The paper investigates how large audio‑language models (LALMs) hallucinate objects by affirming sounds that are not present. Using a black‑box causal test, the authors isolate and remove the claimed evidence region via spectral gating and compare the effect against a random‑window control across a 2×2×2 grid on three LALMs. The study finds that localization of evidence improves detection over random removal, that raw detection values are model‑dependent, and that contrastive decoding combining blank and language‑prior contrasts can significantly boost accuracy, but a unified causal‑grounding profile does not transfer between models.
By Mohammad Khalooei, Mohammad Sabokrou
The study investigates where a large audio language model (LALM) derives its refusal responses to harmful spoken requests. By applying causal interventions at the audio-to-language-model interface and within the language-model residual layers, the authors find that the primary influence on the refusal margin comes from mid-to-late layers of the text language model rather than the audio front end. Ablations of the audio interface have minimal effect, while zeroing the encoder output still reduces the margin, indicating the audio pathway remains active but is not the main source of refusal decisions.
By Leonardo Haw-Yang Foo, Hung-yi Lee
arXiv:2608. 19211v1 Announce Type: cross Abstract: Human speech is richly expressive, with prosody carrying linguistic and emotional information beyond the lexical content.
By Linkai Peng, Baorian Nuchged
The paper investigates how to evaluate audio‑language models by separating the use of acoustic evidence from the need to invoke a generative audio model. Using a controlled call‑decision framework, the authors compare policies that rely on transcript labels, encoder outputs from CLAP, AST, or WavLM, and optional calls to generative models such as Qwen2‑Audio, Qwen2.5‑Omni, or MOSS‑Audio. Results on the VocalSound dataset show that while transcript‑only accuracy is low (0.296), encoder‑only controls achieve high accuracy (≈0.85) without any generative calls, and adding generative calls yields only a marginal improvement (0.925 vs. 0.921).
By Mengzhe Geng
The paper presents the first music‑specific, layer‑wise empirical study of hallucination in audio‑language models, framing it as a hierarchical perceptual grounding failure across five layers: sound events, temporal properties, tonal attributes, style, and emotion. It introduces MuseDiag, a diagnostic framework that evaluates nine models and finds universal vocal misperception, significant tonal perception differences, and identifies Audio‑Flamingo‑3 as the most stable model. The study also proposes two training‑free mitigation methods, ADD‑M and TPA, which reduce hallucination in probing but show variable effectiveness in free‑form generation, highlighting the need for multi‑paradigm evaluation.
By Yu Liu, Jiahui Liu, Zhilin Liu, Cong Cao, Fangfang Yuan, Yuling Yang, Pin Xu, Yanbing Liu
The paper investigates how to evaluate generative audio large language models (Audio‑LLMs) on known closed‑set tasks by separating the decision to call a generative model from the use of acoustic evidence. It introduces a controlled call‑decision framework where a policy can choose between a transcript label, encoder evidence from CLAP, AST, or WavLM, or a generative call to Qwen2‑Audio, Qwen2.5‑Omni, or MOSS‑Audio, and measures the impact of generative calls on accuracy. Results on the VocalSound dataset show that while transcript‑only accuracy is low (0.296), encoder‑based controls achieve high accuracy (≈0.85) without any generative calls, and adding generative calls yields only a marginal improvement (0.925 vs. 0.921).
By Mengzhe Geng
The paper investigates how streaming emotion recognition models can be misled by their own prior predictions, a problem termed previous-belief contamination (PBC). Using a counterfactual diagnostic on CREMA-D-Stream, the authors show that feeding a model’s previous emotion label into its current prediction can drastically lower accuracy and flip many predictions, with the effect varying by label. To mitigate PBC, they propose EmoUpdate, a training‑free framework that isolates current audio perception from historical context through a prior‑blind firewall, a causal belief filter, and a decontamination operator, achieving significant gains across multiple SpeechLMs and benchmarks.
By Haoyue Liu, Zhichao Wang, Ye Chen, Haonan Deng, Xiaoying Tang
arXiv:2610.02976v1 Announce Type: new
Abstract: Audio-Visual Large Language Models (AV-LLMs) remain prone to cross-modal hallucinations, where one modality incorrectly affects predictions about anoth...
By Hyunjae Ra, Aecheon Jung, Jungin Park, Sungeun Hong
The paper introduces SAC‑Bench, a benchmark of 616,031 audio clips that test audio‑language models (ALMs) on how they handle conflicts between spoken claims and acoustic evidence. It covers 188 sound‑event classes across seven semantic, temporal, and spatial tasks, pairing environmental recordings with either compatible or contradictory synthesized spoken claims and including audio‑only controls. Evaluation of five open‑weight ALMs shows that high audio‑only accuracy often coexists with significant drops in performance under conflicting claims, revealing gaps in conflict robustness and the mixed effects of prompt interventions.
By Wenbin Wang, Yuheng Li, Koichiro Yoshino
arXiv:2606. 02642v1 Announce Type: cross Abstract: Despite the success of audio-visual large-language models (LLMs), they can produce plausible but ungrounded outputs, termed hallucination.
By Chenshuang Zhang, Kyeong Seon Kim, Chengxin Liu, Tae-Hyun Oh
Large audio‑language models (LALMs) can understand music but struggle to pinpoint a described moment in time. Using a paired‑query probe on Slakh2100 stem activations, researchers attached state and change queries to the same audio span and varied preceding context. Across frozen LALMs, different extraction methods (full‑track prompting, windowed yes/no scoring, and attention) revealed that state queries perform well without context, while change queries benefit from preceding audio in most short‑segment models; however, 30‑second padded encoders show no comparable recovery, and prompting or attention alone expose almost none of this capability.
By Eunsik Shin
The paper investigates why automatic speech recognition (ASR) systems sometimes generate fluent text that does not correspond to the input audio, a phenomenon termed hallucination. By examining two independently trained Conformer‑Large models—one using CTC and the other RNN‑T—under conditions of environmental noise and speaker‑background shift, the authors identify the final encoder stage as a critical boundary. Bypassing this final block leads to divergence on almost all utterances, while bypassing earlier blocks has minimal effect; at this stage, representations become compact, the decoder can read the text, and grapheme information becomes explicit, yet the output is garbled or repetitive rather than fluent fabrication. The study thus pinpoints a mechanistic precondition for hallucination—failure to produce adequately grounded output—though it does not fully explain naturally occurring hallucinations, and it highlights a consistent terminal‑stage dependency across decoder families and distribution shifts.
By Hamees Sayed, Apoorv Singh, Kumar Aman, Akshat Mandloi