Counterfactual Graph for Multi-Agent LLM Calibration

In three linesMulti-agent LLM systems assume agreement between agents indicates reliability. Authors show communication induces correlated failures and false consensus. They propose CAGE-CAL, a counterfactual agent-graph calibration framework comparing post-communication dependencies with no-communication scenarios to adjust confidence accordingly.

Summary generated by Claude — human-verified