An AI robot-control system that uses a pretrained multimodal large language model (MLLM) as running episode memory reported 60.83% average success on RoboMME at the base data scale and 75.54% with 9x data. Those scores were 16.32 and 17.66 percentage points above FrameSamp+Modul, respectively. The evidence comes from two simulated benchmarks.
The preprint asks whether an MLLM's native causal context can serve as actionable episode memory for robot control through an asynchronous link to a pretrained action model. Its system, PonderPounce, connects the MLLM, called Ponder or System 2, to Pounce or System 1 through recurrent continuous cognition tokens. Ponder retains the evolving context, while Pounce produces the robot's actions.
Two models, trained together
The two components were initialized from pretrained checkpoints and jointly optimized end to end, without a separate bridge-pretraining stage. RoboMME contains 16 tasks. Its base dataset has 1,587 episodes, while the 9x version is a fresh collection of 14,400 episodes. Each RoboMME run evaluated 50 episodes per task, or 800 episodes in all, and the reported PonderPounce results were averaged over three runs.
With the Pounce architecture and interface unchanged, pretrained Qwen3.5-9B reached 60.83% RoboMME success, compared with 50.04% for pretrained Qwen3.5-0.8B, a 10.79-percentage-point difference. This matched-interface comparison is specific to the reported benchmark setup.
Random initialization of the 9B MLLM reached 0.00% RoboMME success, and the run did not converge. The study therefore does not provide a matched-convergence estimate of the benefit of pretraining.
Refresh intervals in the comparison
Refresh timing was compared under a same-checkpoint RoboMME intervention. The every-query refresh condition had 60.83% average success, whereas the transitions-only condition had 1.83%.
A separate diagnostic tracked normalized loss as the carried context aged. For the checkpoint trained with a 1-second refresh, the value was 1.00x at 0.3 seconds, 7.11x at 2.3 seconds and 9.22x at 4.3 seconds. In checkpoints trained with 2- and 4-second refreshes, the 4.3-second value was 3.19x and 1.14x, respectively.
Different conditions, different scores
Other comparisons showed different results across supervision conditions. The full training condition reached 60.83% average RoboMME success. The condition without demonstration-reasoning targets reached 48.21%, while the condition without LM-head grounding reached 27.96%.
On RoboCasa-DC, the second benchmark, PonderPounce averaged 12.5% success, compared with 11.6% for SeeTraceAct. The cognition-disabled condition reached 8.6%, while a no-demonstration control recorded 9.0% in a single run. Evaluation covered five held-out tasks, with 50 episodes per task; PonderPounce and cognition-disabled conditions were averaged over five runs.
What the tests cover
The reported evidence is bounded: the evaluation is limited to two simulated benchmarks and one cognition carrier per query. RoboCasa-DC covers five held-out tasks and is presented as initial interface-transfer evidence rather than broad cross-embodiment generalization. The study does not establish performance beyond these reported conditions.
Optimized serving recorded a p50 cognition-refresh latency of 78 milliseconds and a p50 Pounce invocation latency of 25 milliseconds, supporting 20 Hz action playback.
The work was supported by an IITP grant funded by the Korea government (MSIT), grant RS-2026-25522885. The document is a preprint listed as arXiv:2608.24115v1, dated 25 Aug 2026.
Paper data and sources
Original title: PonderPounce: A Pretrained MLLM as an Episode Context Engine for Robot Control
Authors: Suhwan Choi, Jaeyoon Jung, Sungkyung Kim et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text