Vijay K. Madisetti

When Too Many Cooks Spoil the Broth: Three Failure Modes of Multi-Agent LLM Reliability

Multi-agent LLM systems are widely deployed under three implicit assumptions: that self-confidence calibration transfers to peer evaluation, that errors made by distinct instances are statistically independent, and that any sufficiently calibrated model is an equivalent verifier....