Skip to content

Runtime-control ablations

Each row changes one gMAS control while keeping the BBH samples paired. This separates graph and plan mutation from the policy that decides when to use it.

Entry point: python -m benchmarks.ablations.run.

Summary

Control Accuracy Δ Token Δ Latency Δ Executed agents Δ Statistical note
Early stopping +1.8 pp -52.6% -45.0% -3.0 Accuracy n.s., p = .70
Disabled-node handling +3.4 pp -44.4% -32.1% -2.0 Accuracy p = .002
Reachability filter +0.5 pp -23.1% -34.3% -2.0 Accuracy n.s., p = .90
Adaptive topology policy +3.2 pp -21.2% -18.5% -1.0 Accuracy n.s., p = .251
Hidden channels -2.9 pp +6.2% +7.2% 0.0 Accuracy p = .005
Multi-model routing +9.4 pp -4.6% +4.2% 0.0 Observed accuracy effect, p = .118

Variant means

Each cell is first variant / second variant, matching the comparison column.

Control Comparison Accuracy Latency (ms) Tokens Agents
Early stopping On / Off 74.1% / 72.3% 19,453±8,941 / 35,386±11,203 2,461±1,187 / 5,196±1,305 2.0±1.0 / 5.0±0.0
Disabled nodes On / Off 75.2% / 71.8% 25,022±8,512 / 36,847±12,983 3,063±768 / 5,508±1,574 3.0±0.0 / 5.0±0.0
Reachability filter On / Off 73.8% / 73.3% 36,822±11,204 / 56,087±24,163 5,603±1,392 / 7,289±2,198 5.0±0.0 / 7.0±0.0
Adaptive topology Adaptive / Static 74.8% / 71.6% 29,681±10,752 / 36,429±12,148 4,298±1,662 / 5,453±1,489 4.0±2.0 / 5.0±0.0
Hidden channels On / Off 69.2% / 72.1% 31,978±21,203 / 29,836±18,124 4,117±2,195 / 3,875±1,867 5.0±0.0 / 5.0±0.0
Multi-model routing Strong+weak / Weak-only 74.6% / 65.2% 37,682±6,953 / 36,147±11,489 5,271±913 / 5,528±1,382 5.0±0.0 / 5.0±0.0

Takeaways

  • Early stop, disabled-node handling, and reachability filtering are direct cost controls with the clearest operational effect.
  • The adaptive policy saves less than the direct stopping and filtering controls in this setup.
  • Hidden channels are unfavorable on both cost and accuracy here.
  • Multi-model routing behaves as a quality/cost trade-off, not as a speed optimization.

These results apply to the tested policies and configuration. Enabling every adaptive feature is not a general optimization strategy.