Research map / Adversarial robustness and security / Attacks on text and graphs
Does AI Reviewer See the Full Picture? Attacking and Defending Multimodal Peer Review
Xinyu Zhao, Rana Shahroz, Zhen Xu, Zhen Tan, Tianlong Chen
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 chunk-based embedding search instead of full-document LLM judge or classifiers for detecting injections (ablation) | PaperGuard defense · PaperGuard text injection detection | Detection accuracy (%): 66.70 -> 95.0 | Improved |
| Use global Embedding Classifier instead of chunk-based search (ablation) | defense · PaperGuard text injection detection | Recall (%): 92.86 -> 0.0 | Hurt |
| Chunk-based embedding search on visual attacks (ablation) | defense · PaperGuard visual attack detection | Detection accuracy (%): 35.60 -> 93.50 | Improved |
| Chunk-based search on real arxiv injections (ablation) | defense · 17 real arxiv papers | Recall (%): 64.7 -> 100.0 | Improved |
| Use larger, stronger instruction-following reviewers | AI reviewer · Prompt injection ASR | ASR: Claude-sonnet-4.5 0.80 vs DeepSeek-R1-Distill-Llama-8B 0.46 | Hurt |
| Apply APGD instead of PGD or C&W for visual attack (ablation) | MLLM reviewer · Visual learnable attack | Score inflation: 12.78 -> 14.11 (Janus-Pro-7B PGD to APGD) | Improved |
| Text prompt injection suppresses negative tags | Mistral-Small-3.1 · Prompt injection | #Neg tags shift: -3.26 | Improved |
| GCG attack on weaker reasoning model | DeepSeek-Llama-8B · GCG text attack | ASR: ASR 0.52 vs 0.78 (Qwen-3-8B); score shift 5.74 vs 3.33 | Mixed |
Closest papers by meaning
- Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews (ICML 2024 · similarity 0.88)
- PRISMM-Bench: A Benchmark of Peer-Review Grounded Multimodal Inconsistencies (ICLR 2026 · similarity 0.86)
- From Replication to Redesign: Exploring Pairwise Comparisons for LLM-Based Peer Review (NeurIPS 2025 · similarity 0.86)
- Agent Reviewers: Domain-specific Multimodal Agents with Shared Memory for Paper Review (ICML 2025 · similarity 0.86)
- Position: The Artificial Intelligence and Machine Learning Community Should Adopt a More Transparent and Regulated Peer Review Process (ICML 2025 · similarity 0.85)
- Format Matters: The Robustness of Multimodal LLMs in Reviewing Evidence from Tables and Charts (AAAI 2026 · 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.
