Multi-agent debate — Du et al. 2023, convergence dynamics
round 0
round 1
round 2
round 3
A: 38.0
B: 42.5
C: 51.0
A: 42.89
B: 44.49
C: 45.34
A: 44.15
B: 44.56
C: 44.59
A: 44.43
B: 44.52
C: 44.50
spread 13.0
spread 2.45
spread 0.44
spread 0.09
Du et al. ablations
agents-only plateaus · rounds-only barely helps · agents × rounds produces the big jump
failure modes to guard against
sycophancy cascade: agents defer to loudest confidence, may converge on wrong answer
topic drift: long debates wander from original question — re-inject every round
cost: N × R calls with growing context; cap N=5, R=3