Skip to content
Safety · Nov 3, 2026

Paper proposes benchmark for safety sufficiency in web world models, finds plausible prediction alone insufficient for robust authorization

Evaluation of ten open models shows 86.1% of predictions are semantically plausible, yet authorization-sensitive decisions remain weak; authors introduce SafeWebWorld benchmark and Latent Authorisation Recovery method.

Trust79
HypeLow hype

1 source · cross-referenced

ShareXLinkedInEmail
TL;DR
  • A human evaluation of 3,000 predictions from ten frozen open models found 86.1% were semantically plausible, but authorization behavior was asymmetric and weak.

Web world models are increasingly used to predict the consequences of candidate actions before agents act, but the authors argue that a plausible predicted future does not necessarily support the correct safety decision. In a matched human evaluation of 3,000 predictions from ten frozen open models, 86.1% were judged semantically plausible.

The study reports that several models exhibit highly asymmetric authorization behavior, combining low false-safe execution with poor recall of clarifiable ASK cases and a strong preference for BLOCK. This motivates the concept of safety sufficiency: whether a predictive interface preserves and exposes the policy-decisive information required to distinguish actions that may be EXECUTED, require clarification through ASK, or should be BLOCKED.

The authors introduce SafeWebWorld, a benchmark for evaluating safety sufficiency across five web-safety domains using action-conditioned futures, policy-decisive evidence, and policy-conditioned decisions. Under a controlled Security-Aligned Supervised Fine-Tuning protocol, evidence preservation can improve substantially while authorization-sensitive decisions deteriorate.

To diagnose whether deterioration reflects loss of authorization-sensitive score structure or a failure of the decision readout, the authors propose Latent Authorisation Recovery (LAR). Frozen-score recalibration yields substantial but backbone-dependent recovery, while paired recovery provides only a modest additional gain. Within-family and matched-authorisation evaluations remain poor, indicating that aggregate recoverability does not imply robust fine-grained authorization behavior.

Overall, the paper concludes that plausible prediction, evidence preservation, and aggregate recovery are each insufficient on their own to establish robust authorization.

Sources
  1. 01arXiv cs.AI — Beyond Plausible Prediction: Safety Sufficiency in Web World Models
Also on Safety

Stories may contain errors. Dispatch is assembled with AI assistance and curated by human editors; despite the trust-score filter, mistakes happen. We correct publicly — every article links to its revision history. Nothing here is financial, legal, or medical advice. Verify before relying on any claim.

© 2026 Dispatch. No ads. No sponsorships. No paid placement. Reader-supported via Ko-fi.

Built by a person who cares about honest AI news.