Where World Models Break: Natural-Input Failure Discovery

• URL: http://arxiv.org/abs/2608.22421v1
• Date: Sun, 23 Aug 2026 13:53:15 GMT
• Title: Where World Models Break: Natural-Input Failure Discovery
• Abstract summary: We formalize the natural-input failure discovery problem, finding environment-valid conditions and action prefixes that induce severe prediction risk.<n>We propose BasinLens, which exploits the underlying structure of valid inputs, by pairing uncertainty-guided global search with typed local replacements.
• Score: 13.330611636170103
• License: http://creativecommons.org/licenses/by-sa/4.0/
• Abstract: World models predict action-conditioned futures and serve as critical internal simulators for downstream planning and control. However, catastrophic prediction failures of world models could dangerously propagate through the control pipeline, as subsequent agent or model training and decision-making depend heavily on the continuous environment evolution forecasted by these world models. Existing evaluations overlook this systemic risk: by aggregating average errors over benign generations from general queries, they fail to stress-test the model against catastrophic collapses under rare or unobserved condition-action combinations. To bridge this gap, we formalize the natural-input failure discovery problem: under a finite query budget, finding environment-valid conditions and action prefixes that induce severe prediction risk, verifying whether these failures reproduce on fresh seeds, and testing their persistence under nearby valid edits. Discovering such critical failures is computationally challenging, as valid condition-action combinations explode exponentially, rendering exhaustive search or standard sampling infeasible given the high cost of noisy rollouts. To tackle this, we propose BasinLens, which exploits the underlying structure of valid inputs, where each coordinate possesses environment-defined semantic types and admissible domains, by pairing uncertainty-guided global search with typed local replacements. Across diverse benchmarks and world-model families, BasinLens exposes reproducible and locally persistent failure modes that conventional evaluations fail to reveal, showing that average-case benchmarks can mask important vulnerabilities in world-model-driven control.

Related papers

• FMMO: Detecting the Divergence Between Local Attribution and Global Drift [2.007262412327553]
  Post-deployment drift poses a critical risk to algorithmic accountability.<n>We demonstrate that popular local attribution methods can exhibit misleading stability even as model reliability collapses.<n>We propose a Framework for Model Monitoring and Observability (FMMO) designed to expose the divergence between local explanation stability and global distribution shifts.
  arXiv  Detail & Related papers  (2026-09-05T16:25:59Z)
• Covariance-Boosted Gaussian Processes for Spatiotemporal Irregularities [0.0]
  Motivated by ionospheric modeling for satellite-based augmentation systems, this paper proposes a Covariance-Boosted Gaussian Process (CBGP) framework.<n>CBGP model efficacy and robustness are demonstrated through out-of-sample testing of both simulated and real-world applications.<n>The modeling of an extensive ionospheric storm dataset over South America suggests accurate and reliable means to compute SBAS ionospheric corrections.
  arXiv  Detail & Related papers  (2026-07-25T03:24:04Z)
• Estimating Rare Events in Language Models with Proper Evaluation [16.1901795614153]
  We introduce Gradient Activation Adaptive Multi-Level Splitting (GA-AMLS)<n>GA-AMLS adapts rare-event Monte Carlo methods to the continuous activation space of language models.<n>We also propose the Shifted-Power Bregman (SPB) Loss, a proper scoring rule that remains finite for zero-estimates.
  arXiv  Detail & Related papers  (2026-07-20T19:07:43Z)
• Foundational World Models Accurately Detect Bimanual Manipulator Failures [21.93685012734004]
  We train a probabilistic, history informed, world model within the compressed latent space of a pretrained vision foundation model.<n>The model outputs uncertainty estimates alongside its predictions that serve as non-conformity scores within a conformal prediction framework.<n>We show that our approach requires approximately one twentieth of the trainable parameters as the next-best learning-based approach.
  arXiv  Detail & Related papers  (2026-03-07T02:11:29Z)
• ProbeLLM: Automating Principled Diagnosis of LLM Failures [89.44131968886184]
  We propose ProbeLLM, a benchmark-agnostic automated probing framework that elevates weakness discovery from individual failures to structured failure modes.<n>By restricting probing to verifiable test cases and leveraging tool-augmented generation and verification, ProbeLLM grounds failure discovery in reliable evidence.
  arXiv  Detail & Related papers  (2026-02-13T14:33:13Z)
• Character as a Latent Variable in Large Language Models: A Mechanistic Account of Emergent Misalignment and Conditional Safety Failures [70.48661957773449]
  Emergent Misalignment refers to a failure mode in which fine-tuning large language models on narrowly scoped data induces broadly misaligned behavior.<n>Across multiple domains and model families, we find that fine-tuning models on data exhibiting specific character-level dispositions induces substantially stronger and more transferable misalignment than incorrect-advice fine-tuning.
  arXiv  Detail & Related papers  (2026-01-30T15:28:42Z)
• Revisiting Multivariate Time Series Forecasting with Missing Values [65.30332997607141]
  Missing values are common in real-world time series.<n>Current approaches have developed an imputation-then-prediction framework that uses imputation modules to fill in missing values, followed by forecasting on the imputed data.<n>This framework overlooks a critical issue: there is no ground truth for the missing values, making the imputation process susceptible to errors that can degrade prediction accuracy.<n>We introduce Consistency-Regularized Information Bottleneck (CRIB), a novel framework built on the Information Bottleneck principle.
  arXiv  Detail & Related papers  (2025-09-27T20:57:48Z)
• TrustLoRA: Low-Rank Adaptation for Failure Detection under Out-of-distribution Data [62.22804234013273]
  We propose a simple failure detection framework to unify and facilitate classification with rejection under both covariate and semantic shifts.<n>Our key insight is that by separating and consolidating failure-specific reliability knowledge with low-rank adapters, we can enhance the failure detection ability effectively and flexibly.
  arXiv  Detail & Related papers  (2025-04-20T09:20:55Z)
• Benchmarking the Spatial Robustness of DNNs via Natural and Adversarial Localized Corruptions [49.546479320670464]
  This paper introduces specialized metrics for benchmarking the spatial robustness of segmentation models.<n>We propose region-aware multi-attack adversarial analysis, a method that enables a deeper understanding of model robustness.<n>The results reveal that models respond to these two types of threats differently.
  arXiv  Detail & Related papers  (2025-04-02T11:37:39Z)
• Reshaping the Online Data Buffering and Organizing Mechanism for Continual Test-Time Adaptation [49.53202761595912]
  Continual Test-Time Adaptation involves adapting a pre-trained source model to continually changing unsupervised target domains. We analyze the challenges of this task: online environment, unsupervised nature, and the risks of error accumulation and catastrophic forgetting. We propose an uncertainty-aware buffering approach to identify and aggregate significant samples with high certainty from the unsupervised, single-pass data stream.
  arXiv  Detail & Related papers  (2024-07-12T15:48:40Z)
• Condition Monitoring with Incomplete Data: An Integrated Variational Autoencoder and Distance Metric Framework [2.7898966850590625]
  This paper introduces a new method for fault detection and condition monitoring for unseen data. We use a variational autoencoder to capture the probabilistic distribution of previously seen and new unseen conditions. Faults are detected by establishing a threshold for the health indexes, allowing the model to identify severe, unseen faults with high accuracy, even amidst noise.
  arXiv  Detail & Related papers  (2024-04-08T22:20:23Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.