Risk-Triggered Search adds selective safety intervention to frozen JEPA world models
A new arXiv paper proposes RTS, a method that lets pre-trained visual world models keep their normal plans but run a local search only when predicted process risk rises.
In plain words
World models are AI systems that predict what will happen after each action, but they can still steer into forbidden zones or collisions. RTS watches the model's own predictions for danger and, only when an alarm fires, runs a search for a safer plan. Safe plans run untouched, so behavior stays the same while safety improves.
What was announced
The authors, writing on arXiv, note that purely visual world models may lack analytical dynamics, structured states, or explicit collision geometry, so the system must predict action-conditioned process risk internally. RTS unrolls future latent prefixes inside a frozen Joint-Embedding Predictive Architecture model and trains a lightweight risk head with process-violation supervision. The same predicted prefix-risk score serves both as the trigger gate and as a hard feasibility gate: plans judged safe execute directly, while alarms start a local Cross-Entropy Method search that minimizes deviation from the nominal plan within the feasible set.
Because naive Euclidean correction can shrink obstacle-directed actions into zero-amplitude stalling, the paper introduces an amplitude-preserving metric that keeps tangential progress while holding a unique zero at the nominal plan. Under unified gating and this condition, the authors report that bypassing safe-state search produces closed-loop trajectories strictly identical to running search at every step. On a process-safety benchmark using the Safe Arrival Rate metric across five tasks, RTS improved SAR over Direct execution by 12.67 to 41.67 percentage points on navigation and articulated control, and by 2.08 and 4.17 points on Cube and PushT boundary cohorts, while matching every-step search on 100% of actions and cutting controller time by 17.71% to 70.41% across 1,968 paired trajectories.
What it does not show
- Evidence is limited to the paper's abstract; the full article, experimental setup and peer review status are not available in the supplied passages.
- Reported gains and timing figures are the authors' own benchmark results and have not been independently confirmed.
For security and IT teams
- Scope
- The method targets frozen JEPA world models that take current and goal observations and output multi-step action plans, evaluated on five benchmark tasks including navigation, articulated control, Cube and PushT.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.