# Attention-guided Evidence Grounding for Spoken Question Answering

Source: [Attention-guided Evidence Grounding for Spoken Question Answering](<https://www.sota2.com/research/paper/69ba31830636e5bf4afc075c>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-03-20T05:00:04.947Z

Authors: Ke Yang, Bolin Chen, Yuejie Li, Yueying Hua, Jianhao Nie, Yueping He, Bowen Li, Chengjun Mao

Published: 2026-03-17T09:27:13.000Z

## Abstract

Spoken Question Answering (Spoken QA) presents a challenging cross-modal problem: effectively aligning acoustic queries with textual knowledge while avoiding the latency and error propagation inherent in cascaded ASR-based systems. In this paper, we introduce Attention-guided Evidence Grounding (AEG), a novel end-to-end framework that leverages the internal cross-modal attention of Speech Large Language Models (SpeechLLMs) to explicitly locate and ground key evidence in the model's latent space. To address the diffuse attention distribution in pre-trained models, we propose Learning to Focus on Evidence (LFE), a supervised fine-tuning paradigm that calibrates the model's attention mechanism to distinguish query-relevant segments from irrelevant context. Experiments on SQuAD, HotpotQA, and MuSiQue demonstrate that AEG reduces hallucinations and achieves strong efficiency gains, outperforming large-scale cascaded baselines (Whisper-Large-v3 + Reranker) while reducing inference latency by approximately 62%.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Spoken Question Answering on HotpotQA v1.0 (test)](<https://www.sota2.com/research/sota/spoken-question-answering-on-hotpotqa-v1-0-test.md>): Spoken Question Answering; HotpotQA v1.0 (test); Exact Match (EM): 79.16.

- [Spoken Question Answering on MuSiQue v1.0 (test)](<https://www.sota2.com/research/sota/spoken-question-answering-on-musique-v1-0-test.md>): Spoken Question Answering; MuSiQue v1.0 (test); Exact Match (EM): 53.99.

- [Spoken Question Answering on SQuAD v1.1 (test)](<https://www.sota2.com/research/sota/spoken-question-answering-on-squad-v1-1-test.md>): Spoken Question Answering; SQuAD v1.1 (test); EM: 89.24.

- [Evidence Grounding on SQuAD v1.1 (test)](<https://www.sota2.com/research/sota/evidence-grounding-on-squad-v1-1-test.md>): Evidence Grounding; SQuAD v1.1 (test); F1 Score: 80.02.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2603.16292>)

[PDF](<https://arxiv.org/pdf/2603.16292.pdf>)