Research map / Large language models / Retrieval and knowledge
When Search Goes Wrong: Red-Teaming Web-Augmented Large Language Models
Haoran Ou, Kangjie Chen, Xingshuo Han, Gelei Deng, Jie Zhang, Han Qiu, Tianwei Zhang, Kwok-Yan Lam
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 |
|---|---|---|---|
| use CREST-Search instead of the best baseline (SafeSearch) | web-search LLMs · 4 commercial search LLMs | risk detection rate: 32.4% -> 80.5% | Improved |
| fine-tune the red-teaming model on WebSearch-Harm instead of vanilla few-shot prompting (ablation) | CREST-Search red-teaming model · query generation | risk detection rate: 49.3% -> 80.5% | Improved |
| fine-tune the red-teaming model instead of vanilla zero-shot (ablation) | CREST-Search red-teaming model · query generation cost and time | cost per 100 queries (USD) and time (s): 4.1 -> 1.5 USD; 128.5 -> 49.6 s | Improved |
| fine-tune the red-teaming model instead of few-shot prompting (ablation) | CREST-Search red-teaming model · query toxicity | query toxicity: 12.2% -> 23.6% | Hurt |
| increase the maximum refinement rounds from 1 to 10 (ablation) | CREST-Search · harassment category | risk detection rate: 34.6% -> 73.8% | Improved |
| increase the maximum refinement rounds (ablation) | CREST-Search · harassment category | cost and time: 2.22 USD per 100 queries and 75.4 seconds at 10 rounds | Hurt |
| cyclically apply the three strategies across refinement rounds instead of a single strategy (ablation) | CREST-Search · multi-strategy variant | risk detection rate | No clear effect |
| use keyword injection versus exaggeration or role play (ablation) | CREST-Search · five harmful categories average | risk detection rate: 78.9% (exaggeration) -> 81.6% (keyword injection) | Improved |
| use persona-based query construction (ablation) | query generation · citation risk | citation risk rate: 10.3% | Hurt |
Closest papers by meaning
- Text-Diffusion Red-Teaming of Large Language Models: Unveiling Harmful Behaviors with Proximity Constraints (AAAI 2025 · similarity 0.88)
- Look Before You Leap: Enhance Attention and Vigilance Regarding Harmful Content with GuidelineLLM (AAAI 2025 · similarity 0.87)
- Cannot See the Forest for the Trees: Invoking Heuristics and Biases to Elicit Irrational Choices of LLMs (ICML 2025 · similarity 0.87)
- ReliabilityRAG: Effective and Provably Robust Defense for RAG-based Web-Search (NeurIPS 2025 · similarity 0.86)
- On the Vulnerability of Applying Retrieval-Augmented Generation within Knowledge-Intensive Application Domains (ICML 2025 · similarity 0.86)
- DiveR-CT: Diversity-enhanced Red Teaming Large Language Model Assistants with Relaxing Constraints (AAAI 2025 · similarity 0.86)
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.
