Research map / Large language models / Agents, code and math
Training Language Model Agents to Find Vulnerabilities with CTF-Dojo
Terry Yue Zhuo, Dingmin Wang, Hantian Ding, Varun Kumar, Zijian Wang
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 |
|---|---|---|---|
| fine-tune on 486 CTF-Dojo trajectories | Qwen3-32B · 3-benchmark average | Pass@1 (%): 20.3 -> 31.9 | Improved |
| fine-tune on 486 CTF-Dojo trajectories | Qwen3-8B · InterCode-CTF | Pass@1 (%): 46.5 -> 53.8 | Improved |
| compare 486 CTF-Dojo trajectories with 9,464 Cyber-Zero trajectories | Qwen3-32B · 3-benchmark average | Pass@1 (%): Cyber-Zero 33.4 vs CTF-Dojo 31.9 | Mixed |
| add writeup hints during trajectory collection (ablation) | Qwen3-14B · CTF-Dojo tasks | solved rate (%): 12.9 -> 21.1 | Improved |
| add writeup hints during trajectory collection (ablation) | DeepSeek-V3-0324 · CTF-Dojo tasks | solved challenges: 168 -> 217 | Improved |
| augment runtime with randomized ports, paths, distractors (ablation) | Qwen3-Coder · CTF-Dojo tasks (4 rollouts) | challenges solved: 169 -> 211 | Improved |
| augment runtime with randomized ports, paths, distractors (ablation) | DeepSeek-V3-0324 · CTF-Dojo tasks (4 rollouts) | challenges solved: 156 -> 217 | Improved |
| combine trajectories from Qwen3-Coder and DeepSeek-V3-0324 (ablation) | Qwen3-32B SFT · 3-benchmark average | Pass@1 (%): 31.3 -> 31.9 | Improved |
| combine teachers for 8B model (ablation) | Qwen3-8B SFT · 3-benchmark average | Pass@1 (%): 17.6 -> 18.9 | Improved |
| scale training trajectories from 0 to 486 (ablation) | Qwen3-32B SFT · 3-benchmark average | Pass@1 (%): 22.0 -> 31.9 | Improved |
Closest papers by meaning
- MLE-Dojo: Interactive Environments for Empowering LLM Agents in Machine Learning Engineering (NeurIPS 2025 · similarity 0.90)
- FT-Dojo: Towards Autonomous LLM Fine-Tuning with Language Agents (ICML 2026 · similarity 0.88)
- Offline Training of Language Model Agents with Functions as Learnable Weights (ICML 2024 · similarity 0.87)
- How to Train Your LLM Web Agent: A Statistical Diagnosis (NeurIPS 2025 · similarity 0.87)
- AgentBench: Evaluating LLMs as Agents (ICLR 2024 · similarity 0.87)
- ToolACE: Winning the Points of LLM Function Calling (ICLR 2025 · 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.
