Research map / Adversarial robustness and security / Black-box and transfer attacks
Inducing Overthink: Hierarchical Genetic Algorithm-based DoS Attack on Black-Box Large Language Reasoning Models
Shuqiang Wang, Wei Cao, Jiaqi Weng, Jialing Tao, Licheng Pan, Hui Xue, Zhixuan Chu
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 |
|---|---|---|---|
| use HGA-evolved prompts instead of missing-premise prompts | DeepSeek-R1 · SVAMP | Avg-len (tokens): 2319 -> 3589 | Improved |
| use HGA on GPT-o3 instead of MIP | GPT-o3 · SVAMP | Max-len (tokens): 906 -> 6562 | Improved |
| HGA vs MIP on Gemini-2.5-Flash | Gemini-2.5-Flash · SVAMP | Max-len (tokens): 683 -> 941 | Mixed |
| HGA vs AutoDoS (fewer input tokens) | HGA · MATH | Avg-len (tokens): 11207 -> 25419 | Improved |
| HGA vs AutoDoS | HGA · SVAMP | Max-len (tokens): 12775 -> 10708 | Hurt |
| add overthink-marker score (alpha 0.7 vs pure length alpha 1.0) (ablation) | HGA · GSM8K | Avg-len (tokens): 1535 -> 3646 | Improved |
| add overthink-marker score (alpha 0.7 vs pure length alpha 1.0) (ablation) | HGA · SVAMP | Avg-len (tokens): 3192 -> 6306 | Improved |
| add overthink-marker score (alpha 0.7 vs pure length alpha 1.0) (ablation) | HGA · MATH | Avg-len (tokens): 6258 -> 18315 | Improved |
| add length term to pure overthink-marker objective (alpha 0.7 vs 0.0) (ablation) | HGA · MATH | Avg-len (tokens): 7996 -> 18315 | Improved |
| composite fitness alpha 0.7 vs pure length when transferred from Qwen3-14B (ablation) | Transferred attack · MATH (DeepSeek-R1 target) | Avg-len (tokens): 7018 -> 10893 | Improved |
| composite fitness alpha 0.7 vs pure length when transferred (ablation) | Transferred attack · GSM8K (DeepSeek-R1 target) | Avg-len (tokens): 2778 -> 4038 | Mixed |
| increase generations/population (search budget) (ablation) | HGA · MATH | Max-len/Avg-len: population 5: 5893 -> 11301 max-len over 5 -> 30 generations | No clear effect |
Closest papers by meaning
- BadThink: Triggered Overthinking Attacks on Chain-of-Thought Reasoning in Large Language Models (AAAI 2026 · similarity 0.90)
- Do NOT Think That Much for 2+3=? On the Overthinking of Long Reasoning Models (ICML 2025 · similarity 0.88)
- Let LRMs Break Free from Overthinking via Self-Braking Tuning (NeurIPS 2025 · similarity 0.88)
- Mitigating Overthinking in Large Reasoning Language Models via Reasoning Path Deviation Monitoring (NeurIPS 2026 · similarity 0.88)
- Don’t Think Longer, Think Wisely: Optimizing Thinking Dynamics for Large Reasoning Models (NeurIPS 2025 · similarity 0.87)
- Efficient Reasoning with Balanced Thinking (ICLR 2026 · 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.
