Research map / Large language models / Language, safety and interpretability
Stable-GFlowNet: Toward Diverse and Robust LLM Red-Teaming via Contrastive Trajectory Balance
Minchan Kwon, Sunghyun Baek, Minseo Kim, Jaemyung Yu, Dongyoon Han, Junmo Kim
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 |
|---|---|---|---|
| replace GFN-TB with CTB (pairwise, no Z) (ablation) | GFN with MKS reward · Red-teaming Qwen2.5-1.5B-Instruct | Unique Attacks: 67 -> 108 | Improved |
| add Noisy Gradient Pruning to CTB (ablation) | GFN-CTB · Red-teaming Qwen2.5-1.5B-Instruct | UA and ASR: UA 108 -> 121, ASR 82.9 -> 92.2 | Improved |
| add Min-K Fluency Stabilizer to plain reward (ablation) | GFN-CTB · Red-teaming Qwen2.5-1.5B-Instruct | Unique Attacks: 0 -> 108 | Improved |
| use MKS instead of KL-style reference probability reward (ablation) | GFN-CTB · Red-teaming Qwen2.5-1.5B-Instruct | Unique Attacks: 20 -> 108 | Improved |
| use MKS instead of log-prob sum threshold (ablation) | GFN-CTB · Red-teaming Qwen2.5-1.5B-Instruct | Unique Attacks: 78 -> 108 | Improved |
| replace CTB with batch mean or median baseline (ablation) | Z-free GFN · Red-teaming noisy reward | UA: Mean 106, Median 110, CTB 108 | No clear effect |
| remove learnable Z (mean, median or CTB) versus TB (ablation) | GFN · Red-teaming noisy reward | UA: TB 67 vs 106-110 | Improved |
| increase k in MKS (ablation) | S-GFN · Red-teaming unique prompts | Unique Attacks | Improved |
| set MKS threshold to -20 or below (ablation) | S-GFN · Red-teaming | ASR and UA | Hurt |
| set NGP sigma to 7.0 (ablation) | S-GFN · Red-teaming | Unique Attacks | Hurt |
| NGP sigma below 1.0 (ablation) | S-GFN · Red-teaming | ASR and UA | No clear effect |
| use S-GFN instead of GFN | GFN attacker · Red-teaming Qwen2.5-1.5B-Instruct | Unique Attacks: 17.67 -> 134.00 | Improved |
Closest papers by meaning
- Learning Diverse Attacks on Large Language Models for Robust Red-Teaming and Safety Tuning (ICLR 2025 · similarity 0.90)
- There are Levels to It: Red Teaming LLMs with Hierarchical Reinforcement Learning (NeurIPS 2026 · similarity 0.88)
- Stable GFlowNets with Probabilistic Guarantees (NeurIPS 2026 · similarity 0.87)
- DiveR-CT: Diversity-enhanced Red Teaming Large Language Model Assistants with Relaxing Constraints (AAAI 2025 · similarity 0.87)
- gfnx: Fast and Scalable Library for Generative Flow Networks in JAX (NeurIPS 2026 · similarity 0.85)
- GFlowNet Training by Policy Gradients (ICML 2024 · similarity 0.85)
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.
