Searcharxiv⌕ Search

arXiv subjects

Guangzhi Sun

Publications and source records attributed to Guangzhi Sun.

At least 19 recordsLinked to original sources

AudioTrace: File-Level Membership Detection for Audio Large Language Models

Significant privacy and copyright concerns are raised by the use of publicly available audio recordings to train audio large language models (LLMs), yet effective auditing methods remain limited. Existing membership inference attacks typically assume knowledge of complete training examples, including their associated textual supervision. In practice, however, audio recordings may be paired with privately generated transcripts, captions, or question-answer pairs that are unavailable to owners of the recordings. This motivates audio-file-level membership inference, which determines whether a recording was used for training without assuming knowledge of its training tasks and annotations. This paper provides the first systematic investigation of this challenge and introduces AudioTraceBench, a controlled benchmark comprising 1,285 audio recordings with explicit fine-tuning membership labels. Evaluations across three audio LLMs show that existing text- and video-based membership inference methods adapted to audio perform close to random guessing. To address this challenge, AudioTrace is proposed, where audio-conditioned proxy questions are generated using the target model, and membership evidence is extracted from its responses and aggregated at the recording level. Consistent improvements over existing baselines are observed, with AUROC reaching 0.60 and increasing further under increased fine-tuning exposure or adaptation capacity. Further experiments were conducted to examine the factors influencing membership signals for audio-file-level detection. These findings establish audio-file-level membership inference as a distinct and practically important auditing challenge, with AudioTrace providing an effective initial approach to this challenge.

cs.CR↗

Video-as-Experience: Learning Computer Use via Audio-Visual Understanding

Multimodal large language models (LLMs) have enabled capable computer-use agents that perceive graphical user interfaces, reason over instructions, and execute actions. However, existing approaches to learning from video demonstrations typically convert videos into screenshot-based action trajectories, rather than directly leveraging their rich audio-visual information. We introduce VEX (Video as EXperience), a framework that uses audio-visual understanding to unify learning from external demonstrations and reflecting on an agent's own behavior. Following a watch-act-rewatch paradigm, the agent extracts procedural guidance from tutorial videos containing visual demonstrations and spoken explanations, executes tasks, and verifies its progress by rewatching its execution as video. We instantiate VEX with video-SALMONN-VEX, a 9B-scale audio-visual computer-use agent built on the video-SALMONN 2 architecture with a Qwen3.5-9B backbone. We further introduce Action-Grounded Guide Optimization (AGGO), a reinforcement learning method that optimizes video-derived guidance for downstream action prediction, and augment AgentNet, OSWorld, and CUA-Gym with over 24K real-world tutorial videos. Experiments on OSWorld and a complementary subset of CUA-Gym demonstrate consistent improvements, with video-SALMONN-VEX achieving a 22% relative increase in task success over its Qwen3.5 backbone on OSWorld. Moreover, its generated guidance improves Gemini-3.5-flash by 10% relatively on OSWorld. These results highlight the potential of audio-visual understanding as a unified interface for procedural learning and self-verification in computer-use agents.

cs.AI↗

Disentangling Linguistic and Paralinguistic Information with Routed Sparse Autoencoders

Self-supervised speech encoders contain linguistic and paralinguistic information in a shared, entangled representation space. We combine a TopK sparse autoencoder with route-specific supervision and cross-factor adversaries. Across frozen SPEAR and WavLM encoders, independent probes show factor-specific retention and suppression: linguistic information remains stronger in the linguistic route, while paralinguistic factors, including speaker identity, emotion, and prosody, are retained in the paralinguistic route and substantially reduced in the linguistic route. The route organisation learned on LibriSpeech persists on MSP-Podcast without representation-side retraining. Feature-space route interventions further transfer the swapped factor while largely preserving the information carried by the unchanged route. These results show consistent route-selective separation across encoders, corpora, independent probes, and representation-level interventions.

cs.CL↗

Multiplexing Neural Audio Watermarks with Adaptive Routing

Audio watermarking supports speech authenticity verification. We study whether combining heterogeneous watermarks can retain at least one detectable provenance signal when their failure modes differ. This Any-survival objective concerns complementary evidence retention, not simultaneous survival of all constituent marks or recovery of all payloads. To our knowledge, this is the first systematic study of neural audio watermark multiplexing, with a scoped benchmark covering five released systems, parallel and sequential baselines, and 14 evaluation conditions. The benchmark shows complementary failure modes, but naive composition does not reliably turn them into system-level robustness under the Any-survival objective. We therefore formulate multiplexing as watermark allocation and study perceptual-adaptive time-frequency multiplexing (PA-TFM), a training-free routing method, and MaskNet, a learned time-domain router for separately trained systems with native detectors. We use the five-system scoped benchmark to characterize multiplexing behavior and the AudioSeal-PerTh pair as a representative heterogeneous pair for adaptive watermark routing. Compared with direct parallel composition, MaskNet improves average TPR@1%FPR from 0.75 to 0.88 and SNR from 15.20 dB to 25.36 dB; compared with PA-TFM, it gives a clear system-level robustness gain under the Any-survival objective while retaining similar high-fidelity behavior. A SpeechTokenizer-aware case study further improves average TPR@1%FPR to 0.91 and SpeechTokenizer robustness from 0.20 to 0.60, showing that channel-adapted constituents can enter the same routing framework.

eess.AS↗

Talk2Agent: Benchmarking Voice Interfaces for Text Agents

Large language model (LLM) computer-use agents are typically evaluated with clean written instructions, despite speech being an increasingly popular interface for interacting with such systems. Speech input introduces an additional failure point: transcription errors can alter task-critical entities, constraints, or targets before the agent begins reasoning, while conventional ASR metrics do not directly measure whether the information required for successful execution has been preserved. We introduce Talk2Agent, a benchmark for evaluating how effectively voice interfaces convey human-spoken instructions to LLM-based computer-use agents. Talk2Agent builds human-spoken versions of tasks from WildClawBench and OSWorld and evaluates a range of voice interfaces, including dedicated ASR models, audio-capable LLMs, contextual biasing, and LLM-based ontology repair. Because repeatedly executing long-horizon computer-use tasks is costly and stochastic, we further propose an execution-free, task-conditioned evaluation framework that projects the original task grader onto prompt-addressable intentions and measures how much task-relevant information is retained after the voice interface. On WildClawBench, Talk2Agent's execution-free native projection provides a practical, execution-grounded measure of voice-interface quality, correlating with downstream task completion and improving Pearson correlation by 0.246 over WER/CER on 32 hours of real human speech.

cs.AI↗

SALMONN-2: Advancing General-Purpose Hearing Abilities with Self-Supervised Representations

Recent audio large language models (ALLMs) are typically built upon audio encoders trained with large amounts of supervised data. Since self-supervised learning (SSL) audio encoder models are known to learn general-purpose and transferable representations, we investigate whether general-purpose SSL audio representations can serve as an effective foundation for ALLMs. We present SALMONN-2, an ALLM built upon a unified SSL encoder. To better exploit the hierarchical representations learned by SSL encoders, we propose a multi-layer feature fusion (MLF) adapter that aggregates information from all encoder layers before projecting them into the language model. Beyond conventional audio understanding tasks, we further explore multimodal in-context learning (MICL) in ALLMs and study how this capability can be acquired through contextual biasing training. Experimental results show that a general-purpose SSL encoder achieves performance comparable to, or better than, specialised supervised audio encoders while providing a more balanced capability across speech, audio, music and paralinguistic tasks. SALMONN-2 further achieves state-of-the-art performance among comparable-scale open-weight models on ALLM understanding benchmarks, obtaining the best results on MMAU-Pro, MMAR and MMSU. We also show that MICL does not emerge naturally in ALLMs, but can be effectively acquired through targeted contextual biasing training.

eess.AS↗

SLVMBench: Skill Learning from Video Memory

We introduce Skill Learning from Video Memory (SLVMBench), the first benchmark that jointly evaluates whether video large language models (video-LLMs) can learn skills from long video memory and apply them to real-time tasks. SLVMBench presents models with 2-3 hour video streams that contain a tutorial video embedded in a stream of arbitrary irrelevant videos, resembling real-world human learning practices. Video-LLMs are asked to apply the acquired skill to answer real-time questions about an ongoing video. Unlike long-video understanding benchmarks that emphasize passive comprehension and skill-learning benchmarks that rely on short, immediate demonstrations, SLVMBench tests the full pipeline of memorizing and extracting procedural knowledge, as well as transferring it to real-time tasks. Moreover, rigorous human annotations feature sub-second-level temporal calibration, manually engineered questions eliminating common-sense guessing, and collated tutorials to ensure coverage of the required skills. Evaluations on state-of-the-art proprietary and open-source video LLMs show that video-LLMs struggle substantially with learning and applying skill knowledge from videos. Moreover, performance degrades markedly when the skill knowledge is placed within a long video memory. These results reveal a key limitation of existing video LLMs and position SLVMBench as the first benchmark for studying real-time skill acquisition and application from long-context video memory.

cs.CV↗

video-SALMONN-R$^3$: Learning to ReWatch, ReAsk, and ReAnswer for Efficient Video Understanding

Video large language models (LLMs) are often constrained by computation and memory budgets, leading them to use reduced frame rates and spatial resolutions, which may cause them to miss critical information for question answering (QA). A practical and efficient solution is a two-stage paradigm: first perform coarse video understanding to localize relevant segments, and then re-watch these segments at higher temporal or spatial fidelity. In this paper, we present video-SALMONN-R$^3$, the first end-to-end video-LLM that enables re-watch through reinforcement learning without relying on chain-of-thought (CoT) cold-start. This design removes the need for costly CoT data annotations and avoids CoT-based supervised fine-tuning (SFT), which can otherwise degrade the pretrained video understanding abilities. To address the mismatch between the reasoning-first behavior induced by re-watch and the answer-first tendency of pretrained video-LLMs, we propose a re-answer strategy, in which the model first produces a direct answer in the first watch and then refines it after re-watching. Finally, to improve question adherence during re-watching, we propose a re-ask mechanism that re-injects the query when revisiting localized segments. Experimental results show that video-SALMONN-R$^3$ consistently outperforms both the base model and the QA-SFT baseline, while surpassing prior re-watch-based approaches with significantly lower computational cost. Code, models, and data will be publicly released upon acceptance.

cs.CV↗

CASE-Bench: Context-Aware SafEty Benchmark for Large Language Models

Aligning large language models (LLMs) with human values is essential for their safe deployment and widespread adoption. Current LLM safety benchmarks often focus solely on the refusal of individual problematic queries, which overlooks the importance of the context where the query occurs and may cause undesired refusal of queries under safe contexts that diminish user experience. Addressing this gap, we introduce CASE-Bench, a Context-Aware SafEty Benchmark that integrates context into safety assessments of LLMs. CASE-Bench assigns distinct, formally described contexts to categorized queries based on Contextual Integrity theory. Additionally, in contrast to previous studies which mainly rely on majority voting from just a few annotators, we recruited a sufficient number of annotators necessary to ensure the detection of statistically significant differences among the experimental conditions based on power analysis. Our extensive analysis using CASE-Bench on various open-source and commercial LLMs reveals a substantial and significant influence of context on human judgments (p<0.0001 from a z-test), underscoring the necessity of context in safety evaluations. We also identify notable mismatches between human judgments and LLM responses, particularly in commercial models within safe contexts.

cs.CL↗

Measuring the Redundancy of Decoder Layers in SpeechLLMs

Speech Large Language Models route speech encoder representations into an LLM decoder that typically accounts for over 90% of total parameters. We study how much of this decoder capacity is actually needed for speech tasks. Across two LLM families and three scales (1-8B), we show that decoder redundancy is largely inherited from the pretrained LLM: text and speech inputs yield similar redundant blocks. We then measure excess capacity by pruning decoder layers and analysing post-pruning healing to increase robustness. Our findings show that 7-8B models retain good ASR performance with only 60% of decoder layers, and the same trend extends to smaller scales with reduced pruning tolerance. We then generalise to speech translation, and show that the same blocks of layers are redundant across speech encoders, tasks and languages, indicating that a more global redundancy structure exists, enabling a single pruned and multi-tasks SpeechLLM backbone to be deployed.

cs.CL↗

Uncertainty-based Debiasing and Unlearning for Decontamination

Benchmark-based evaluation is the dominant paradigm for assessing large language model (LLM) capabilities, yet data contamination inflates reported performance and undermines fair comparison. Existing decontamination methods are evaluated solely through aggregate accuracy, which can obscure substantial differences in per-sample model behaviour, and many require access to an uncontaminated model. In this paper, we propose a sample-level evaluation framework for decontamination that complements accuracy-based assessment with distributional distance metrics, measuring how closely a decontaminated model recovers the output distribution of an uncontaminated model on each sample. Building on this framework, we introduce Uncertainty-Based Decontamination (UBD), a family of methods that leverage deep ensembles of the contaminated model to estimate per-sample memorization without requiring a uncontaminated model or knowledge of which samples are contaminated. UBD estimates a per-sample correction scalar from ensemble uncertainty, which is used to construct a debiased target distribution that suppresses the inflated probability mass on correct answers induced by contamination. This target is then used either as a post-hoc output correction (debiasing) or as a soft training signal for parameter update (unlearning). Experiments on MMLU-Pro and MATH-MCQA across multiple LLM backbones demonstrate that UBD produces per-sample output distributions substantially closer to those of an uncontaminated model than paraphrasing or choice-permutation baselines, while preserving model performance on uncontaminated data.

cs.CY↗

Who can we trust? LLM-as-a-jury for Comparative Assessment

Large language models (LLMs) are increasingly applied as automatic evaluators for natural language generation assessment often using pairwise comparative judgements. Existing approaches typically rely on single judges or aggregate multiple judges assuming equal reliability. In practice, LLM judges vary substantially in performance across tasks and evaluation aspects, and their judgment probabilities may be biased and inconsistent. Furthermore, human-labelled supervision for judge calibration may be unavailable. We first empirically demonstrate that inconsistencies in LLM comparison probabilities exist and show that it limits the effectiveness of direct probability-based ranking. To address this, we study the LLM-asa-jury setting and propose BT-sigma, a judge-aware extension of the Bradley-Terry model that introduces a discriminator parameter for each judge to jointly infer item rankings and judge reliability from pairwise comparisons alone. Experiments on benchmark NLG evaluation datasets show that BT-sigma consistently outperforms averaging-based aggregation methods, and that the learned discriminators strongly correlate with independent measures of the cycle consistency of LLM judgments. Further analysis reveals that BT-sigma can be interpreted as an unsupervised calibration mechanism that improves aggregation by modelling judge reliability.

cs.CL↗

OmniMem: Perturbation-aware Memory Compression for Streaming Audio-Visual LLMs

Audio-visual large language models (LLMs) hold strong promise for long-form video understanding, yet their long-video inference is fundamentally limited by the linear growth of video tokens and key-value (KV) caches. We present OmniMem, a memory-efficient streaming framework designed specifically for audio-visual LLMs. Unlike existing compression methods that treat all tokens uniformly, OmniMem introduces a modality-aware memory allocation strategy that separately manages visual and audio contexts, addressing the severe token imbalance between the two modalities. OmniMem further preserves informative and non-redundant KV states through perturbation-aware memory selection, enabling compact memory without sacrificing long-range understanding. To strengthen compression under realistic deployment constraints, we also explore budget-aware fine-tuning, which encourages the model to consolidate useful information into retained memory. Experiments on VideoMME Long, LVBench, and LVOmniBench with video-SALMONN 2+ and Qwen-2.5-Omni show that OmniMem consistently improves over strong training-free compression baselines by 2-4% absolute accuracy under the same memory budgets, with an additional 1-2% gain after fine-tuning.

cs.AI↗

Protecting Bystander Privacy via Selective Hearing in Audio LLMs

Audio Large language models (LLMs) are increasingly deployed in the real world, where they inevitably capture speech from unintended nearby bystanders, raising privacy risks that existing benchmarks and defences did not consider. We introduce SH-Bench, the first benchmark designed to evaluate selective hearing: a model's ability to attend to an intended main speaker while refusing to process or reveal information about incidental bystander speech. SH-Bench contains 3,968 multi-speaker audio mixtures, including both real-world and synthetic scenarios, paired with 77k multiple-choice questions that probe models under general and selective operating modes. In addition, we propose Selective Efficacy (SE), a novel metric capturing both multi-speaker comprehension and bystander-privacy protection. Our evaluation of state-of-the-art open-source and proprietary LLMs reveals substantial bystander privacy leakage, with strong audio understanding failing to translate into selective protection of bystander privacy. To mitigate this gap, we also present Bystander Privacy Fine-Tuning (BPFT), a novel training pipeline that teaches models to refuse bystander-related queries without degrading main-speaker comprehension. We show that BPFT yields substantial gains, achieving an absolute 47% higher bystander accuracy under selective mode and an absolute 16% higher SE compared to Gemini 2.5 Pro, which is the best audio LLM without BPFT. Together, SH-Bench and BPFT provide the first systematic framework for measuring and improving bystander privacy in audio LLMs.

cs.SD↗

Audio-Conditioned Diffusion LLMs for ASR and Deliberation Processing

Diffusion-based large language models (DLLMs) have recently attracted growing interest as an alternative to autoregressive decoders. In this work, we present an empirical study on using the diffusion-based large language model LLaDA for automatic speech recognition (ASR). We first investigate its use as an external deliberation-based processing module for Whisper-LLaMA transcripts. By leveraging the bidirectional attention and denoising capabilities of LLaDA, we explore random masking, low-confidence masking, and semi-autoregressive strategies, showing that Whisper-LLaDA substantially reduces WER compared with the baseline. On LibriSpeech, the best cascade system achieves 2.25%/4.94% WER on test-clean/test-other, representing a 12.3% relative improvement over the Whisper-LLaMA baseline on the test-other split. In contrast, a plain-text LLaDA without acoustic features fails to improve accuracy, highlighting the importance of audio-conditioned embeddings. We further evaluate Whisper-LLaDA as a standalone decoder for ASR with diffusion-based and semi-autoregressive decoding. Most experimental configurations achieve faster inference than the Whisper-LLaMA baseline, although recognition accuracy is slightly lower. These findings offer an empirical view of diffusion-based LLMs for ASR and point to promising directions for improvements. Code and model are open-sourced at https://github.com/liuzhan22/Diffusion-ASR.

eess.AS↗

Cross-Lingual Interleaving for Speech Language Models

Spoken Language Models (SLMs) aim to learn linguistic competence directly from speech using discrete units, widening access to Natural Language Processing (NLP) technologies for languages with limited written resources. However, progress has been largely English-centric due to scarce spoken evaluation benchmarks and training data, making cross-lingual learning difficult. We present a cross-lingual interleaving method that mixes speech tokens across languages without textual supervision. We also release an EN-FR training dataset, TinyStories (~42k hours), together with EN-FR spoken StoryCloze and TopicCloze benchmarks for cross-lingual semantic evaluation, both synthetically generated using GPT-4. On 360M and 1B SLMs under matched training-token budgets, interleaving improves monolingual semantic accuracy, enables robust cross-lingual continuation, and strengthens cross-lingual hidden-state alignment. Taken together, these results indicate that cross-lingual interleaving is a simple, scalable route to building multilingual SLMs that understand and converse across languages. All resources will be made open-source to support reproducibility.

cs.CL↗

Scaling Open Discrete Audio Foundation Models with Interleaved Semantic, Acoustic, and Text Tokens

Current audio language models are predominantly text-first, either extending pre-trained text LLM backbones or relying on semantic-only audio tokens, limiting general audio modeling. This paper presents a systematic empirical study of native audio foundation models that apply next-token prediction to audio at scale, jointly modeling semantic content, acoustic details, and text to support both general audio generation and cross-modal capabilities. We provide comprehensive empirical insights for building such models: (1) We systematically investigate design choices -- data sources, text mixture ratios, and token composition -- establishing a validated training recipe. (2) We conduct the first scaling law study for discrete audio models via IsoFLOP analysis on 64 models spanning $3{\times}10^{18}$ to $3{\times}10^{20}$ FLOPs, finding that optimal data grows 1.6$\times$ faster than optimal model size. (3) We apply these lessons to train SODA (Scaling Open Discrete Audio), a suite of models from 135M to 4B parameters on 500B tokens, comparing against our scaling predictions and existing models. SODA serves as a flexible backbone for diverse audio/text tasks -- we demonstrate this by fine-tuning for voice-preserving speech-to-speech translation, using the same unified architecture.

cs.SD↗

Speech-Audio Compositional Attacks on Multimodal LLMs and Their Mitigation with SALMONN-Guard

Recent progress in LLMs has enabled understanding of audio signals, but has also exposed new safety risks arising from complex audio inputs that are inadequately handled by current safeguards. We introduce SACRED-Bench (Speech-Audio Composition for RED-teaming) to evaluate the robustness of LLMs under complex audio-based attacks. Unlike existing perturbation-based methods that rely on noise optimization or white-box access, SACRED-Bench exploits speech-audio composition to enable effective black-box attacks. SACRED-Bench adopts three composition mechanisms: (a) overlap of harmful and benign speech, (b) mixture of benign speech with harmful non-speech audio, and (c) multi-speaker dialogue. These mechanisms focus on evaluating safety in settings where benign and harmful intents co-occur within a single auditory scene. Moreover, questions in SACRED-Bench are designed to implicitly refer to content in the audio, such that no explicit harmful information appears in the text prompt alone. Experiments demonstrate that even Gemini 2.5 Pro, a state-of-the-art proprietary LLM with safety guardrails fully enabled, still exhibits a 66% attack success rate. To bridge this gap, we propose SALMONN-Guard, the first guard model that jointly inspects speech, audio, and text for safety judgments, reducing the attack success rate to 20%. Our results highlight the need for audio-aware defenses to ensure the safety of multimodal LLMs. The dataset and SALMONN-Guard checkpoints can be found at https://huggingface.co/datasets/tsinghua-ee/SACRED-Bench.

cs.SD↗