit geometric or force state. e-world-training
Generate closed-loop imagination
Start from an observation sampled from a real trajectory. The policy samples an action using the instruction and predicted observation, the world model predicts the next observation, and that prediction feeds the next policy step. This produces complete synthetic trajectories. e-loop
Filter and refit the policy
The reward model scores each trajectory against its instruction. Retain predicted successes and combine them with successful real trajectories for policy training. Repeat real collection, world-model updating, imagination, and policy updating. e-filter-policy e-schedule

6. Experiments & results

VLAW improves a robot policy by first teaching its video world model what the policy actually does, including failures, then learning from reward-filtered imagined successes. Separate π0.5, Ctrl-World, and reward models form an iterative training pipeline. Real-robot results support useful synthetic supervision on five task categories, with unresolved reporting and evaluation details (e-loop, e-policy-results, e-identity).

6.1 Read the original evidence

Three world-model variants compared on five video metrics and four event-confusion counts; adding online data improves every video metric and reduces false-positive interactions from 11 to 1 relative to expert-only tuning. — **Table 1.** Online experience improves visual prediction and reduces falsely predicted interaction success. Original paper, p. 6 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

Real-robot success rates for six policy variants across stacking, wiping, opening books, scooping and drawing; Ours-2 reaches a mean of 0.868 compared with 0.752 for Filtered BC-2. — **Table 2.** After two iterations, VLAW reaches 86.8% mean real-robot success. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

6.2 Results and evaluation conditions

Task & protocol	Reported result	Comparison & interpretation
Real-robot success across five manipulation categories DROID Franka Panda with Robotiq gripper; multitask training for two iterations. Appendix B states 50 evaluations per task, except DSRL with 10. No separately specified held-out evaluation split or repeated-seed uncertainty.	Ours-2: stacking 92%, wiping 86%, open book 86%, scooping 92%, drawing 78%; mean 86.8%. Ours-1 mean: 74.4%. Task success rate and unweighted mean	Base 46.0%; DSRL 50.0%; Filtered BC-1 64.8%; Filtered BC-2 75.2%. Ours-2 exceeds Filtered BC-2 by 11.6 percentage points. These are real execution results. Table 2 implies a 40.8-point gain over base; the abstract instead claims 39.2 points, an unresolved discrepancy. Section 5.3 claims matched online rollout budgets, but Appendix B explicitly gives fewer DSRL evaluations. e-setup e-policy-results e-identity
World-model action replay fidelity 256 five-second wrist-view clips; 50 interaction clips for event counts. Recorded real actions are replayed from matching starting frames. Train/test separation is not specified.	Expert-plus-online: 21.77, 0.784, 0.136, 9.58, 64.12; TP/FN/TN/FP = 26/4/19/1. PSNR ↑; SSIM ↑; LPIPS ↓; FID ↓; FVD ↓; event confusion counts	Expert-only: 19.87, 0.748, 0.189, 12.76, 99.98; 28/2/9/11. Pretrained: 16.32, 0.634, 0.347, 41.03, 225.13; event counts unreported. Online data reduces false optimism substantially, but increases false negatives from 2 to 4. Video fidelity and event agreement measure model prediction, not robot success. e-wm-metrics
First-round drawing data ablation Drawing task; 50 real rollouts and either 500 or 250 synthetic rollouts before successful-trajectory filtering.	Real plus 500 synthetic: 68%; real plus 250 synthetic: 60%. Success rate	Synthetic only: 50%; real only: 46%; base: 22%. Combining the sources works best here. Accepted sample counts and matched optimization exposure are not reported, so this is not a pure test of data diversity at fixed training-data size. e-drawing e-filter-policy
Reward-model success filtering 40 manually labeled trajectories; direct yes/no versus probability threshold greater than 0.8.	Threshold: 10/12/2/16. TP/FN/FP/TN counts	Direct output: 15/7/8/10. Fewer false acceptances come with fewer retained true successes. The diagnostic does not measure the downstream policy effect of changing the threshold. e-reward

6.3 Ablations and diagnostic examples

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

Wrist-view sequences for stacking, drawing and scooping, with blue real-world borders, dark pretrained borders, yellow expert-only borders and green online-adapted borders. — **Figure 6.** Matched actions expose outcome differences that visual realism alone can hide. Original paper, p. 7 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

First-round drawing success bars: base 22%, real only 46%, synthetic only 50%, real plus half synthetic 60%, and real plus full synthetic 68%. — **Figure 9.** Drawing benefits from both real successes and a larger synthetic rollout pool. Original paper, p. 8 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

Two confusion matrices over 40 trajectories: direct yes/no yields 15 true positives, 7 false negatives, 8 false positives and 10 true negatives; thresholding yields 10, 12, 2 and 16 respectively. — **Table 3.** A conservative reward threshold removes false successes but discards additional true successes. Original paper, p. 14 ↗
Excerpt from the authors’ paper; cropped without altering the figure or table.

7. Analysis & limitations

7.1 What the evidence leaves open

Source description

Evaluation covers five task categories and world-model adaptation uses online data from those downstream tasks. Broader task coverage is explicitly future work. e-limits

Reader analysis

Long-horizon examples show imagined trajectories up to 20 seconds, but quantitative replay uses five-second clips. Selected plausible videos do not establish long-horizon physical correctness or transfer to unseen tasks. e-long-rollouts e-wm-metrics

Reader analysis

No uncertainty bars, repeated-seed analysis, or explicit held-out split accompanies the main results. Appendix B relates evaluation counts to collected online rollouts, leaving evaluation/training separation unclear. e-policy-results

7.2 Questions for discussion

Does better failure-event calibration predict physical policy gains when synthetic training volume is held fixed (e-wm-metrics, e-drawing)?
How should reward precision and successful-trajectory coverage be traded off across tasks (e-reward)?

8. Reproducibility audit

8.1 Requirements and known gaps

Reader analysis

Reproduction requires the π0.5, Ctrl-World and Qwen3-VL-4B-Instruct initializations, original DROID data, a three-camera Franka/Robotiq setup, expert demonstrations and success-labeled online interactions. The described schedules alone do not specify learning rates, optimizer settings, λ, frozen modules, accelerator counts, wall-clock cost, software versions or complete action/state encoding. e-setup e-world-training e-schedule e-reward

Reader analysis

Operationalize the outcome criteria before evaluation: Appendix B leaves the stack holding duration, book-opening angle, minimum scooped amount and drawing gap tolerance qualitative. Preserve the explicitly specified reward threshold of 0.8 rather than confusing it with those missing task thresholds. e-setup e-reward

Reader analysis

Proposed checks: hold accepted training volume fixed when comparing real/synthetic mixtures, and independently audit reward thresholds for both false acceptance and lost true successes before measuring physical policy outcomes. These are proposed experiments, not reproduced results. e-drawing e-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: Does the drawing gain survive matched training exposure?

Reader-proposed experiment: repeat the first-round drawing comparison from the same warm-start policy and real rollout pool. Compare real-only training, real plus 250 synthetic candidates, and real plus 500 candidates. Record acceptance counts, then add controls that equalize accepted transitions, real/synthetic sampling ratios where applicable, update steps and batch size; oversample real successes for a matched-exposure real-only control. Evaluate independent reset trials with a preregistered circle-closure tolerance and repeated seeds. A remaining mixture advantage would support useful additional synthetic coverage. If the gain disappears after exposure is matched, the original result would be explained more narrowly by training quantity or sampling. e-drawing e-filter-policy e-schedule e-setup

Check 2: Does conservative reward filtering improve physical learning?

Reader-proposed experiment: freeze one world model and candidate rollout pool, then compare direct yes/no filtering with strict probability thresholds of 0.5, 0.8 and 0.9. Use a separate human-labeled set, stratified by task and excluded from reward-model fitting, to measure false acceptances and rejected true successes. Train matched policy copies with both natural accepted-set sizes and a size-matched control, keeping update exposure fixed. Evaluate real-robot success on independent trials. The mechanism is supported if reduced label contamination improves control without unacceptable loss of successful coverage; it is weakened if stricter filters merely reduce false positives while yielding equal or worse physical success. e-reward e-filter-policy e-policy-results

8.3 Reading coverage

Visual audit: Visually inspected the title, authors, affiliations and both date markings on page 1; overview and pipeline figures on pages 1–3; method equations and Algorithm 1 on page 4; setup and regularized-RL formulation on page 5; world-model table and rollout figure on page 6; replay and policy plots on page 7; candidate rollouts, drawing ablation and limitations on page 8; derivation and task criteria on page 13; and evaluation counts, detailed policy results, reward configuration and confusion matrices on page 14. All six final original crops were inspected. Figure 3 was cross-checked against its caption, Equations (3)–(4), Algorithm 1 and Appendix A: its compact synthetic-data arrow omits the explicitly described policy/world-model alternation, and its advantage-weighted label is interpreted through the implemented binary filtering. The strict reward inequality and all table orientations were checked. Pages 9–12 contain reference material read as text; external videos and separate supplements were not inspected.

PDF pages inspected for this edition: 1, 2, 3, 4, 5, 6, 7, 8, 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 (pp. 1–2)
2. Related Works, including 2.1–2.2 (pp. 2–3)
3. Preliminaries (p. 3)
4. Co-Improvement of VLA and World Model, including 4.1–4.3 and Algorithm 1 (pp. 3–5)
5. Experiments, including 5.1–5.3 (pp. 5–8)
6. Conclusions and discussions (p. 8)
Impact Statement; Acknowledgment; References (pp. 9–12)
Appendix A. Relation to Regularized Reinforcement Learning (p. 13)
Appendix B. Task Details (pp. 13–14)
Appendix C. Reward Model Details (p. 14)

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.
All five supplied text chunks were read, covering the complete 14-page PDF. Text extraction does not reconstruct figure images; this gap was addressed by inspecting PDF pages 1–8 and 13–14 and all final crops. Reference-only pages 9–12 were read as text.
Separate supplemental material availability has not been fully verified. No separate supplements, linked videos, code, or checkpoints were inspected, and no experiments were reproduced.
Identity and edition: the observed title and all six authors match the catalog. The title page identifies arXiv:2602.12063v2 [cs.RO], 15 February 2026, while its footer says Preprint, February 17, 2026. Both dates are preserved without reconciliation. The catalog lists ICML 2026; a proceedings edition and other revisions were not supplied or compared.

The visual audit above records the subsequent illustrated pass.

8.4 Traceable evidence

e-identityPDF p. 1, title/author block, arXiv margin, preprint footer and AbstractInspect

Title and six authors match; affiliations are Stanford University and Tsinghua University. Margin identifies v2, 15 February 2026; footer dates the preprint February 17. Abstract claims 39.2% absolute improvement over base and 11.6% from synthetic rollouts.

Go to primary source ↓

e-problemPDF pp. 1–3, Introduction and Section 4.1 openingInspect

Real collection is costly; demonstration-heavy world models can be optimistic and miss contact/deformation details. Online policy experience supplies successes and failures.

Go to primary source ↓

e-loopPDF p. 3, Figure 3 and caption; Sections 3–4Inspect

Separate policy and world model interact from real initial observations to generate imagined trajectories. Real rollouts ground the world model; real and synthetic data feed policy learning.

Go to primary source ↓

e-world-trainingPDF p. 4, Section 4.1, Eqs. (1)–(2); Algorithm 1 lines 4–7Inspect

Growing real rollout data, including failures, fine-tunes pretrained Ctrl-World through future-observation denoising; DROID co-training regularizes it with λ.

Go to primary source ↓

e-filter-policyPDF p. 4, reward-model paragraph, Eqs. (3)–(4); Algorithm 1 lines 9–12Inspect

First-iteration labels fine-tune the reward model. A strict yes-probability threshold filters synthetic success. Binary trajectory weights retain only successful real and synthetic samples for flow matching.

Go to primary source ↓

e-rlPDF p. 5, Section 4.3, Eqs. (5)–(7); p. 13, Appendix A, Eqs. (8)–(11)Inspect

The authors motivate weighted flow matching using a regularized policy-improvement projection and offline approximation; binary filtering is related through limiting discount and failure-penalty assumptions.

Go to primary source ↓

e-setupPDF p. 5, Figure 4 and Section 5.1; pp. 13–14, Appendix B, Success CriteriaInspect

Franka Panda, Robotiq gripper, two third-person cameras and one wrist camera serve stacking, opening books, wiping, scooping and drawing. Several success tolerances remain qualitative.

Go to primary source ↓

e-schedulePDF pp. 5–6, Section 5.1, Base Models and HyperparametersInspect

π0.5 and Ctrl-World are the initial models. Twenty-five expert demonstrations warm-start each category; two iterations use 50 real rollouts/category, 50K world-model updates, 500 synthetic rollouts/task, and 2K policy updates at batch 256.

Go to primary source ↓

e-wm-metricsPDF p. 6, Table 1, all rows/columns and caption; Section 5.2Inspect

Metrics use 256 five-second wrist-view clips and event counts use 50 interaction clips. Online adaptation improves all reported video metrics and changes expert-only event TP/FN/TN/FP from 28/2/9/11 to 26/4/19/1.

Go to primary source ↓

e-replayPDF p. 7, Figure 6 and captionInspect

Colored borders compare real trajectories, pretrained predictions, expert-only predictions and online-adapted predictions using identical starting frames and five recorded action chunks; only wrist views are shown.

Go to primary source ↓

e-long-rolloutsPDF p. 6, Figure 5 and caption; p. 7, Section 5.2, Policy-in-the-loop rollout; p. 8, Figure 8 and captionInspect

Selected policy-in-the-loop examples run for 20 iterations/seconds. Figure 8 displays diverse imagined candidates from real initial frames, with real failures and some imagined successes.

Go to primary source ↓

e-policy-resultsPDF p. 14, Appendix B evaluation paragraph and Table 2, all rows/columns; p. 7, Figure 7 and Section 5.3 baselinesInspect

Table 2 means are 0.460, 0.500, 0.648, 0.752, 0.744 and 0.868. Appendix B states 50 evaluations per task except 10 for DSRL; Section 5.3 states equal online rollout budgets. No repeated-seed uncertainty is given.

Go to primary source ↓

e-drawingPDF p. 8, Figure 9, legend, bars and caption; Section 5.3 AblationsInspect

First-round drawing rates are 22, 46, 50, 60 and 68 percent for base, real only, synthetic only, real plus half synthetic and full mixture. Caption specifies 500 versus 250 synthetic and 50 real rollouts.

Go to primary source ↓

e-rewardPDF p. 14, Appendix C and Table 3, both confusion matricesInspect

Reward inputs contain 16 frames; Qwen3-VL-4B-Instruct trains for 200 steps at batch 128. On 40 labeled trajectories, strict p(yes)>0.8 changes TP/FN/FP/TN from 15/7/8/10 to 10/12/2/16.

Go to primary source ↓

e-limitsPDF p. 8, Section 6, Conclusions and discussionsInspect

The authors explicitly limit evaluation to five categories and identify broader online task coverage as future work.

Go to primary source ↓

8.5 Primary sources

VLAW: Iterative Co-Improvement of Vision-Language-Action Policy and World Model ↗

PDF · 8,054 extracted words

Source fingerprint

34ba87492a0fca9d382c366a80884eef99bc1101d973bbbb64125ae78bedffb8