PMR
A personalized credit assignment framework for long-term memory AI agents that decomposes terminal feedback into token-level fine-grained signals through a two-tier scoring rubric (a general layer and a user-adaptive layer), enabling memory policy optimization with zero inference overhead.
Timeline (last 90 days)
PMR (Personalized Memory Rubric) is a framework that converts terminal feedback into fine-grained, step-by-step, per-user credit assignment.
PMR decomposes evaluation into two layers: a general layer that captures failure modes across user types, and a per-user adaptive layer that models individual characteristics as specialized increments on top of the general baseline.
Through rubric-guided credit assignment, PMR redistributes credit at the token level: selectively amplifying verified decisions on successful trajectories and falling back to standard policy optimization on failed trajectories.
PMR's rubrics and memory policy co-evolve in a closed loop, with anti-cycling guarantees.
At deployment, PMR requires no rubrics or external judges, achieving zero inference overhead.
Extensive experiments across four benchmarks show that PMR achieves consistent performance improvements over strong memory-augmented baselines.
PMR's rubrics are not manually set but co-evolve with the memory policy: the general layer captures common failure modes across users, while the user-adaptive layer learns individual preferences from a small amount of user interaction data.
All Facts (7)
PMR (Personalized Memory Rubric) is a framework that converts terminal feedback into fine-grained, step-by-step, per-user credit assignment.
50%UnverifiedPMR decomposes evaluation into two layers: a general layer that captures failure modes across user types, and a per-user adaptive layer that models individual characteristics as specialized increments on top of the general baseline.
50%UnverifiedThrough rubric-guided credit assignment, PMR redistributes credit at the token level: selectively amplifying verified decisions on successful trajectories and falling back to standard policy optimization on failed trajectories.
50%UnverifiedPMR's rubrics and memory policy co-evolve in a closed loop, with anti-cycling guarantees.
50%UnverifiedAt deployment, PMR requires no rubrics or external judges, achieving zero inference overhead.
50%UnverifiedExtensive experiments across four benchmarks show that PMR achieves consistent performance improvements over strong memory-augmented baselines.
50%UnverifiedPMR's rubrics are not manually set but co-evolve with the memory policy: the general layer captures common failure modes across users, while the user-adaptive layer learns individual preferences from a small amount of user interaction data.
50%