ConfAL-WM

01 · Overview

Confidence closes the loop.

Estimate where the world model is unreliable, spend the post-training budget there, then strengthen supervision on the difficult frames and patches.

Turn on sound for the full experience

UNet-native confidence

Decoder-feature confidence probe with an EMA-calibrated training target for dense latent-space reliability estimation.

Efficient data selection

Mean risk becomes an acquisition signal for staged, budget-aware active post-training on new task and scene distributions.

Effective data enhancement

Frame-and-patch confidence weighting focuses training on unreliable spatiotemporal regions instead of treating every target equally.

Better post-training

Confidence improves selection and weighted retraining over scalar reward, progress, preference, and judge-based scoring baselines.

02 · Method

Dense confidence, from UNet features to retraining.

One lightweight probe serves two downstream roles: data acquisition and localized data enhancement.

TODO · Figure 2
Training and inference of the confidence probe
assets/figures/figure2_confidence_probe.webp

Decoder-feature probe

Tap selectable UNet decoder features: enough spatial locality for patch-wise confidence, while retaining global context.

EMA-calibrated target

Local latent denoising error is converted into binary confidence supervision using a stable adaptive threshold band.

Mean-risk acquisition

Dense risk maps are aggregated across future frames and spatial patches to prioritize informative tasks and scenes.

Frame + patch weighting

The same dense confidence output becomes a training weight, emphasizing difficult frames and local regions during EVAC-v2 retraining.

03 · Confidence Visualization

Where does the model fail?

50 tasks are organized into 10 groups. Each group shows five tasks × six synchronized video views. Toggle between pixel space (GT / predicted frames) and latent space (latent magnitudes) for the first, second, and last columns.

04 · Active Learning Results

Post-training numerical results.

Switch between the main bar-chart comparison and three table views. The seed selector updates the same result space.

05 · Qualitative Evolution

Base EVAC → v1 → confidence-guided v2.

Each episode pairs five synchronized videos with a compact metric table. Ten episodes across four pages; the first six match the paper.

06 · More Experiments

From confidence signals to design choices and data budgets.

Explore confidence diagnostics, controlled probe ablations, and fixed-update budget comparisons.

Patch / Frame / Task Spearman 0.540 / 0.590 / 0.595Top-5% patch AUROC 0.761Adjacent-frame top-region IoU 0.740Risk flicker 0.005Peak temporal correlation 0.602

**High-error detection**Risk is used to find the patches with the largest prediction errors. It performs above chance in latent and pixel space; for the top 5% latent-error patches, AUROC reaches 0.761.

**Task-level risk and error**Each point compares a task’s average risk with its average prediction error. Higher risk generally means larger error, with a stronger relationship in latent space (Spearman ≈ 0.60) than in pixels (≈ 0.30).

Appendix B.3 · How feature choice and threshold supervision affect confidence quality.

Table 9 · Feature and supervision comparison
Feature	Supervision	Brier ↓	BCE ↓	ECE ↓	AUROC ↑
Bottleneck	Tuned fixed	0.089487	0.288859	0.00567	0.78379
Bottleneck	EMA	0.089551	0.289220	0.00639	0.78487
Decoder	Tuned fixed	0.084181	0.271915	0.00402	0.82237
Decoder	EMA	0.084243	0.272314	0.00643	0.82184

With the same probe size and training setup, decoder features reduce Brier by about 5.93% compared with bottleneck features. EMA and calibrated fixed thresholds perform similarly, with tuned fixed slightly ahead for decoder features.

**Figure 15 · Shared-threshold performance**All four probes answer the same nine error thresholds. Fixed and EMA curves nearly overlap, while changing from bottleneck to decoder features produces the larger improvement.

Table 10 · Complete threshold-range ablation
Feature	Supervision	Cal. Brier ↓	Test Brier ↓	Test ECE ↓	Test AUROC ↑	Low-noise Brier ↓
Bottleneck	Fixed [0.20, 0.70]	0.154102	0.153483	0.12854	0.77596	0.090326
Bottleneck	Fixed [Q_.10, Q_.90]	0.092173	0.089468	0.00670	0.78338	0.057385
Bottleneck	Fixed [Q_.05, Q_.95] †	0.092149	0.089487	0.00567	0.78379	0.057210
Bottleneck	Fixed [Q_.20, Q_.80]	0.092776	0.090206	0.01182	0.78185	0.058627
Bottleneck	EMA	0.092239	0.089551	0.00639	0.78487	0.057544
Bottleneck	EMA-Qinit	0.092101	0.089477	0.00694	0.78517	0.057503
Decoder	Fixed [0.20, 0.70]	0.133136	0.133257	0.10630	0.81088	0.085256
Decoder	Fixed [Q_.10, Q_.90]	0.086885	0.084235	0.00617	0.82313	0.054462
Decoder	Fixed [Q_.05, Q_.95] †	0.086872	0.084181	0.00402	0.82237	0.054344
Decoder	Fixed [Q_.20, Q_.80]	0.087667	0.085248	0.01142	0.82040	0.056637
Decoder	EMA	0.086867	0.084243	0.00643	0.82184	0.054400
Decoder	EMA-Qinit	0.086840	0.084266	0.00548	0.82235	0.054522

All 12 fits compare manually chosen, training-quantile, and adaptive threshold ranges. Quantile-based ranges and EMA are much stronger than [0.20, 0.70]; † marks the fixed range selected using calibration data.

Q denotes a training-error quantile. EMA-Qinit starts at [Q_.10, Q_.90]; bold values mark the best result within each feature group.

Table 11 · Paired task-level Brier differences (×10³)
Contrast	Training noise · Δ	Training noise · 95% CI	Low noise · Δ	Low noise · 95% CI
Decoder − bottleneck (fixed)	−5.331	[−5.657, −4.987]	−2.767	[−3.156, −2.243]
Decoder − bottleneck (EMA)	−5.294	[−5.637, −4.934]	−3.141	[−3.475, −2.762]
EMA − fixed (decoder)	+0.072	[+0.011, +0.134]	−0.016	[−0.151, +0.091]
EMA − fixed (bottleneck)	+0.035	[−0.071, +0.143]	+0.358	[+0.270, +0.449]
Feature × supervision	+0.037	[−0.090, +0.170]	−0.374	[−0.562, −0.229]

Each task receives equal weight, and the 95% intervals come from 3,000 task-bootstrap resamples at seed 42; negative differences favor the first named configuration. Decoder gains remain clear in both noise ranges, while EMA-versus-fixed differences are much smaller.

**Figure 16 · Feature and supervision effects**The error bars compare paired Brier differences across held-out tasks. Decoder features show a much larger benefit than the difference between EMA and tuned fixed supervision; the two panels use different horizontal scales.

Appendix B.5 · Confidence versus random selection across data budgets, with the same training steps.

**Figure 21 · Nine metrics across data budgets**Confidence and random selection receive the same 750 updates, with the full-pool model shown as a reference. More data does not consistently improve every metric under this short schedule, and confidence’s broadest advantage appears at 10%.

Table 13 · Budget sensitivity under 750 updates
Budget	Selection	Episodes	Reconstruction ↑	Scene ↑	Semantics ↑	Motion ↑	Train GPU-h
10%	Random	1,824	0.6341	0.8772	0.5007	0.1701	4.47
10%	Confidence	1,824	0.6995	0.9088	0.5900	0.2502	4.45
20%	Random	3,649	0.6679	0.9003	0.5391	0.2360	4.45
20%	Confidence	3,649	0.6642	0.8812	0.6167	0.1970	4.44
40%	Random	7,298	0.6874	0.8786	0.5576	0.2004	4.44
40%	Confidence	7,298	0.6611	0.8981	0.5987	0.2305	4.47
100%	Full pool	18,244	0.6902	0.9030	0.5263	0.2023	4.46

With the same 750 updates, confidence beats random on all four aggregate scores at the 10% budget; larger budgets bring metric-dependent trade-offs. Training costs stay around 4.44–4.47 GPU-hours, so selecting fewer episodes does not reduce compute in this experiment.

Bold marks the higher score within each confidence/random pair. Reconstruction, Scene and Motion use 62 episodes; CLIP/BLEU use 56, and Semantics averages the available component means.

Table 14 · Confidence minus random · paired mean [95% CI]
Budget	Δ Reconstruction ↑	Δ Scene ↑	Δ Semantics ↑	Δ Motion ↑
10%	+0.0655 [0.0534, 0.0772]	+0.0316 [0.0229, 0.0396]	+0.0914 [0.0247, 0.1563]	+0.0801 [0.0293, 0.1391]
20%	−0.0037 [−0.0186, 0.0112]	−0.0190 [−0.0296, −0.0089]	+0.0768 [0.0119, 0.1425]	−0.0389 [−0.1023, 0.0185]
40%	−0.0263 [−0.0411, −0.0120]	+0.0195 [0.0031, 0.0362]	+0.0440 [−0.0164, 0.1021]	+0.0301 [−0.0570, 0.1390]

Paired episode comparisons show confidence’s clearest advantage at 10%: all four intervals lie above zero. At 20% and 40%, some intervals cross zero or favor random, so the gain is not uniform across budgets and metrics.

10,000 episode-bootstrap resamples. Reconstruction, Scene and Motion use 62 common episodes; Semantics uses the 56-episode intersection, so paired differences can differ slightly from subtracting Table 13.

07 · Models & Data

Everything needed to reproduce the loop.

Model weights, confidence probes, trajectory detector, prescreen outputs, v2 scoring artifacts, and YOLO annotations.

Models

Planned public checkpoints.

EVAC · Warmup v1RoboTwin2.0 domain-adapted warmup

🤗 Download

EVAC-v2 · Weighting NoneMean-risk selection-only checkpoint

🤗 Download

EVAC-v2 · FrameConfidence-guided frame weighting

🤗 Download

EVAC-v2 · Frame + PatchDense confidence-guided enhancement

🤗 Download

Confidence Probe · RoboTwin2.0Main probe used in the paper

🤗 Download

Confidence Probe · AgiBot WorldAdditional confidence checkpoint

🤗 Download

YOLO · RoboTwin2.0Trajectory-metric detector for EWMBench-style evaluation

🤗 Download

Data & Evaluation Artifacts

Precomputed outputs that avoid expensive repeated inference.

50-task prescreen packageEVAC-v1 inference, confidence scores, risk maps, and JSON metadata

🤗 Download

EVAC-v2 training · inference + JSONPrecomputed v1 inference outputs + episode metadata

🤗 Download

EVAC-v2 training · dense confidence + JSONDense confidence/risk outputs and metadata for confidence-guided retraining

🤗 Download

Baseline selection · v1 inference resultsEVAC-v1 inference outputs for tasks/scenes selected by acquisition baselines

🤗 Download

Baseline weighting · v2 frame-scoring dataFrame-level scoring artifacts used by the additional-weighting baseline experiments

🤗 Download

YOLO RoboTwin2.0 annotationsRobot-arm trajectory labels estimated from RoboTwin2.0 action conditions

🤗 Download

Evaluation tables & bootstrap JSONMean / seed-wise metrics and pooled paired-bootstrap statistics

🤗 Download

08 · Citation

Cite ConfAL-WM.

The copy button writes the BibTeX block directly to the clipboard.

@article{confalwm2026,
  title   = {ConfAL-WM: Confidence-Guided Active Learning for Action-Conditioned World Models},
  author  = {Anonymous Authors},
  journal = {arXiv preprint},
  year    = {2026},
  url     = {https://ConfAL-WM.github.io}
}

% TODO: replace author / arXiv identifier / venue when public.