← News

Models & toolsSources: arXiv · 18 Nov 2026

Executable perceptual memory aims to keep world models grounded after contact is lost

A new arXiv paper proposes executable perceptual memory, a world-model framework that carries remembered sights and touches into imagined rollouts, reporting stronger reward prediction and control results on memory benchmarks.

In plain words

Robots and other physical agents need to remember things they saw or touched even after those things are no longer visible. The paper introduces a method that keeps those memories active inside a world model, a system that imagines future outcomes to learn behaviour. In the authors' tests, the memory-aware version predicted rewards better and scored higher on control tasks.

What was announced

The authors describe executable perceptual memory (EPM) as a framework that advances an explicit perceptual state alongside neural latent dynamics during imagination, so remembered facts evolve as the agent's actions change their implications. The evolving read-out of that state informs reward prediction, value estimation and policy learning, with reward inputs aligned to the transitions that produced them and the reward head supervised on open-loop rollouts of recorded actions.

On the Memory Maze 13x13 benchmark, the paper reports a return of 30.37 at 1M interactions versus 23.37 when the same read-out informs only policy and value on a shared perceptual front end, a gain of 7.00 with a reported interval of [3.80, 10.21]. The complete method also improved control on the 9x9 variant and in MiniWorld, and probes of remembered out-of-view arrivals showed stronger reward anticipation in both visual implementations. The authors add that concentration, tactile recovery and real tactile frames illustrate complementary roles of remembered facts in imagined-policy training, motion-dependent decisions and sensory-content retention.

What it does not show

  • The evidence is the paper's abstract only; full experimental setup, baselines and peer-review status cannot be verified from this excerpt.
  • All reported scores and comparisons come from the authors themselves; no independent confirmation is available.

Sources

Source-support review completed 19 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 19 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.