Ablate ADM retrieval and confidence parameters

Investigate the individual effects of the fixed retrieval capacity, source-reliability weights, confidence threshold, and recency-decay threshold in the attention-enhanced dynamic-memory planner, whose separate contributions remain unresolved and require calibration.

Background

The attention-enhanced dynamic-memory planner uses a hand-designed confidence–distance heuristic to select a bounded set of active object records for the language-model planner. The evaluation reports strong aggregate performance, but the authors note that the heuristic can suppress a distant entity that may be important to a future subgoal.

The planner also depends on several fixed hyperparameters: the retrieval capacity K, source-reliability weights, confidence thresholds, and recency-decay thresholds. The paper does not isolate the contribution of each parameter, leaving their individual effects unresolved and identifying their calibration as an open empirical problem.