Hierarchical Dual-Graph Retrieval for Long-form Audio Reasoning
Organizations: Division of Artificial Intelligence & Software, Ewha Womans University
Abstract
Large audio language models (LALMs) struggle with long-form audio reasoning, as processing entire recordings often limits to capture query-relevant content distributed across distant segments while incurring substantial computational cost. We introduce AudioHelix, a hierarchical DualGraph framework tailored for LALMs' retrieval-augmented generation (RAG) systems. Unlike existing text-only or non-graph RAG approaches, AudioHelix connects related textual and acoustic evidence across temporally distant segments for multi-hop reasoning. Extensive experiments on TraceAV-Bench, MMAU-Pro, and AudioMarathon demonstrate that AudioHelix consistently outperforms full-audio processing and existing RAG methods. The project site is https://github.com/ewha-mmai/audiohelix.
Figures & tables
| Method | TraceAV | MMAU-Pro | AM |
| Audio-Flamingo-3 [ 2 ] | 56.9 | 50.7 | 57.3 |
| + LinearRAG [ 26 ] | 54.4 | 48.0 | 56.1 |
| + PlanRAG-Audio [ 13 ] | 37.5 | 34.0 | 30.8 |
| + AudioHelix | 59.4 | 52.3 | 58.5 |
| Voxtral-3B [ 3 ] | 48.1 | 40.9 | 56.4 |
| + LinearRAG [ 26 ] | 38.9 | 53.1 | 52.3 |
| Method | TraceAV | MMAU-Pro | AM |
| Audio-Flamingo-3 | 56.9 | 50.7 | 57.3 |
| + AudioHelix (w/o Text) | 55.3 | 48.2 | 54.4 |
| + AudioHelix (w/o Audio) | 58.8 | 50.7 | 56.5 |
| + NaïveRAG | 58.1 | 58.0 | 62.6 |
| + LinearRAG (w/ Our text) | 60.0 | 50.0 | 54.7 |
| + AudioHelix | 59.4 | 52.3 | 58.5 |
| Method | Indexing (s) | Latency (s) | TFLOPs |
| Full Audio | - | 17.52 | |
| LinearRAG [ 26 ] | 169.77 | 2.57 | |
| PlanRAG-Audio [ 13 ] | 476.99 | 18.18 | |
| AudioHelix | 391.45 | 1.58 |
Audio Reasoning
Audio reasoning studies how AI systems infer events, relationships, intent, and context from sound, beyond recognizing words or matching labels.
See all 71 papers