Research map / Adversarial robustness and security / Attacks on text and graphs
LoopLLM: Transferable Energy-Latency Attacks in LLMs via Repetitive Generation
Xingyu Li, Xiaolei Liu, Cheng Liu, Yixiao Xu, Kangyi Ding, Bangzhou Xin, Jia-Li Yin
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 |
|---|---|---|---|
| induce repetitive generation instead of delaying EOS (LoopLLM-t vs Engorgio) | energy-latency attack · Dialogue prompts, Llama3-8B | ASR: 1% -> 94% | Improved |
| add token-aligned ensemble of Llama3-3B and Llama3-8B surrogates (ablation) | LoopLLM-t transfer attack · Transfer to Phi4-mini | ASR: 49% -> 76% | Improved |
| add Llama2-7B & 13B ensemble (ablation) | LoopLLM-t transfer attack · Transfer to Vicuna-7B | ASR: 16% -> 42% | Improved |
| ensemble of Llama2-7B & 13B instead of single Llama3-8B surrogate (ablation) | LoopLLM-t transfer attack · Transfer to Mistral-7B | ASR: 33% -> 27% | Hurt |
| use shorter cyclic segment (length 1 vs 5 vs 10) (ablation) | LoopLLM · GLM4-9B | ASR: 88% -> 67% | Improved |
| choose specific tokens instead of random vocabulary tokens for the cyclic segment (ablation) | LoopLLM-t · GLM4-9B | ASR: 64% -> 88% | Improved |
| add repeated-token halting defense | LoopLLM-t · Llama3-8B | ASR: 93% -> 14% | Hurt |
| add repeated-token halting defense | LoopLLM-p · Llama3-8B | ASR: 79% -> 76% | No clear effect |
| use beam search instead of greedy search (ablation) | LoopLLM-t · GLM4-9B | ASR: 68% -> 94% | Improved |
| raise sampling temperature to 1.2 (ablation) | LoopLLM-t · GLM4-9B | ASR: 84% -> 61% | Hurt |
| increase suffix length (ablation) | LoopLLM · Avg-len and ASR vs suffix length | Avg-len / ASR | Improved |
| add more optimization steps (ablation) | LoopLLM · Avg-len and ASR vs steps | Avg-len / ASR | Improved |
Closest papers by meaning
- LingoLoop Attack: Trapping MLLMs via Linguistic Context and State Entrapment into Endless Loops (ICLR 2026 · similarity 0.88)
- The VLLM Safety Paradox: Dual Ease in Jailbreak Attack and Defense (NeurIPS 2025 · similarity 0.88)
- Catastrophic Jailbreak of Open-source LLMs via Exploiting Generation (ICLR 2024 · similarity 0.87)
- Exploiting LLM Quantization (NeurIPS 2024 · similarity 0.87)
- SilentStriker: Toward Stealthy Bit-Flip Attacks on Large Language Models (NeurIPS 2025 · similarity 0.87)
- An Engorgio Prompt Makes Large Language Model Babble on (ICLR 2025 · similarity 0.87)
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.
