Same World, Different Knowledge: When Isolated Audits Misjudge World-Model Repairs

Rui Min; Xianyao Li; Fang Xu; Sofiane Lachab; Jing Du

cs.ROSep 17, 2026

Same World, Different Knowledge: When Isolated Audits Misjudge World-Model Repairs

Authors: Rui Min, Xianyao Li, Fang Xu, Sofiane Lachab, Jing Du

Organizations: Department of Mechanical and Aerospace Engineering, University of Florida, Gainesville, FL, USA · Department of Civil and Coastal Engineering, University of Florida, Gainesville, FL, USA · Molinaroli College of Engineering and Computing, University of South Carolina, Columbia, SC, USA

Abstract

A repair favored under an isolated input fault can be inferior when deployed modules share the faulty information. We introduce an information-interface audit for world models, distinguishing fidelity gaps, where exact inputs become estimates, from availability gaps, where inputs are missing. Fixed-weight interventions measure prediction error, input dependence, and paired closed-loop benefit, including dependencies introduced by reconstruction. In simulated quadrotor model predictive control, coupled, opposite-sign 10% mass/thrust calibration errors reduce a physics-anchored model's success from 69% to 8%; uncertainty training restores 65%. Wind reconstruction recovers control benefit but inherits calibration dependence. For a positive calibration offset, reconstruction-only corruption favors uncertainty-trained reconstruction, whereas shared corruption favors the baseline. Acceleration diagnostics reveal compensation between reconstruction bias and nominal-model error, also observed with a disturbance observer. Repair selection therefore depends on the information paths used in deployment.

Figures & tables

Fig. 1: The information-interface audit. (a) Fidelity and availability gaps arise when exact calibration inputs become estimates and wind is not directly measured. A reconstruction module recovers wind from state–action history and also uses calibration inputs. C, E, and F mark corruption, erasure, and freezing; the audit measures prediction error, input dependence, and paired closed-loop benefit. (b) Re-auditing wind reconstruction in its deployed composition (Table IV ) changes the preference: at $e=+0.1$ , reconstruction-only corruption favors DCC-trained reconstruction, while shared corrupted calibration favors baseline reconstruction. Scene images: simulation render (top) and generated deployment illustration (bottom).

Fig. 2: Representative held-out scenes: the first library entry of each evaluation topology. Scene IDs follow the frozen library. Blue circles and orange stars mark start and goal; transparency reveals the geometry.

	Error (cm)		Outcome /100
Model	common	own	Success	Crash	Timeout	Time (s)
DKI	32.8	31.0	68	13	19	4.25
WM-A	8.2	9.5	69	21	10	2.75
WM-U	6.3	7.4	69	22	9	3.00
True-parameter	2.6	2.6	69	16	15	3.20
With ground-truth clearance

TABLE I: Prediction-accuracy comparison on the held-out obstacle library (100 scenes,

v_{\max}

3.4 m/s, learned models use seed 0; success across three seeds is 69–70 for WM-U and 68–71 for WM-A). Error: median 20-step position error on common trajectories collected with the true-parameter model and on each model’s own trajectories. Time: median over successful runs.

Fig. 3: Success/100 versus calibration error

e

on the selection library in free space (seed 0). WM-U+DCC completes all scenes across the tested range; WM-A+DCC broadens the nominal peak but still fails at the most negative offset.

Matched WM-A comparison:
Model	$e{=}{-}0.1$	$e{=}0$	$e{=}{+}0.1$
WM-A	18	69	8
WM-A+DCC	68	68	65
+DCC minus baseline (percentage points):
Difference	$49.7$	$-1.0$	$57.3$
95% paired CI	$[40.0,59.3]$	$[-3.3,1.0]$	$[48.3,66.3]$

TABLE II: Calibration audit on held-out obstacles. Matched comparisons report rounded three-model-seed mean success/100; differences use unrounded means. Paired confidence intervals (CIs) bootstrap scene-wise mean differences, conditional on the trained models. The final panel uses WM-A seed 0: errors are median 20-step position errors on common held-out trajectories collected with the true-parameter model; success uses the same model seed.

Training	Deployment input	L0	L2	L3
Held-out, free space:
wind-withheld	none	100	98	82.7 a
wind-conditioned	zero	100	96	78.7 a
wind-conditioned	$\hat{w}_{t}$ (trajectory GRU)	100	100	96.7 b
wind-conditioned	$\hat{w}_{t}$ (GRU)	100	100	100
wind-conditioned	$w_{t}$ (ground truth)	100	100	100

TABLE III: Held-out success/100 under wind, with and without obstacles. Predictor training and deployment wind inputs are listed separately. Entries without a footnote use seed 0. GRU denotes the collection-trained reference reconstruction module; the trajectory-trained variant is labeled separately. At L0, ground-truth wind is zero; those cells reuse the zero-input runs.

GRU+DCC minus GRU (pp):
Interface	Correct	Reconstruction-only	Shared
$(e_{\mathrm{R}},e_{\mathrm{P}})$	$(0,0)$	$(+0.1,0)$	$(+0.1,+0.1)$
GRU	67.0	49.3	67.0
GRU+DCC	57.3	56.0	56.0
DOB	66	44	63
Difference	$-9.7$	$+6.7$	$-11.0$

TABLE IV: Wind reconstruction on held-out L3 obstacles (100 scenes). GRU and GRU+DCC average three reconstruction seeds; predictor seed 0 and DOB are fixed.

(e_{\mathrm{R}},e_{\mathrm{P}})

: reconstruction and predictor calibration errors. CIs bootstrap scene-wise mean differences; success differences are percentage points (pp). DCC changes reconstruction training.

DCC repair / fault path	$\Delta_{0}$ (pp)	$\Delta_{f}$ (pp)	$p^{\star}$ (%) [95% CI]
WM-A	$-1.0$	$+57.3$	$2\,[0,6]$
WM-U	$-1.3$	$+40.0$	$3\,[0,8]$
GRU: reconstruction only	$-9.7$	$+6.7$	$59\,[33,80]$
GRU: shared with predictor	$-9.7$	$-11.0$	No crossing

TABLE V: DCC decision record on held-out obstacles at

e=+0.1

(predictors: L0; GRU: L3). Success differences are +DCC minus baseline under correct (

\Delta_{0}

) and faulty (

\Delta_{f}

) inputs; their CIs appear in Tables II and IV .

(a) Analytic acceleration error (m/s 2 )
$(e_{\mathrm{R}},e_{\mathrm{P}})$	$(0,0)$	$(+0.1,0)$	$(0,+0.1)$	$(+0.1,+0.1)$
GRU	1.10	1.93	1.84	1.00
GRU+DCC	1.49	1.64	1.89	1.57
DOB	0.54	2.32	2.08	0.53

TABLE VI: Calibration compensation on 100 held-out L3 obstacle scenes. (a) Mean of per-seed window medians over three reconstruction seeds; DOB fixed. The 1,461 common windows were recorded with nominal GRU seed 0. (b) Success/100, reconstruction and predictor seed 0.

World Models

World models represent and predict how environments change, helping agents plan, act, and reason about possible futures.

See all 371 papers

The Intervention Gap in Latent World Models
Aug 30, 2026Donna VakalisWorld Model-Based Planning Model-Based Planning
Mila – Quebec AI Institute · University of Montreal
World Models Dream of Success: Diagnosing and Repairing Failure Insensitivity in Robot World Models
Oct 6, 2026Jiuyi Xu, Xiao Hu, Meida Chen +3World Models for Robotics Action-Conditioned World Models
Colorado School of Mines · Northeastern University · Institute for Creative Technologies, University of Southern California +1
Where World Models Break: Natural-Input Failure Discovery
Aug 23, 2026Zhanpeng Shi, Zi Liang, Rong Feng +3Adversarial Attacks on World Models Uncertainty-Aware Exploration
School of Computer Science, Northwestern Polytechnical University · The Hong Kong Polytechnic University · Department of Data Science, City University of Hong Kong +3

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

The Intervention Gap in Latent World Models

Donna Vakalis

Mila – Quebec AI Institute · University of Montreal

World Model-Based Planning Model-Based Planning Latent World Models World Model Evaluation

Open paper AskPDF arXiv

Planning-time intervention fidelity is a distinct, measurable property of a learned world model: whether the model's own open-loop transitions move task variables the way matched environment interventions do. In the settings we test, it is neither revealed by reward fit nor ensured by task-anchored training. Across released TD-MPC2 checkpoint sizes, episode return falls as an operator-error diagnostic on task observables grows, while reward-prediction error stays small and nearly flat, and a self-supervised world model trained without task signal preserves the same operator substantially better than a task-anchored model on the shared task. A capture-gated matched-intervention audit then localizes what fails. On Cheetah, three LeWorldModel checkpoints capture the current task query and support decodable real intervention effects; however, their imagined five-step effects are worse than predicting no effect and worse than an environment-endpoint oracle. The failure is task-direction rotation with excess gain, not feature collapse. This severe pattern is conditional: five PreJEPA seeds retain an oracle-relative deficit without it, Finger Spin experiments extend the deficit beyond locomotion with heterogeneous severity across seeds, and shared-bank effect geometry is both candidate- and support-dependent. We also test practice-side questions. In DreamerV3 the posterior distribution, not its sample, carries the current query; ensemble disagreement ranks error only near training support; and a frozen support-aware score degrades held-out error ranking in both tested transfer directions while native disagreement remains informative in both. We conclude that intervention fidelity must be audited directly, capture-first, on the model's native interface.

Abstract

World Models

Explore similar work

The Intervention Gap in Latent World Models

World Models Dream of Success: Diagnosing and Repairing Failure Insensitivity in Robot World Models

Where World Models Break: Natural-Input Failure Discovery