Research map / Architectures and efficiency / RNNs and normalization
Sparse Tokens Suffice: Jailbreaking Audio Language Models via Token-Aware Gradient Optimization
Zheng Fang, Xiaosen Wang, Shenyi Zhang, Shaokang Wang, Zhijin Ge
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| enforce token-level sparsity during optimization (zeta=0.25) instead of pruning the dense perturbation afterwards (ablation) | TAGO · AdvBench-50, Qwen3-Omni | ASRl (%): 1 -> 86 | Improved |
| reduce token retention ratio from 1.0 to 0.25 (ablation) | TAGO · AdvBench-50, Qwen3-Omni (rho=0.9) | ASRl (%): 87 -> 86 | No clear effect |
| reduce token retention ratio from 1.0 to 0.25 (ablation) | TAGO · AdvBench-50, Qwen2.5-Omni | ASRl (%): 65 -> 53 | Hurt |
| reduce token retention ratio from 1.0 to 0.25 (ablation) | TAGO · AdvBench-50, LLaMA-Omni | ASRl (%): 71 -> 72 | No clear effect |
| lower early-stopping confidence rho from 0.9 to 0.7 (ablation) | TAGO (zeta=1.0) · AdvBench-50, Qwen3-Omni | ASRl (%): 87 -> 32 | Hurt |
| lower early-stopping confidence rho from 0.9 to 0.7 (ablation) | TAGO (zeta=1.0) · AdvBench-50, Qwen3-Omni | average iterations: 256.64 -> 75.39 | Improved |
| reduce token retention ratio from 1.0 to 0.25 (ablation) | TAGO · AdvBench-50, Qwen3-Omni (rho=0.9) | average iterations: 256.64 -> 323.16 | Hurt |
| extreme sparsity, token retention ratio 0.05 instead of 1.0 (ablation) | TAGO · AdvBench-50, Qwen3-Omni (rho=0.9) | ASRl (%): 87 -> 38% | Hurt |
| token-aware sparse optimization (zeta=1.0) versus dense baselines SpeechGuard and AdvWave | TAGO · AdvBench-50, Qwen3-Omni | ASRl (%): 42 -> 87 | Improved |
| add uniform waveform noise to input audio | TAGO (zeta=0.25) · AdvBench-50, LLaMA-Omni | ASRl (%): 72.0 -> 65.0 | Mixed |
| switch TTS speaker/backend (Gemini-TTS six speakers) | TAGO (zeta=0.25) · AdvBench-50, Qwen2.5-Omni | ASRl (%): 53.0 -> 55.7±2.1 | No clear effect |
Closest papers by meaning
- ALMGuard: Safety Shortcuts and Where to Find Them as Guardrails for Audio–Language Models (NeurIPS 2025 · similarity 0.90)
- Jailbreak-AudioBench: In-Depth Evaluation and Analysis of Jailbreak Threats for Large Audio Language Models (NeurIPS 2025 · similarity 0.89)
- AdvWave: Stealthy Adversarial Jailbreak Attack against Large Audio-Language Models (ICLR 2025 · similarity 0.88)
- ALMTokenizer: A Low-bitrate and Semantic-rich Audio Codec Tokenizer for Audio Language Modeling (ICML 2025 · similarity 0.86)
- Efficient LLM Jailbreak via Adaptive Dense-to-sparse Constrained Optimization (NeurIPS 2024 · similarity 0.86)
- JailPO: A Novel Black-Box Jailbreak Framework via Preference Optimization Against Aligned LLMs (AAAI 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
