Personalized Question Answering on LaMP-QA
53.27Accuracy (Arts & Entertainment)Upper Bound: RAG-Personalization Server-Side
Evaluation Results
| Method | Links | ||||
|---|---|---|---|---|---|
| Upper Bound: RAG-Personalization Server-SideModel Family=Gemma 3, Parameters=12B2026.01 | 53.27 | 67.22 | 71.41 | 63.96 | |
| P3Model Family=Gemma 3, Server-Side Parameters=12B, Client-Side Parameters=4B2026.01 | 48.11 | 61.99 | 65.92 | 58.67 | |
| RAG-Personalization Client-SideModel Family=Gemma 3, Parameters=4B2026.01 | 42.58 | 57.65 | 60.13 | 53.45 | |
| Non-Personalized Server-SideModel Family=Gemma 3, Parameters=12B2026.01 | 41.32 | 59.54 | 62.99 | 54.61 | |
| Upper Bound: RAG-Personalization Server-SideModel Family=Qwen 2.5, Parameters=14B2026.01 | 39.2 | 51.07 | 57.43 | 49.23 | |
| Non-Personalized Speculative DecodingModel Family=Gemma 3, Verifier Parameters=12B, Draft Parameters=4B2026.01 | 39.12 | 57.99 | 59.23 | 52.11 | |
| P3Model Family=Qwen 2.5, Server-Side Parameters=14B, Client-Side Parameters=3B2026.01 | 36.8 | 48.88 | 53.06 | 46.24 | |
| Non-Personalized Server-SideModel Family=Qwen 2.5, Parameters=14B2026.01 | 32.66 | 45.87 | 48.7 | 42.41 | |
| Non-Personalized Speculative DecodingModel Family=Qwen 2.5, Verifier Parameters=14B, Draft Parameters=3B2026.01 | 32.26 | 44.96 | 47.7 | 41.64 | |
| RAG-Personalization Client-SideModel Family=Qwen 2.5, Parameters=3B2026.01 | 30.13 | 42.78 | 45.67 | 39.52 |