Reproduce “$Σ$-Mem: An Online Reliability Memory for LLM-based Multi-Agent Systems” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Peilin Feng, Suorong Yang, Soujanya Poria
Researchers at Nanyang Technological University introduced Σ-Mem, an online reliability memory for LLM-based multi-agent systems, designed to model and accumulate trustworthiness evidence rather than just historical content. This system improves peer selection accuracy, for example, boosting Qwen3-0.6B's performance from 46.22% to 71.10% on counterfactual reliability shifts and showing generalization across unseen peers and out-of-distribution tasks.
Key insights
- Modeling explicit peer reliability (who to trust, and when) is crucial for effective multi-agent coordination, moving beyond simple content-based memory.
- Representing reliability evidence as symmetric matrices with bounded updates enables stable online learning, preventing single events from distorting long-term accumulated evidence.
- Combining individual peer competence with inter-peer relationship evidence provides complementary trust signals that enhance decision-making across varied task contexts.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.