An approach that uses a model's own confidence estimates to choose training episodes was associated with higher aggregate benchmark values than EVAC-v1 in a selection-only comparison. The reported relative differences were 5.0% for Reconstruction, 1.6% for Scene, 13.7% for Motion and 3.6% for Semantics.
Those results were not equally settled across the four measures. A paired bootstrap, a resampling method used to estimate how much a comparison may vary, used 10,000 percentile-bootstrap resamples at a 95% confidence level. The intervals for Reconstruction, Scene and Motion were +0.031 to +0.037, +0.012 to +0.018 and +0.018 to +0.070. The Semantics interval ran from -0.002 to +0.038, so it included zero.
A confidence score for choosing training data
The research concerns a robot world model, a system that predicts how a scene changes after an action. It tests whether confidence is a meaningful indicator of prediction error and whether confidence-guided active learning could improve post-training efficiency and final model quality under a fixed data budget. The document is arXiv:2608.25572v1, a preprint dated 26 Aug 2026.
ConfAL-WM uses a lightweight confidence probe built from decoder features. The probe produces dense confidence maps in latent space, the model's internal working representation, then aggregates them into risk scores for patches, frames and tasks. The pipeline includes a warmup stage, task prescreening and retraining on selected episodes, with optional frame or patch loss weighting.
The target data came from the Aloha-AgileX dual-arm robot across 50 manipulation tasks, with 500 randomized scenes per task. The paper reports 24,992 videos in total, with lengths ranging from 98 to 578 frames. It reports the protocol subsets as 6,248 episodes for confidence-probe training and EVAC-v1 warmup, 18,244 in the candidate pool and 7,298 selected for EVAC-v2 retraining.
To compare selection methods, the researchers summarized aggregate Reconstruction, Scene, Motion and Semantics values. Learned scoring methods were averaged over random seeds 42, 3407 and 123. Paired differences were estimated with 10,000 percentile-bootstrap resamples at a 95% confidence level.
The score tracked error, but not every detail
Before the retraining comparison, predicted risk was checked against a reference, or oracle, error measure in latent space. Spearman correlations, which compare the ordering of values, were 0.540 at patch level, 0.590 at frame level and 0.595 at task level. For the highest-error patches, AUROC, a ranking measure, was 0.761 and AUPRC was 0.146, compared with random baselines of 0.5 and 0.05. The score therefore ranked likely high-error patches better than random on the reported measures.
Risk maps showed greater consistency over time than exact spatial matching. Adjacent frames had a top-region IoU, an overlap measure, of 0.740 and a flicker score of 0.005. Risk and latent error reached a peak temporal correlation of 0.602 near zero lag, while the reported top-region spatial IoU was 0.130. The pattern indicates that risk tended to stay stable and line up in timing, but only partly matched the precise boundaries of high-error regions.
What the reported comparisons showed
Under selection-only retraining, the confidence method reported aggregate values of 0.7219 for Reconstruction, 0.9196 for Scene, 0.3643 for Motion and 0.5897 for Semantics. Relative to EVAC-v1, those corresponded to 5.0%, 1.6%, 13.7% and 3.6% higher, respectively.
The paper also reports a configuration with additional frame-and-patch weighting. Its aggregate values were higher than EVAC-v1 for Reconstruction, Motion and Semantics, and lower for Scene.
In a default-seed ablation without weighting, mean risk had the highest value on seven of nine component metrics. Tail risk had the highest Sem.-BLEU result, while persistent risk had the highest Sem.-CLIP result.
A benchmark result with clear boundaries
The paired mean differences for Reconstruction, Scene, Motion and Semantics were +0.034, +0.015, +0.044 and +0.018. Their 95% intervals were +0.031 to +0.037, +0.012 to +0.018, +0.018 to +0.070 and -0.002 to +0.038, respectively. The intervals for the first three dimensions stayed above zero, while the Semantics interval included zero.
The paper's own caveats are important. The confidence probe is tied to UNet decoder features and to EVAC training on RoboTwin2.0, and confidence lacks an independent representation measure. The authors also describe the world-model evaluation as incomplete because visual fidelity and motion accuracy may conflict.
Taken together, the paper reports a risk signal that ranked latent error better than random on the reported measures, alongside higher selection-only values in Reconstruction, Scene, Motion and Semantics. The evidence is tied to EVAC post-training on the reported RoboTwin2.0 Aloha-AgileX data, with the authors' caveats limiting how broadly the result can be read.
Paper data and sources
Original title: ConfAL-WM: Confidence-Guided Active Learning for Action-Conditioned World Models
Authors: Xiang Liu, Sen Cui, Changshui Zhang
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text