# Want VLA to be more reliable? Is the cost more replanning?

Want VLA to be more reliable? Is the cost more replanning?

Full article: https://haiguangboy.com/posts/chunktrust

## Core Method

The training-free AHS keeps the base weights and action values, and evaluates two types of evidence for candidate prefixes:
↳ Intra-chunk: during flow generation, whether the spectrum is stable relative to the early denoising baseline.
↳ Inter-chunk: after connecting to already executed actions, whether the velocity is irregular. A prediction that looks very stable on its own may still fail to connect with historical motion.

A Beta state with forgetting accumulates within-episode length preferences, mitigating single-instance evidence noise; it resets each episode. Optional QHA offline-learns a conditional prior, with no online updates at deployment.

## Key Results

· On the full 50 tasks of RoboTwin, the same π0.5 improves its success rate from 56.70% to 63.50%.
· On 24 RoboCasa tasks, Qwen3GR00T goes from 47.83% to 57.50%; on the independent eight-task Fast-WAM evaluation group, from 86.81% to 88.13%.
· On four real-robot tasks with 15 trials each, the normalized process score goes from 50.4% to 57.5%, which is not the full success rate.

## Comparison with Related Lines of Work

- Opposite line · [SmoothRL: only settle accounts for actions actually executed](https://haiguangboy.com/posts/smoothrl_online_reinforcement_learning_during_asynchronous_execution) `smoothrl_online_reinforcement_learning_during_asynchronous_execution_2026_09`: SmoothRL: only settle accounts for actions actually executed
- Same line · [Agile-WAM: lightweight visuotactile joint prediction](https://haiguangboy.com/posts/agile-wam) `agile_wam_an_agile_tactile_world_action_model_for_contact_rich_robot_control_2026_09`: evaluation scope and cost accounting must be stated separately validates that small-sample and best-of-n evaluations cannot be extrapolated into stable general capability
- Same line · [DexTouch-WM: human tactile pretraining](https://haiguangboy.com/posts/dextouch-wm) `dextouch_wm_learning_action_conditioned_tactile_world_models_from_human_touch_fo_2026_09`: internal quality and the Beta posterior are not calibrated success probabilities validates that correlation does not equal success rate, nor absolute score calibration
- Same line · [HyperWorld: take it apart, generalize away 31 points](https://haiguangboy.com/posts/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo) `hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo_2026_09`: HyperWorld: take it apart, generalize away 31 points
- Same line · [Introducing Gemini Robotics ER 2](https://haiguangboy.com/posts/google_blog_gemini_robotics_er_2_20260730) `google_blog_gemini_robotics_er_2_20260730_2026_07`: average gains cannot hide regressions, nor can they claim to comprehensively beat other selectors validates that the numbers lack controls: no named competitors, no baseline values, and comparison charts that cannot be verified
- Same line · [OpenWAM: the right answer changes with scale](https://haiguangboy.com/posts/openwam_an_open_modular_exploration_towards_systematic_world_action_model_pretra) `openwam_an_open_modular_exploration_towards_systematic_world_action_model_pretra_2026_09`: OpenWAM: the right answer changes with scale

## Boundaries

⚠️ Balance does not equal optimality, let alone a safety guarantee
Internal quality and the Beta state are not calibrated success probabilities. During execution of the selected prefix there is no new-observation monitoring, and sudden disturbances still require other mechanisms. QHA has task regressions; the above latency experiments use a controlled clock and cannot be directly treated as native deployment time. A low selector overhead also does not mean the whole episode is more compute-efficient.

## Author's Judgment (not part of the paper; a cross-paper synthesis view)

💡 Putting this into the technical lineage of the execution layer
Legato improves action-chunk transitions through training; SmoothRL learns residual corrections for executed actions; ChunkTrust keeps action values and adjusts when feedback is reacquired.

The three address different stages. To judge whether an execution scheme is worth deploying, success rate, feedback timeliness, model calls, and total compute should be considered together, rather than reporting only the selector's millisecond overhead.