Research map / Large language models / Language, safety and interpretability
Capability-Based Scaling Trends for LLM-Based Red-Teaming
Alexander Panfilov, Paul Kassianik, Maksym Andriushchenko, Jonas Geiping
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 |
|---|---|---|---|
| increase attacker general capability (MMLU-Pro) | LLM-based jailbreak attackers (PAIR, TAP, PAP, Crescendo) · HarmBench | average attacker ASR: Spearman rho > 0.84 | Improved |
| increase target capability beyond attacker capability | target models · HarmBench | attack success rate: R^2 = 0.83; ASR <= 0.2 when gap < -3.5 | Hurt |
| use a more capable inner judge in PAIR (ablation) | PAIR with Vicuna-7b and Llama3-8b attackers · HarmBench | ASR@25 | No clear effect |
| use a more capable inner judge in PAIR (ablation) | PAIR with Vicuna-7b and Llama3-8b attackers · HarmBench | ASR@1 | Improved |
| use TAP instead of Crescendo or PAP under same query budget | attack method · HarmBench | ASR | Improved |
| LoRA unlocking of safety-tuned attackers | open-weight models · MMLU-Pro | MMLU-Pro: biggest absolute change 4% | No clear effect |
| LoRA unlocking of safety-tuned attackers | open-weight models · IFEval | IFEval loose prompt accuracy: Qwen-2.5-72B 0.57 (delta -0.18) | Hurt |
Closest papers by meaning
- CoP: Agentic Red-teaming for Large Language Models using Composition of Principles (NeurIPS 2025 · similarity 0.90)
- Learning Diverse Attacks on Large Language Models for Robust Red-Teaming and Safety Tuning (ICLR 2025 · similarity 0.89)
- Co-RedTeam: Orchestrated Security Discovery and Exploitation with LLM Agents (ICML 2026 · similarity 0.89)
- TreeTeaming: Autonomous Red-Teaming of Vision-Language Models via Hierarchical Strategy Exploration (CVPR 2026 · similarity 0.89)
- There are Levels to It: Red Teaming LLMs with Hierarchical Reinforcement Learning (NeurIPS 2026 · similarity 0.89)
- DiveR-CT: Diversity-enhanced Red Teaming Large Language Model Assistants with Relaxing Constraints (AAAI 2025 · similarity 0.89)
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.
