Research map / Large language models / Retrieval and knowledge
From Volume to Value: Preference-Aligned Memory Construction for On-Device RAG
Changmin Lee, Jaemin Kim, Taesik Gong
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 |
|---|---|---|---|
| add semantic-based coarse filtering alone (ablation) | Contriever RAG, Llama-3.1-8B · PrefWiki | preference-following accuracy (%): 40.88 -> 37.26 | Hurt |
| add semantic-based coarse filtering alone (ablation) | Contriever RAG, Llama-3.1-8B · PrefRQ | preference-following accuracy (%): 68.11 -> 69.00 | No clear effect |
| add coarse filtering alone: memory reduction (ablation) | Contriever RAG, Llama-3.1-8B · PrefWiki | memory usage (MB): 142.16 -> 1.83 | Improved |
| add preference-aligned fine verification to coarse filtering (ablation) | EPIC, Llama-3.1-8B · PrefWiki | preference-following accuracy (%): 37.26 -> 52.53 | Improved |
| add preference-aligned fine verification to coarse filtering (ablation) | EPIC, Llama-3.1-8B · PrefEval | preference-following accuracy (%): 27.89 -> 61.58 | Improved |
| add fine verification: memory (ablation) | EPIC, Llama-3.1-8B · PrefWiki | memory usage (MB): 1.83 -> 0.27 | Improved |
| add preference-guided query steering (ablation) | EPIC, Llama-3.1-8B · PrefWiki | preference-following accuracy (%): 52.53 -> 54.07 | Improved |
| add preference-guided query steering (ablation) | EPIC, Llama-3.1-8B · PrefELI5 | preference-following accuracy (%): 85.48 -> 87.95 | Improved |
| add preference-guided query steering (ablation) | EPIC, Llama-3.1-8B · PrefEval | preference-following accuracy (%): 61.58 -> 65.61 | Improved |
| add fine verification (F) to HippoRAG 2 (ablation) | HippoRAG 2, Llama-3.1-8B · PrefWiki | preference-following accuracy (%): +10.57%p | Improved |
| add coarse filtering to F in HippoRAG 2 (ablation) | HippoRAG 2 + F · PrefWiki | memory and indexing latency: 6.08x memory reduction, 8.31x indexing latency reduction | Improved |
| add query steering to HippoRAG 2 + C + F (ablation) | HippoRAG 2 + C + F · PrefWiki | preference-following accuracy (%): +6.49%p | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- NV-Embed-v2 — NV-Embed: Improved Techniques for Training LLMs as Generalist Embedding Models (ICLR 2025) · better on 4 of 4 shared results
- HippoRAG 2 — HippoRAG: Neurobiologically Inspired Long-Term Memory for Large Language Models (NeurIPS 2024) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Personalize-then-Store: Benchmarking and Learning Personalized Memory for Long-Horizon Agents (NeurIPS 2026 · similarity 0.88)
- AMemGym: Interactive Memory Benchmarking for Assistants in Long-Horizon Conversations (ICLR 2026 · similarity 0.88)
- MemArena: An Ego-Centric Benchmark for On-Device Agentic Personal Memory Assistants at Scale (NeurIPS 2026 · similarity 0.87)
- Mem-PAL: Towards Memory-based Personalized Dialogue Assistants for Long-term User-Agent Interaction (AAAI 2026 · similarity 0.87)
- Aligning LLMs by Predicting Preferences from User Writing Samples (ICML 2025 · similarity 0.86)
- How Does Personalized Memory Shape LLM Behavior? Benchmarking Rational Preference Utilization in Personalized Assistants (NeurIPS 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.
