Coupled Spiking Memory for Budgeted Long-Dialogue Question Answering with Small Language Models
Year: 2026
Type: preprint
Abstract: Longitudinal conversational assistance requires access to information that is scattered across sessions and changes over time. A small language model can retrieve this information from an external log, but independently ranking each record ignores the sequence in which the record was produced. We present Coupled Spiking Memory (CSM), an evidence-supervised log controller that augments each record with a causal temporal trace before query-dependent retrieval. CSM adapts two-compartment spiking dynamics, reads the post-reset somatic voltage, and learns a residual correction to lexical relevance while leaving the language model frozen. Whole-record selection restricts the retrieved context to a fixed token budget and preserves chronological presentation. With a frozen 3B reader, CSM obtains 30.33 mean QA F1 on 192 EvoEmo questions across five training seeds, exceeding BM25, BGE-M3, and a matched GRU controller by 0.31, 1.28, and 0.76 points, respectively. It uses 123,907 trainable controller parameters. Frozen transfer to LoCoMo and a three-budget analysis reveal that the benefit depends on the readout, evidence budget, and balance between answering and abstention. These results establish a compact spiking-controller approach to conversational evidence retrieval and identify the conditions that limit its transfer.
Source: Qeios
DOI: 10.32388/89xyao
Authors: Yiwen Wang, Jiahao Qin
Institutions: Fudan University, University of Liverpool
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Keywords: spiking neural networks, token budget, LoCoMo
Open Access status: diamond
OpenAlex ID: W7235690358
