Research map / Adversarial robustness and security / Attacks on text and graphs
Tree-based Dialogue Reinforced Policy Optimization for Red-Teaming Attacks
Ruohao Guo, Afshin Oroojlooy, Roshan Sridhar, Miguel Ballesteros, Alan Ritter, Dan Roth
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 RL stage after SFT (ablation) | DialTree attacker · HarmBench, closed-source avg | ASR (%): 29.6 -> 82.7 | Improved |
| add RL stage after SFT (ablation) | DialTree attacker · HarmBench, open-source avg | ASR (%): 44.3 -> 80.2 | Improved |
| add format pruning (ablation) | DialTree · HarmBench (Llama-3.1-8B) | ASR (%): 62.5 -> 71.5 | Improved |
| add topic pruning (ablation) | DialTree · HarmBench (Mistral-7B) | ASR (%): 83.0 -> 86.5 | Improved |
| add both pruning mechanisms (ablation) | DialTree · HarmBench (Gemma-2-9B) | ASR (%): 37.0 -> 70.5 | Improved |
| add tree rollout versus single-path GRPO (ablation) | DialTree · HarmBench subset, open-source avg | ASR (%): 68.8 -> 77.3 | Improved |
| use adaptive masking of format tokens instead of none/static (ablation) | DialTree · RL training stability | malformed trajectory rate: ~100% -> below 50% | Improved |
| increase branching factor from 1 to 4 (ablation) | DialTree · Trajectory diversity | Self-BLEU: 0.554 -> 0.160 | Improved |
| increase branching factor from 4 to 8 (ablation) | DialTree · Trajectory diversity | Self-BLEU: 0.160 -> 0.269 | Hurt |
| increase group size from 8 to 32 (ablation) | DialTree · HarmBench (Mistral-7B) | ASR (%): 71.5 -> 86.5 | Improved |
| increase tree depth from 5 to 7 turns (ablation) | DialTree · HarmBench | ASR (%) | Hurt |
| increase training goals from 200 to 1200 (ablation) | DialTree · HarmBench (GPT-4o) | ASR (%): 77.5 -> 59.5 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AutoDAN-Turbo — AutoDAN-Turbo: A Lifelong Agent for Strategy Self-Exploration to Jailbreak LLMs (ICLR 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Learning Diverse Attacks on Large Language Models for Robust Red-Teaming and Safety Tuning (ICLR 2025 · similarity 0.88)
- TreeTeaming: Autonomous Red-Teaming of Vision-Language Models via Hierarchical Strategy Exploration (CVPR 2026 · similarity 0.88)
- There are Levels to It: Red Teaming LLMs with Hierarchical Reinforcement Learning (NeurIPS 2026 · similarity 0.88)
- RedDebate: Safer Responses Through Multi-Agent Red Teaming Debates (ICML 2026 · similarity 0.88)
- Automated Red Teaming with GOAT: the Generative Offensive Agent Tester (ICML 2025 · similarity 0.87)
- The Alignment Waltz: Jointly Training Agents to Collaborate for Safety (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.
