Confidence-Conditioned Slot Composition for Structured Retrieval under Noisy Grounding

Reliability-weighted fusion has been widely used, from products of experts (Hinton, 2002) and gated multimodal units (Arevalo et al., 2017) to trusted multi-view classification (Han et al., 2021). Yet, how per-slot confidence should enter a structured retrieval query, and whether it should be used during training, is rarely examined. In this work, we systematically study CCSC (Confidence-Conditioned Slot Composition) for intent-to-knowledge-graph retrieval, comparing eight alternative composers over 37 matched runs on five benchmarks built from BINS, CoDEx-M, WN18RR, and FB15k-237. Synthetic confidence that perfectly separates correct from wrong slots isolates how confidence is used from how well it is estimated. Our central finding is that a simple design, CCSC, which adds slot confidence as a log-bias to query-side attention logits, consistently improves retrieval: it achieves higher mean reciprocal rank (MRR) than role-wise matching, pooling, SumMLP, and learned scalar weighting on all five benchmarks, and than a model trained without confidence on all five when both receive the same confidence. Ablations trace the gain to two factors: (1) training with confidence in the attention logits, since a confidence-only variant already recovers most of the gain, and (2) learned content scores, which help further when correct and wrong slots are mixed. Notably, we find that a model trained with confidence should also receive it at deployment, as removing it reverses its advantage over the confidence-free model on four benchmarks. We release code and checkpoints to facilitate future research.