Title: Bridging the Gap: Evaluating Reflective Memory in Extended Conversations

Abstract:

While long-context modeling has advanced significantly, current evaluation metrics are largely restricted to factual memory, focusing exclusively on explicit recall. These benchmarks fail to assess reflective memory, which is essential for synthesizing fragmented, multimodal signals into sophisticated interpretations. To bridge this gap, we present RefMem-Bench, a new benchmark designed to evaluate reflective memory within long-horizon dialogues. The benchmark comprises 26,000 annotated question-answering instances, organized around eight dimensions of reflective memory and three distinct task formats. It demands that models transcend simple surface-level retrieval to infer latent meanings from evidence scattered throughout interaction histories. To bolster reflective memory capabilities, we introduce REflective Memory INDuction (REMIND), a hierarchical framework that conceptualizes reflective memory as an incremental process of meaning construction. REMIND integrates question-conditioned evidence retrieval, salience-aware grounding, and abstraction-level supervision, while employing Progressive Reflective Alignment to distill high-level reflective reasoning into the factual inference pathway. Our experiments demonstrate that RefMem-Bench presents a significant challenge to existing models. Meanwhile, REMIND consistently enhances both answer accuracy and memory recall by facilitating progressive evidence perception, grounding, and abstraction.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC