cs.SDNov 19, 2026

When Sounds Turn Heads: Recovering Event Timing That Audio LLMs Fail to Say

Authors: Jongbhin Woo, Jaewook Jeong, Joon Son Chung

Organizations: KAIST

Abstract

Audio LLMs can know when a queried sound occurs but fail to say it. We find a few attention heads that track the queried sound: they shift when the query changes, follow the sound when it moves, and removing or blocking them often hurts the model's timestamps. Yet the answer does not express what they track. Intervals read directly from these heads outperform the model's original answer and recover events that its timestamps miss. To bring this timing into the answer, we introduce head-guided decoding, which needs no training. It keeps the model's decision on whether the sound is present and rescores its candidate timestamps with a prior from these heads. Heads selected on development data transfer unchanged to new recordings and free-text queries, where head-guided decoding improves the timestamps of three audio LLMs, scores above their head readout on every benchmark, and, in most settings, outperforms random-head guidance. On recordings released after the models were trained, head-guided decoding raises mIoU from 0.07-0.36 to 0.48-0.56. These results suggest that poor temporal answers in audio LLMs partly reflect a failure to say rather than to hear, and that the model's own attention can close part of this gap.

Figures & tables

Explore similar work

Jul 11, 2026

GigaChat Audio: Time-aware Large Audio Language Model

Temporal grounding in long recordings remains challenging for audio-conditioned LLMs. We present a time-aware audio LLM that answers questions with explicit timestamps over up to 120 minutes of input. Our approach interleaves periodic time markers with continuous audio tokens using large-scale synthetic supervision from a cascaded pipeline. Our model achieves strong temporal-grounding accuracy on short and long benchmarks and supports time-anchored fragment descriptions and summaries. Extensive ablations examine how time representation, marker frequency, tokenization, and duration-mixture design affect accuracy and computational cost. We release model weights and datasets to support further research on time-aware audio understanding, available at https://huggingface.co/ai-sage/GigaChat3.1-Audio-10B-A1.8B.
Sep 4, 2026

Tracing Audio Grounding and Answer Selection in Audio LLMs

Audio Large Language Models (Audio LLMs) have advanced in audio understanding, yet they can still predict the answer by reasoning from textual cues or linguistic priors rather than the provided audio. A common remedy is to train models on data whose answers cannot be inferred from text alone. This approach can improve performance, but what changes within the model remains unclear. In this paper, we ask what must happen inside the model for the audio to actually determine the answer. Our findings are threefold. (1) Replacing the audio with silence or unrelated audio causes substantially larger performance degradation in the trained model than in the pretrained model. (2) Acoustic information most strongly shapes the model's representations of the answer choices in early-to-middle layers, while training mainly increases the influence of audio information on the final prediction in middle-to-late layers. (3) The weights learned during training have their largest impact in specific layer bands. Together, these results provide a mechanistic account of how training strengthens the use of acoustic evidence in Audio LLMs.
Feb 16, 2026

Event-Grounded Question Answering over Long Audio via Structured Retrieval

Answering natural-language questions over multi-hour audio requires reliable event recognition, temporal grounding, and efficient retrieval. We present LA-RAG (Long Audio Retrieval-Augmented Generation), a structured framework that converts audio into timestamped event records, stores them in an event database, and answers questions using intent-aware retrieval and LLM-based generation. LA-RAG supports two deployment settings: offline grounding mode, in which long recordings are pre-indexed for low-latency querying, and inference-time grounding mode, which performs query-conditioned grounding over shorter, open-ended clips. We evaluate LA-RAG on two synthetic 24-hour benchmarks, Home-IoT and Industrial-IoT, and on CASTELLA-QA and SpotSound-Bench-QA, which are QA versions of two public audio temporal grounding benchmarks that we construct. The results demonstrate effective long-audio question answering with low query-time latency after grounding and indexing. They also reveal a substantial gap between event detection and temporal localization in current large audio-language models. Explicit timestamped grounding consistently improves temporal localization while maintaining competitive detection accuracy. These findings establish structured grounding and retrieval as a practical complement to generative audio-language models for deployment-oriented long-audio understanding.