TRACE-Memory: Selective Personalization via Utility-Aware Evidence Admission
A recent paper on arXiv (2608.08446) presents TRACE-Memory, a dual-phase framework designed for tailored generation that utilizes personal memory selectively, only when it enhances the response beyond what is available publicly. The framework identifies user-specific details absent from the request and public context, then incorporates a concise set of traceable evidence units based on the incremental utility of the response. The training process includes structured SFT initialization, a reduced-space GRPO warm-up in stages, and a nested multi-sample Joint GRPO. Evaluation covers 4,500 tasks from Controlled and Natural categories sourced from Goodreads and Amazon datasets. This framework tackles issues related to relevant history that may involve incorrect preferences, redundant public information, or inadequate support.
Key facts
- TRACE-Memory is a two-stage framework for selective personalization.
- Stage 1 queries for user-specific information missing from the request and public context.
- Stage 2 admits a compact subset of source-traceable evidence units or the empty set.
- Training uses structured SFT initialization, reduced-space stage-wise GRPO warm-up, and nested multi-sample Joint GRPO.
- Evaluation includes 4,500 Controlled and Natural tasks from Goodreads and Amazon datasets.
- The paper is available on arXiv with ID 2608.08446.
- The framework argues personal memory should be used only when it adds utility beyond a public-only response.
- The announcement type is 'new'.
Entities
Institutions
- arXiv
- Goodreads
- Amazon