Research map / Large language models / Agents, code and math
Just Ask: Curious Code Agents Reveal System Prompts in Frontier LLMs
Xiang Zheng, Yutao Wu, Hanxun Huang, Yige Li, Xingjun Ma, Bo Li, Yu-Gang Jiang, Cong 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 |
|---|---|---|---|
| add UCB-guided skill selection instead of random selection (ablation) | JustAsk · Framework ablation on Claude Sonnet 4, B=5 | Sim-GT: 0.757 -> 0.867 | Improved |
| add UCB-guided skill selection instead of random selection (ablation) | JustAsk · Framework ablation on Claude Sonnet 4, B=10 | Sim-GT: 0.867 -> 0.889 | Improved |
| add skill framework to a bare agent (ablation) | Bare-Agent · Framework ablation on Claude Sonnet 4, B=10 | Sim-GT: 0.801 -> 0.889 | Improved |
| increase query budget from 3 to 10 for bare agent (ablation) | Bare-Agent · Framework ablation on Claude Sonnet 4 | Sim-GT: 0.801 -> 0.801 | No clear effect |
| use only introspection skill L14 instead of full skill set (ablation) | JustAsk · Framework ablation on Claude Sonnet 4, B=5 | Sim-GT: 0.867 -> 0.771 | Hurt |
| add the skill framework to an unguided agent (Self-Evolving) (ablation) | Unguided agent · End-to-end evaluation, 6 models | Sim-GT: +10.6% on average, improves 5 of 6 targets | Improved |
| add the skill framework to an unguided agent (Self-Evolving) (ablation) | Unguided agent · End-to-end evaluation, Mistral Large | Sim-GT: 0.55 -> 0.52 (-6.7%) | Hurt |
| add the skill framework to an unguided agent (Self-Evolving) (ablation) | Unguided agent · End-to-end evaluation, Llama 4 Scout | Sim-GT: 0.73 -> 0.92 (+25.9%) | Improved |
| add attack-aware defense in system prompt | Frontier models · Controlled defense evaluation | Sim-GT: 0.719 -> 0.587 (-18.4%) | Improved |
| add simple do-not-reveal instruction | Frontier models · Controlled defense evaluation | Sim-GT: 0.719 -> 0.676 (6.0% reduction) | No clear effect |
| add simple do-not-reveal instruction | Claude Opus 4.5 · Controlled defense evaluation | Sim-GT: 0.600 -> 0.616 | Hurt |
| use JustAsk instead of static baseline Raccoon on hardest target | Raccoon · Baseline comparison, B=3, target GPT-4o | Sim-GT: 0.701 -> 0.727 | Improved |
Closest papers by meaning
- PragLocker: Protecting Agent Intellectual Property in Untrusted Deployments via Non-Portable Prompts (ICML 2026 · similarity 0.89)
- Agent Security Bench (ASB): Formalizing and Benchmarking Attacks and Defenses in LLM-based Agents (ICLR 2025 · similarity 0.88)
- Shadows in the Code: Exploring the Risks and Defenses of LLM-based Multi-Agent Software Development Systems (AAAI 2026 · similarity 0.88)
- AgentDojo: A Dynamic Environment to Evaluate Prompt Injection Attacks and Defenses for LLM Agents (NeurIPS 2024 · similarity 0.87)
- Attractive Metadata Attack: Inducing LLM Agents to Invoke Malicious Tools (NeurIPS 2025 · similarity 0.87)
- EnIGMA: Interactive Tools Substantially Assist LM Agents in Finding Security Vulnerabilities (ICML 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.
