Reading the Room: Implicit Confusion Encoding in Recurrent World Model States

Donald Aadithiyan

cs.LGAug 21, 2026

Reading the Room: Implicit Confusion Encoding in Recurrent World Model States

Authors: Donald Aadithiyan

Organizations: Department of Computer Science and Engineering, University of Moratuwa, Moratuwa, Sri Lanka

Abstract

World models built on the RSSM architecture, such as DreamerV3, keep a recurrent hidden state $h_t$ trained only to reduce prediction error. We show this state also tracks its own confusion, hiding in plain sight: nearly orthogonal to $h_t$ 's directions of greatest variance, invisible to any variance-based method. It is functionally distinct from ensemble disagreement, which flags new inputs, and reconstruction error, which flags bad predictions right now. On a test holding prediction error fixed while confusion varies, a linear probe on $h_t$ finds the signal (AUROC 0.72, 5 runs), while an ensemble baseline scores below chance. A discounted count of recent high-error steps explains 80% of the probe's output ( $R^2=0.80$ ). We confirm the signal is causally used, not merely present, by editing $h_t$ directly and watching behaviour change, including a check using real values from other trajectories instead of synthetic edits. Its geometry and closed form generalize across three control tasks; the decisive dissociation test itself holds cleanly on only one, and its practical use, deciding when to check reality instead of trusting imagination, generalizes to only two of the three tasks.

Figures & tables

Figure 1: The overall pipeline: train once, build the three evaluation sets, fit the probe, test it causally against an empirical null, then repeat across tasks and seeds and check two operational uses.

Method	Set A	Set C	Direct novelty
Probe on $h_{t}$	0.863	0.723	0.49
Ensemble (5 models)	0.868	0.550	0.966
Reconstruction error	—	—	0.996

Table 1: Confusion and novelty are dissociated. Ensemble and probe are scored on the same Set C construction throughout.

Method	$\Delta$ probe score	Null percentile
Ablation (synthetic edit)	$-0.586$	100th ( $z\approx-22$ )
Real-value substitution	$-0.761$	100th

Table 2: Two structurally different causal edits agree.

	Cartpole	Reacher	Pendulum
Geometry angle	$88.0^{\circ}$	$89.4^{\circ}$	$88.1^{\circ}$
$C_{t}$ $R^{2}$	0.80	0.26	0.86
Set C AUROC	0.72	0.62–0.67	0.32–0.40
Within-bin correlation	$+0.39$	$-0.09$	$\mathbf{-0.12}$

Table 3: Cross-task comparison. Reacher and pendulum entries are ranges across 4 seeds each.

Figure 2: The routing decision made at every step: threshold the probe’s confusion score to decide whether to query a real observation or continue trusting imagination. Resulting recall against a reconstruction-error baseline, per task, is in Table 4 .

Task	Recall gain vs. recon-error baseline
Cartpole	$+0.04$
Pendulum	$\mathbf{+0.30}$
Reacher	$-0.08$

Table 4: Routing recall gain at a 30% query budget.

Appendix figures & tables16 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Value
Env steps per model	100,000
Sequence length / batch size	16 / 8
Learning rate / grad clip	3e-4 / 1.0
Replay capacity	600 episodes
Episode max steps	500
OOD noise $\sigma$ (Set B)	0.1

Appendix

Table 5: Selected training hyperparameters; full list in the companion supplement.

Figure 3: Set A (held-out, in-distribution) ROC curve, AUROC 0.872.

Figure 4: Set B (near-OOD, Gaussian-noised) ROC curve, AUROC 0.807.

Figure 5: Set C (the KL-matched contrastive test) ROC curve, AUROC 0.714.

Figure 6: Set C AUROC per seed (points) with mean and range (bar and whisker), for each of the three tasks. Individual seed values are shown rather than only a bar, since a bar with error bars can visually imply more precision than 4–5 points actually support. Pendulum’s inversion below chance is visible directly, not just in the mean.

Figure 7: The probing and causal-test setup summarised (Section 4 .3).

Look-ahead $k$	Observed $\|\Delta_{k}\|/\|\Delta_{0}\|$	Predicted $\gamma^{k}$ ( $\gamma=0.95$ )
0	1.00	1.00
1	0.81	0.95
5	0.60	0.77
10	0.50	0.60

Appendix

Table 6: Ablation decay versus the closed-form prediction, the same

\gamma

fitted independently in Section 4 .2.

Figure 8: Ablation effect against the 50-direction empirical null at look-ahead $k=0$ (the immediate effect). The confusion direction (marked, $\Delta_{0}=-0.586$ ) sits at the 100th percentile of the null, $z\approx-22$ .

Figure 9: Ablation effect against the 50-direction empirical null at look-ahead $k=10$ . The confusion direction (marked) sits at the 100th percentile of the null, $z\approx-13$ , showing the effect persists but decays with $k$ .

Gate $z$	Full probe	$\\|h\\|$ only	$1-z$
0.50	1.000	0.971	0.50
0.70	1.000	0.888	0.30
0.90	1.000	0.622	0.10
0.99	1.000	0.508	0.01

Appendix

Table 7: Forced gate value vs. AUROC of the full probe and the magnitude-only (

\|h\|

) separator, against the gate’s overwrite fraction

1-z

. The magnitude-only separator collapses toward chance as overwriting vanishes; the full probe does not.

Task	$r(\mathrm{recon},C_{t})$	Best kinematic $r$	Dominant driver
Cartpole	$+0.40$	$+0.50$ (pole $\omega$ )	kinematic
Reacher	$+0.40$	$+0.33$ (velocity)	$C_{t}$
Pendulum	$+0.33$	$+0.27$ (upright)	$C_{t}$

Appendix

Table 8: Reconstruction error correlates more with

C_{t}

than with the best available kinematic feature on the two tasks where it matters; cartpole is the one task where kinematics dominate recon, and its Set C test works fine regardless.

Figure 10: Cartpole: reconstruction error against $C_{t}$ within a single KL bin. The within-bin trend is positive, the recon-based Set C label agrees with confusion.

Figure 11: Pendulum: reconstruction error against $C_{t}$ within a single KL bin. The within-bin trend is negative, the sign flip that inverts Set C despite pendulum having the strongest $C_{t}$ encoding of the three tasks.

Quantity	deter=256 (XS)	deter=512
Null-space angle	$\sim 88^{\circ}$	$89.4^{\circ}$
Variance in top-10 PCs	$\sim 0.5\%$	$0.13\%$
Ablation $\Delta$ (confusion, $k=0$ )	$-0.586$	$-0.287$
Ablation $\Delta$ (random, $k=0$ )	—	$-0.050$

Appendix

Table 9: Both load-bearing findings hold at double the width: the direction stays near-orthogonal to the top PCs, and ablating it collapses the probe far beyond random.

Task	Probe-A	$C_{t}$ -direct	KL (prior)	vs. recon
Cartpole	0.682	0.664	0.721	$+0.040$
Reacher	0.511	0.411	0.259	$-0.077$
Pendulum	0.778	0.744	0.808	$+0.297$

Appendix

Table 10: Recall at a 30% query budget, single held-out split (

N=40{,}000

/task). ”KL (prior)” is the KL-only router using

\mathrm{KL}_{t-1}

; ”vs. recon” is Probe-A’s gap against the reconstruction-error oracle.

Task (seeds)	Probe-A	KL (prior)	$\Delta$ recall
Cartpole (5)	$0.626\pm 0.047$	$0.632\pm 0.068$	$-0.006\pm 0.025$
Reacher (3)	$0.539\pm 0.013$	$0.267\pm 0.007$	$\mathbf{+0.273\pm 0.020}$
Pendulum (3)	$0.729\pm 0.017$	$0.760\pm 0.023$	$-0.030\pm 0.006$

Appendix

Table 11: The probe’s advantage (mean

\pm

sd) over a KL threshold, repeated across every existing multi-seed collection. Only reacher shows a reliable, sign-stable win.

World Model Learning

World model learning builds predictive models of how environments change under actions to support planning, simulation, and decision-making.

See all 419 papers

MEND: Label-Free Detection, Localisation, and Correction of Latent Hallucination in World Models
Sep 30, 2026Ali J Alrasheed, Aryan Yazdan Parast, Basim Azam +2Hallucination Detection Latent World Models
University of Melbourne Melbourne, Australia · Monash University Melbourne, Australia
Latent State Design for World Models under Sufficiency Constraints
May 3, 2026Keon Woo KimRepresentation Learning Predictive Representation Learning
Om Labs
Lucid Dreaming for World Models: Learning to Doubt Imagination and Decide by Trust
Sep 29, 2026Ziqi Wen, Ting Xu, Lianyu Wang +5Uncertainty Quantification World Model-Based Planning
Centre for Innovation and Precision Eye Health, Yong Loo Lin School of Medicine, National University of Singapore, Singapore 119228, Singapore · Department of Ophthalmology, Yong Loo Lin School of Medicine, National University of Singapore, Singapore 119228, Singapore · Bioengineering Program, Biomedical Sciences Division (BioMed), King Abdullah University of Science and Technology (KAUST), Thuwal, Saudi Arabia +3

Sep 30, 2026J/K move · Enter open · S save

MEND: Label-Free Detection, Localisation, and Correction of Latent Hallucination in World Models

Ali J Alrasheed, Aryan Yazdan Parast, Basim Azam, James Bailey+1

University of Melbourne Melbourne, Australia · Monash University Melbourne, Australia

Hallucination Detection Latent World Models Hallucination Mitigation

Open paper AskPDF arXiv

World Models are appearing as the next major frontier in computer vision. However, their robustness is currently largely unexplored. We identify the phenomenon of hallucination in latent World Models: given a state and an action, the predicted next latent can decode to a scene that never occurs. Because the prediction is statistically ordinary and is fed back autoregressively by the model, the error is both silent and compounding. We study whether such latent hallucination can be detected, localised, and corrected at inference time, on a frozen self-supervised world model in the absence of ground-truth error labels. We introduce Masked Empirical-Bayes Neural Denoising (MEND), a single conditional score network trained by denoising score matching on real transitions, whose score field serves three roles: its magnitude detects hallucination, its per-token field localises it to specific image patches, and it defines an inference-time correction direction. On two navigation environments MEND detects hallucination with an AUROC of up to 0.80 without using actions, exceeding a single-Gaussian density baseline while also localising the error (per-token AUPRC up to 0.87) and correcting it, all from one score field. Our correction reliably reduces single-step latent error and improves predictions. We identify that a part of the error is tangent to the data manifold, hence, we focus on detection and localisation while highlighting promises of the correction.

Abstract

World Model Learning

Explore similar work

MEND: Label-Free Detection, Localisation, and Correction of Latent Hallucination in World Models

Latent State Design for World Models under Sufficiency Constraints

Lucid Dreaming for World Models: Learning to Doubt Imagination and Decide by Trust