HeLM lifts VLA robot success 12.8pp with 4.5x less memory

A new robotics preprint argues that the dominant Vision-Language-Action design is losing performance to a modelling choice almost everyone makes: treating each camera frame as if it carried all the context the robot needs. The authors of HeLM report their fix beats the strongest baseline by 12.8 percentage points on intra-episode success while using up to 4.5x less memory at long horizons.

The paper, posted to arXiv by Gyeonghun Kim, Donghoon Kim, Minsoo Choi and Suyun Lee, targets a specific failure mode. "Most existing VLA architectures operate under a Markovian assumption, mapping only the current observation to an action, which leads to failure in non-Markovian environments," the authors write. HeLM instead maintains a running log in "explicit language through factorized event detection and event-triggered memory updates," an architecture the authors say is inspired by the human "inner monologue."

The evaluation covers "10 non-Markovian real-world tasks with 56 variants, including five newly formulated inter-episode tasks that require persistent context across task boundaries." On compositional generalization the system reaches 78% subtask success. The authors also claim HeLM runs "2.5x faster than the keyframe baseline at long horizons."

The abstract names no base VLA, no robot platform, and no per-task breakdown behind the 12.8-point headline. The comparison is anchored on "the strongest baseline" and "the keyframe baseline" without identifying either.