MemCalib: Benchmarking and Optimizing Memory Use in LLM Agents
MemCalib-RL is proposed, an ordered bidirectional counterfactual credit-assignment algorithm that separates over- and under-use signals and localizes their credit to response tokens through exact atom ablation and achieves the best overall performance while better balancing over-use and under-use.
Rui-Ke Cao, Fan-Yu Zhao, Fu-Gen Yao et al.
· 0 citations