Research map / Large language models / Language, safety and interpretability
CIMemories: A Compositional Benchmark For Contextual Integrity In LLMs
Niloofar Mireshghallah, Neal Mangaokar, Narine Kokhlikyan, Arman Zharmagambetov, Manzil Zaheer, Saeed Mahloujifar, Kamalika Chaudhuri
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 |
|---|---|---|---|
| increase number of tasks from 1 to 40 | GPT-5 · CIMemories | Violation@1: 0.1% -> 9.6% | Hurt |
| sample 5 generations instead of 1 for the same prompt | GPT-5 · CIMemories | Violation: 9.6% -> 25.1% | Hurt |
| scale Qwen-3 model size from 1.7B to 32B (ablation) | Qwen-3 family · CIMemories | Violation@5 and completeness | Mixed |
| enable reasoning chain (ablation) | Qwen-3 30B · CIMemories | Violation@5 | Improved |
| enable reasoning chain (ablation) | Qwen-3 30B · CIMemories | Completeness | No clear effect |
| add conservative privacy prompt (medium, high) (ablation) | GPT-5 · CIMemories | Violation@5 | Improved |
| add conservative privacy prompt (medium, high) (ablation) | GPT-5 · CIMemories | Completeness | Hurt |
| add light (task-success-emphasizing) prompt (ablation) | GPT-5 · CIMemories | Completeness | Improved |
| increase number of inappropriate attributes in memory (ablation) | GPT-5 · CIMemories | Violations@5 | Hurt |
| increase number of inappropriate attributes in memory (ablation) | GPT-5 · CIMemories | Completeness | No clear effect |
| choose GPT-4o over Qwen-3 32B | model choice · CIMemories (10 profiles) | Violation@5 vs completeness: 14.82% vs 69.14% violation; 43.95% vs 57.63% completeness | Mixed |
Closest papers by meaning
- Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-Horizon Agents (NeurIPS 2026 · similarity 0.88)
- Position: Contextual Integrity is Inadequately Applied to Language Models (ICML 2025 · similarity 0.87)
- Can LLMs Keep a Secret? Testing Privacy Implications of Language Models via Contextual Integrity Theory (ICLR 2024 · similarity 0.87)
- CASE-Bench: Context-Aware SafEty Benchmark for Large Language Models (ICML 2025 · similarity 0.87)
- It Takes Two: Complementary Self-Distillation for Contextual Integrity in LLMs (NeurIPS 2026 · similarity 0.87)
- AMemGym: Interactive Memory Benchmarking for Assistants in Long-Horizon Conversations (ICLR 2026 · 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.
