Hierarchical Dual-Graph Retrieval for Long-form Audio Reasoning

Eunsang Lee*,  Taein Ju*,  Jiyoung Lee†
Division of Artificial Intelligence & Software, Ewha Womans University
* Equal contribution · † Corresponding author

Abstract

Large audio language models (LALMs) struggle with long-form audio understanding, as processing entire recordings makes it difficult to focus on query-relevant content distributed across distant segments while incurring substantial computational cost. We introduce AUDIOHELIX, a hierarchical DualGraph retrieval-augmented generation (RAG) framework. Unlike existing unimodal or non-graph RAG approaches, AUDIOHELIX constructs a multimodal DualGraph, that 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 with substantially lower inference-time computational cost.

Qualitative Examples

Mel spectrogram of the full original audio. Blue bands are gold evidence; green ticks are AudioHelix's retrieved 5 s units.
Play the audio or click the spectrogram to seek.

video300

Mel spectrogram of video300 with gold evidence (blue) and AudioHelix's retrieved 5s units (green)
Gold Evidence Ours: retrieved 5 s unit
ORIGINAL AUDIO
Query

As the performance shifts from the large male dancer group to the solo unicyclist and then into the later mixed unicycle/dancer rope and pyramid segments, how does the background music behave across these transitions?

A Each segment (male dancers, solo unicyclist, and later mixed unicycle/dancer acts) uses its own distinct track, with noticeable changes in tempo and instrumentation at every transition.
B The soundtrack starts as full EDM for the male dancers, then drops to a sparse, drum-only percussive section during the solo unicyclist, before adding melodies back in only for the later unicycle pyramids.
C A single, upbeat, driving electronic dance track with synth melodies and strong percussion continues across all of these segments without changing to a different musical style.
✓ Correct
D The music switches from upbeat EDM during the large male dancer group to a slower, more acoustic-style track for the solo unicyclist, then returns to EDM for the mixed unicycle/dancer segments.

video339

Mel spectrogram of video339 with gold evidence (blue) and AudioHelix's retrieved 5s units (green)
Gold Evidence Ours: retrieved 5 s unit
ORIGINAL AUDIO
Query

According to the narrator, how did Miuccia Prada’s early use of nylon eventually lead to the Linea Rossa/Prada Sport concept, and how was that line later positioned within Prada’s collections and strategy?

A Miuccia first reinvented Prada by unveiling waterproof nylon bags, and this nylon innovation later evolved into the high-end, functional Prada Sport line in 1997; decades afterward, its Linea Rossa identity was folded back into the Fall/Winter 2018 collections and explicitly cited as a key ‘reintroduction’ in Prada’s future-focused investment strategy.
✓ Correct
B Miuccia’s early nylon accessories directly inspired the launch of Miu Miu as a playful sub-label, which then introduced the red-line branding; this Miu Miu red stripe was subsequently absorbed into Prada’s main runway in 2018 and mentioned alongside Raf Simons as the sole driver of Prada’s future.
C After nylon bags boosted Prada’s popularity, the material was largely abandoned when Neil Barrett arrived, as Prada Sport was framed as a leather-based luxury line; Linea Rossa resurfaced only as a nostalgic capsule for Fall/Winter 2018 and was portrayed as separate from Prada’s broader strategic investments.
D Miuccia’s nylon work laid the groundwork for Neil Barrett’s minimalist menswear, but the narrator stresses that Prada Sport was purely about elegant tailoring with no real link to nylon, and that Linea Rossa’s 2018 appearance was a brief revival not referenced in the discussion of Prada’s long-term plans.

video364

Mel spectrogram of video364 with gold evidence (blue) and AudioHelix's retrieved 5s units (green)
Gold Evidence Ours: retrieved 5 s unit
ORIGINAL AUDIO
Query

According to the narration, what combination of reasons explains why humpback whales make their annual trip between Hawaii and Alaska, and what special ability helps them stay connected during that long journey?

A They leave Hawaii each year to feed in its rich coastal waters and then travel to Alaska mainly to avoid predators, maintaining contact along the way with short‑range clicks that only carry a few miles.
B They leave Alaska each year for Hawaii’s warmer, predator-sheltered waters where they breed and nurse calves despite there being no food, then return once the calves are strong enough for the 3,000‑mile journey, staying in constant contact with low‑frequency calls that can travel up to 10,000 miles.
✓ Correct
C They shuttle back and forth between Hawaii and Alaska throughout the year primarily to follow shifting fish schools for food, relying on high‑pitched songs that echo off the seafloor rather than long‑distance calls to stay in touch.
D They migrate from Alaska to Hawaii mainly to feed in clear, shallow waters and to clean parasites like the sea turtle does at the reef, then head back once food runs out, using visual cues instead of sound to keep track of one another over the open ocean.