Research map / Large language models / Language, safety and interpretability
Exploiting Synergistic Cognitive Biases to Bypass Safety in LLMs
Xikang Yang, Biyu Zhou, Xuehai Tang, Jizhong Han, Songlin Hu
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 |
|---|---|---|---|
| add PPO reinforcement learning after SFT (ablation) | CognitiveAttack (GPT-4o-mini target) · AdvBench | ASR (%): 30.0 -> 64.9 | Improved |
| add PPO reinforcement learning after SFT (ablation) | CognitiveAttack (Llama-2 target) · AdvBench | ASR (%): 36.5 -> 94.3 | Improved |
| add PPO reinforcement learning after SFT (ablation) | CognitiveAttack (DeepSeek-V3 target) · AdvBench | ASR (%): 55.7 -> 94.9 | Improved |
| add PPO reinforcement learning after SFT (ablation) | CognitiveAttack (Qwen-turbo target) · AdvBench | ASR (%): 35.6 -> 88.7 | Improved |
| add SFT on bias-annotated rewrites over prompting alone (ablation) | CognitiveAttack (DS-v3 target) · AdvBench | ASR (%): 41.5 -> 55.7 | Improved |
| use learned multi-bias attack instead of PAP | CognitiveAttack · HarmBench (30 LLMs) | ASR (%): 31.6 -> 60.1 | Improved |
| apply paraphrase defense | CognitiveAttack (GPT-4o-mini) · HarmBench | ASR (%): 56.5 -> 31.5 | Hurt |
| apply retokenization defense | CognitiveAttack (GPT-4o-mini) · HarmBench | ASR (%): 56.5 -> 49.3 | Hurt |
| apply perplexity filter | CognitiveAttack (GPT-4o-mini) · HarmBench | ASR (%): 56.5 -> 50.8 | Hurt |
| apply LlamaGuard detector | CognitiveAttack (GPT-4o-mini) · HarmBench | ASR (%): 56.5 -> 4.5 | Hurt |
| apply Perspective API detector | CognitiveAttack (GPT-4o-mini) · HarmBench | ASR (%): 56.5 -> 54.8 | No clear effect |
Closest papers by meaning
- LLM Agents Can Be Choice-Supportive Biased Evaluators: An Empirical Study (AAAI 2025 · similarity 0.89)
- Navigating the Safety Landscape: Measuring Risks in Finetuning Large Language Models (NeurIPS 2024 · similarity 0.88)
- Cannot See the Forest for the Trees: Invoking Heuristics and Biases to Elicit Irrational Choices of LLMs (ICML 2025 · similarity 0.88)
- Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (ICLR 2024 · similarity 0.88)
- Anchoring the Mind of Multimodal Reasoners: Cognitive Bias as a Vector for Jailbreak Attacks (CVPR 2026 · similarity 0.88)
- Safeguarding LLMs via Model-Agnostic Latent Safety Signals from Dark Knowledge (NeurIPS 2026 · similarity 0.88)
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.
