Latent Dynamics · 2026-10-01

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

Ali J Alrasheed, Aryan Yazdan Parast, Basim Azam, James Bailey, Naveed Akhtar

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors identify latent hallucination in latent world models: a predicted next latent can decode to a scene that never occurs, with error that is silent and compounds in autoregressive rollouts. They propose MEND, a label-free inference-time detect–localise–correct loop using a single conditional score network trained by denoising score matching on real transitions. On Wall and PointMaze, MEND detects hallucination (AUROC up to 0.80) and localises it (per-token AUPRC up to 0.87), and correction reduces latent error.

Why it matters

The paper targets robustness of latent world models by defining and measuring hallucination in latent space and proposing an inference-time mitigation method that does not require ground-truth error labels or retraining. A single differentiable score field is used for detection, per-token localisation, and correction, enabling a unified framework for handling latent hallucinations when collecting and retraining on failures is undesirable.

Method

Define latent hallucination as predicted next latents that lie outside the set reachable under the true transition dynamics, including unreachable-state (S) and incorrect-action-outcome (A) types.
Introduce a generic detect–localise–correct loop using a differentiable validity score; instantiate it label-free with MEND: a conditional score network trained by denoising score matching on real transitions, whose score magnitude detects, per-token field localises, and score direction corrects at inference time.
Apply masked or full latent correction using Tweedie’s formula via iterative refinement; detection and localisation are produced from one score field evaluation.

Limitation

The authors state that correction is bounded by a component of error tangent to the data manifold, which is indistinguishable from valid alternatives; they also note their approach assumes hallucinations lie off-manifold for the navigation environments studied.

Abstract (from 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.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-09-30 · Indexed by Pub-AI: 2026-10-01 UTC

preprint