- Who it helps
- Researchers evaluating embodied agents and multimodal memory.
- Availability
- The October 16 paper and public project explanation are available. This entry explains evaluation design; no agent service or reproduction has been tested.
A staged question
The project separates scene setup, target information, action execution and final state. A plausible answer can be misleading if the requested change never happened.
Execution Before Memory / MIST paper ↗MIST project and evaluation cautions ↗
Account for privileged information
The project flags target leakage and a best-of-five selection process that uses privileged judging. Such selection can overstate what an autonomous system would achieve without that information.
What you can do
- For your own evaluation, retain the observed state before and after the action and score task execution separately from the final answer. Open reference ↗
Where the evidence stops
No benchmark was reproduced. The described results apply to the authors’ evaluation setup.
Sources read. Primary text checked October 19, 2026 UTC. The external product or research workflow was not executed; reported results remain attributed to their authors.
Sources & further reading
Paper · paper
Execution Before Memory / MIST paper ↗
October 16 abstract read; followed the linked project for execution, leakage and selection limitations.
MIST authors · page checked 2026-10-19 · Published 2026-10-16
Web · official
MIST project and evaluation cautions ↗
Project explanation read: scene setup, requested change, final state and answer leakage are separated; best-of-five selection uses privileged judging.
HKU MMLab and collaborators · page checked 2026-10-19
Coverage note: The linked sources establish the documented scope. Unlisted social platforms and independent execution remain unverified.
The record over time (1)
Dated source checks and published explanations. An observation date can differ from the original release date.
- · Added to Current
MIST examines whether agents actually perform a requested scene change before judging later answers about it. This separates failed action from failed memory.