Privacy-preserving Query Rewriting on DelegateCI-Bench General
100UtilizationDirect LLM
Evaluation results
| Method | Links | ||||||
|---|---|---|---|---|---|---|---|
| Direct LLMRewriter=raw query2026.06 | 100 | — | — | — | 89.8 | — | |
| PUFTRewriter=Qwen-2.5-7B, Local aggregator=Qwen-2.5-7B-Instruct2026.06 | 96 | 7.6 | 2.5 | 50.1 | 72.5 | 3.9 | |
| CI-guided reinforcement learning frameworkRewriter=Qwen-2.5-3B (shared), Local aggregator=Qwen-2.5-7B-Instruct2026.06 | 95.3 | 8 | 2.2 | 59.8 | 78.7 | 10.1 | |
| CI-guided reinforcement learning frameworkRewriter=Qwen-2.5-3B (shared), Local aggregator=Llama-3.1-8B-Instruct2026.06 | 94.3 | 8 | 2.2 | 59.8 | 80.5 | 6.6 | |
| PapillonRewriter=Qwen-2.5-7B, Local aggregator=Qwen-2.5-7B-Instruct2026.06 | 93.5 | 6.9 | 2 | 63.4 | 76.3 | 7.7 | |
| PUFTRewriter=Llama-3.1-8B, Local aggregator=Llama-3.1-8B-Instruct2026.06 | 92.7 | 5.1 | 2.4 | 40.6 | 77.8 | 3.8 | |
| PapillonRewriter=Llama-3.1-8B, Local aggregator=Llama-3.1-8B-Instruct2026.06 | 92.2 | 3 | 4.6 | 52.4 | 70.7 | 3.3 | |
| Local OnlyRewriter=original query, Local aggregator=Qwen-2.5-7B-Instruct2026.06 | 89.9 | — | — | — | 68.6 | 0 | |
| CI-guided reinforcement learning frameworkRewriter=Qwen-2.5-3B (shared), Local aggregator=Qwen-2.5-1.5B-Instruct2026.06 | 89.5 | 8 | 2.2 | 59.8 | 65.4 | 12.5 | |
| Presidio(PII Only)Rewriter=rule-based PII2026.06 | 89.2 | 89.6 | 53.7 | 96.2 | 76.7 | — | |
| Local OnlyRewriter=original query, Local aggregator=Llama-3.1-8B-Instruct2026.06 | 87.2 | — | — | — | 74 | 0 | |
| PapillonRewriter=Qwen-2.5-1.5B, Local aggregator=Qwen-2.5-1.5B-Instruct2026.06 | 80.3 | 38.5 | 25.5 | 61.5 | 57.7 | 4.8 | |
| PUFTRewriter=Qwen-2.5-1.5B, Local aggregator=Qwen-2.5-1.5B-Instruct2026.06 | 80 | 19.8 | 12 | 44.5 | 56.3 | 3.4 | |
| Local OnlyRewriter=original query, Local aggregator=Qwen-2.5-1.5B-Instruct2026.06 | 71.8 | — | — | — | 52.9 | 0 |