UpliftMem: Learning Set-Level Uplift for Agent Memory Retrieval
Mengkun Liang, Haoran Qiang, Guannan Liu, Junjie Wu
原文摘要
Large language model (LLM) agents draw on external memory to guide new tasks, raising the question of which memories to retrieve together to improve execution. Learning set values requires outcome feedback, but evaluating each untested alternative costs an additional rollout. We introduce \textsc{UpliftMem}, which learns set-level execution uplift relative to the same executor without memory. Our analysis of how retrieval preferences limit feedback coverage motivates targeted probes. Under a correlated Gaussian surrogate, we derive a closed-form expected value of sample information (EVSI) criterion that prioritizes probes by their expected benefit to local retrieval decisions. The feedback trains a shared scorer with the executor frozen, and inference requires no probing. Across ALFWorld, WebShop, and BigCodeBench, \textsc{UpliftMem} achieves best or joint-best success rates among evaluated methods. On fixed-store MemSyco-Bench evaluations, decision accuracy increases from 41.7\% to 66.5\% with Qwen3-4B and from 44.4\% to 68.8\% with Qwen3-8B. At matched probe budgets, EVSI improves decision accuracy over random probing by 10.6 and 13.9 percentage points, respectively.
cs.AI