EquiMem: Calibrating Shared Memory in Multi-Agent Debate via Game-Theoretic Equilibrium
Researchers have introduced EquiMem, a novel inference-time calibration mechanism designed to secure shared memory in Multi-Agent Debate (MAD) systems. While MAD systems utilize shared memory for long-horizon reasoning, they are vulnerable to contamination from single corrupted entries, which standard debate processes fail to filter. Existing safeguards often rely on heuristics or Large Language Model (LLM) judgments, which share similar failure modes and ignore cross-agent dynamics. EquiMem addresses this by modeling memory updates as a zero-trust game where no agent is assumed honest. It uses the game's equilibrium to determine optimal memory trust, algorithmically quantifying updates based on agents' retrieval queries and traversal paths rather than soliciting new LLM judgments. Compatible with both embedding- and graph-based memory architectures, EquiMem consistently outperforms existing safeguards across diverse benchmarks. It demonstrates robustness against adversarial agents and incurs negligible inference overhead, offering a significant advancement in securing AI reasoning frameworks without relying on potentially flawed AI validation methods.
Wire timeline
EquiMem: Calibrating Shared Memory in Multi-Agent Debate via Game-Theoretic Equilibrium
Researchers have introduced EquiMem, a novel inference-time calibration mechanism designed to secure shared memory in Multi-Agent Debate (MAD) systems. While MAD systems utilize shared memory for long-horizon reasoning, they are vulnerable to contamination from single corrupted entries, which standard debate processes fail to filter. Existing safeguards often rely on heuristics or Large Language Model (LLM) judgments, which share similar failure modes and ignore cross-agent dynamics. EquiMem addresses this by modeling memory updates as a zero-trust game where no agent is assumed honest. It uses the game's equilibrium to determine optimal memory trust, algorithmically quantifying updates based on agents' retrieval queries and traversal paths rather than soliciting new LLM judgments. Compatible with both embedding- and graph-based memory architectures, EquiMem consistently outperforms existing safeguards across diverse benchmarks. It demonstrates robustness against adversarial agents and incurs negligible inference overhead, offering a significant advancement in securing AI reasoning frameworks without relying on potentially flawed AI validation methods.
cs.AI updates on arXiv.org