Latent Preference Modeling on MPT Context-Free, Preference Recall
76.1PrecisionPREFINE
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| PREFINEBase LLM=Gemma-3-12B2026.04 | 76.1 | 63.66 | 69.3 | |
| PREFINEBase LLM=GPT-52026.04 | 74.46 | 82.99 | 78.41 | |
| PREFINEBase LLM=GPT-5-mini2026.04 | 73.23 | 83.43 | 77.9 | |
| PREFINEBase LLM=Gemini-3-Flash2026.04 | 71.45 | 85.37 | 77.75 | |
| LangMemMethodology=Memory-Augmented, Backbone=Gemini-3-Flash2026.04 | 69.25 | 86.7 | 77 | |
| Gemini-3-FlashMethodology=Base Prompting2026.04 | 63.27 | 87.81 | 73.55 | |
| PREFINEBase LLM=GPT-4o-mini2026.04 | 62.11 | 66.7 | 64.25 | |
| GPT-5-miniMethodology=Base Prompting2026.04 | 61.42 | 88.64 | 72.56 | |
| GPT-5Methodology=Base Prompting2026.04 | 59.39 | 86.7 | 70.5 | |
| RAG (Top-5)Methodology=Memory-Augmented, Backbone=Gemini-3-Flash2026.04 | 52.42 | 60.11 | 56 | |
| Mem0Methodology=Memory-Augmented, Backbone=Gemini-3-Flash2026.04 | 52.36 | 55.4 | 53.84 | |
| GPT-4o-miniMethodology=Base Prompting2026.04 | 50.09 | 76.18 | 60.44 | |
| Gemma-3-12BMethodology=Base Prompting2026.04 | 47.78 | 38.78 | 42.81 | |
| PREFINEBase LLM=R1-Distill-Llama-8B2026.04 | 44.72 | 71.3 | 54.95 | |
| PREFINEBase LLM=R1-Distill-Qwen-7B2026.04 | 36 | 57.23 | 44.19 | |
| PREFINEBase LLM=CodeGemma-7B2026.04 | 35.4 | 81.22 | 49.31 | |
| R1-Distill-Llama-8BMethodology=Base Prompting2026.04 | 32.29 | 71.47 | 44.48 | |
| R1-Distill-Qwen-7BMethodology=Base Prompting2026.04 | 21.12 | 56.51 | 30.75 | |
| CodeGemma-7BMethodology=Base Prompting2026.04 | 19.63 | 67.31 | 30.39 |