MIST shows video world models don't forget; they fail to carry out the command

A low score for a video world model on a memory test is usually read as forgetfulness. The authors of "Execution Before Memory" say that conclusion is wrong. They propose MIST, an evaluation that separates scene setup, execution of the requested change, and the outcome in the final frame. Their diagnosis is different: the model builds the scene, but does not perform the action it was asked to perform. The M3-Agent (Make the Missing Move) method refines the instruction and more than triples the MIST reveal score. On the WRBench benchmark, metric T rises 26.1%, and on MBench-T, metric U rises 52.7%.

One score for three questions

Memory tests for video models usually work like this. The model is shown a scene and asked to change something, and the final frame is then graded with a single number. That number blends three different things: whether the scene was set up, whether the change was carried out, and whether the result is visible at the end.

The flaw is logical. If the change never happens, there is nothing for a hidden state transition to come from. There is then nothing to test: memory did not fail, it was never engaged. The score comes out low, and the reader chalks it up to forgetfulness, even though the cause sits earlier in the chain.

How MIST works

MIST separates the change from the memory it is supposed to leave behind. The model first sets up the scene while the change stays hidden. Only then is it asked to reveal it. Leakage of information about the hidden change before that point is kept under control, since otherwise the result cannot be interpreted unambiguously.

This separation shows exactly where the process breaks. According to the authors' results, the model handles scene setup but does not carry out the instructed change. They find the same gap in general memory evaluations, so the problem is not confined to one special-purpose test.

M3-Agent: fix the instruction, not the model

Building on this, the authors propose M3-Agent. It recursively revises the details of the input instruction until the model does what was asked. The base model stays frozen throughout. The method attacks the problem from the execution side: it shows what the model can do when the required action is spelled out explicitly.

Reported results:

• MIST reveal score: more than three times the baseline;
• WRBench, metric T: +26.1%;
• MBench-T, metric U: +52.7%;
• model weights do not change; the work happens at the instruction level.

The abstract gives no absolute values and no list of tested models. The relative gains need to be checked against the full text of the paper. The authors also describe the result as a stronger baseline for future work: any method that claims to improve memory now has to beat simple instruction refinement.

How to use this in practice

Teams that evaluate video world models should revisit their protocol. Before concluding that memory is poor, it helps to check whether the change itself was carried out, that is, whether it appears in the frame at all. If it does not, the instruction or the execution needs work, not the memory architecture.

For those building products on video generation, the takeaway is more down-to-earth. Recursive prompt refinement, as in M3-Agent, is cheap to layer on top of an existing model without fine-tuning. Researchers, meanwhile, will have to compare new memory methods against this tougher baseline. Otherwise, some of the claimed improvements will turn out to be improvements in execution, recorded as improvements in memory.