Moving Beyond Naive Consensus

In multi-agent systems, memory arbitration—the process of deciding which information to trust when agents hold conflicting data—is traditionally handled via simple majority voting or basic heuristic averaging. The authors argue that these methods are fundamentally flawed because they treat all memory sources as equally credible, ignoring the context of how that information was generated. Latent-Source Reasoning shifts this paradigm by treating memory arbitration as a reasoning task rather than a counting task.

The Latent-Source Reasoning Framework

Instead of aggregating raw outputs, the system analyzes the 'latent source' of each memory fragment. This involves:

• Provenance Tracking: Mapping the chain of inference that led to a specific memory entry.
• Reliability Weighting: Dynamically adjusting the influence of an agent's memory based on its historical accuracy in similar domains and the logical consistency of its reasoning path.
• Arbitration via Reasoning: Using a meta-agent or a reasoning layer to evaluate the competing claims. By examining the underlying evidence and logical steps provided by each agent, the system can identify which agent is more likely to be correct, even if that agent is in the minority.

Impact on System Reliability

By implementing Latent-Source Reasoning, developers can significantly reduce the 'hallucination propagation' that occurs when a majority of agents converge on an incorrect but plausible-sounding answer. This approach is particularly effective in high-stakes environments where the quality of information is more important than the quantity of agents agreeing on it. The framework effectively transforms memory from a static database into a dynamic, verifiable knowledge graph that evolves based on the quality of the reasoning performed by the constituent agents.