benchmarks evals
Seeing the Room Is Not Remembering the Room
Ten visual representations could expose visible structure without reliably preserving objects, predicting updates or supporting continuous action.
Summary
Ten visual representations could expose visible structure without reliably preserving objects, predicting updates or supporting continuous action.
Researchers probed ten frozen visual representations across four increasingly demanding capabilities: spatial structure, object persistence, predictive updating and actionability under streaming egocentric views. Strong readout of the visible scene did not guarantee that a model retained objects outside the current view or retrieved useful future features. Even when predicted actions changed with the goal, supplying the correct goal did not reliably improve progress. The work separates decodability from operational spatial state and argues for tests that ask whether information survives motion and guides action, not merely whether a linear probe can find it.
Why it matters
Ten visual representations could expose visible structure without reliably preserving objects, predicting updates or supporting continuous action.
Limits and context
- Strong readout of the visible scene did not guarantee that a model retained objects outside the current view or retrieved useful future features.
- Even when predicted actions changed with the goal, supplying the correct goal did not reliably improve progress.
- The work separates decodability from operational spatial state and argues for tests that ask whether information survives motion and guides action, not merely whether a linear probe can find it.
Key claims
Ten visual representations could expose visible structure without reliably preserving objects, predicting updates or supporting continuous action.
Qualification: Strong readout of the visible scene did not guarantee that a model retained objects outside the current view or retrieved useful future features.
Evidence: source-2026-11-01-006
Sources
- arXiv preprint 2610.40636, original v1arXiv · primary research
Corrections
No corrections have been recorded for this story.