SIGMA: Signed Graph Modeling for Conflict-Resilient Multi-Agent Reasoning
A new framework called SIGMA (Signed Graph-informed Multi-Agent reasoning) addresses limitations in LLM-based multi-agent systems by explicitly modeling trust, conflict, and neutral relations among agents. Existing graph-based MAS frameworks propagate errors from conflicting signals and lack structural awareness. SIGMA constructs a signed interaction graph with confidence-weighted edges, enabling robust aggregation even under disagreement. The approach selects relevant and diverse agents for each query, then builds the graph to capture inter-agent dynamics. This improves reasoning and decision-making beyond naive cooperative aggregation.
Key facts
- SIGMA is introduced for conflict-resilient multi-agent reasoning.
- It uses a signed relational graph to model trust, conflict, and neutral relations.
- Existing graph-based MAS frameworks propagate errors from conflicting signals.
- SIGMA selects relevant and diverse agents per query.
- The signed interaction graph is confidence-weighted.
- The framework improves reasoning and decision-making.
- It addresses lack of explicit modeling of conflicting inter-agent relations.
- SIGMA stands for SIgned Graph-informed Multi-Agent reasoning.
Entities
—