Parameter-Efficient Retriever-Reader for Trustworthy Table-Text Question Answering Toward Agentic Edge Intelligence

Abstract

Resource-constrained edge intelligence increasingly requires lightweight and trustworthy access to heterogeneous knowledge, where decisions often involve reasoning over structured tables and associated text. In this paper, trustworthiness is operationalized as robustness under weak supervision, imperfect evidence selection, and noisy hybrid contexts. Table–text Hybrid Question Answering (HQA) is representative of this need, yet existing methods often face a trade-off between specialization-driven systems with high engineering cost and generalization-driven LLM pipelines that are flexible but may lag behind the strongest specialized systems. We propose RRS, a parameter-efficient retriever–reader framework that formulates HQA as two dialog-style tasks, row retrieval and answer generation, handled by a shared pretrained language model with task-specific low-rank adapters. RRS combines modular decomposition, noise-robust training against weak-supervision errors, and a natural-language interface without specialized feature engineering. Sharing over 99.8% of backbone parameters, it introduces no more than 0.2% task-specific trainable parameters. With modest-scale public models of 7B and 4B parameters, RRS achieves 71.9%/79.6% and 71.6%/79.2% EM/F1 on the HybridQA official blind test, surpassing the previous best result of 67.9%/75.5%. These results demonstrate a lightweight and noise-robust table–text reasoning approach that is well suited to resource-constrained deployment and provides a useful building block for agentic edge intelligence.