Skip to content

Benchmarks and reported results

These pages record the evaluation snapshot from the 2026 gMAS manuscript. The paper tracks and the independent Terminal-Bench harness are listed separately:

Track Question Scope Details
Matched topology How do gMAS and LangGraph behave under the same roles, models, prompts, and graph shapes? BBH, GSM8K, MMLU-Pro; 68 matched model–dataset–topology configurations Results
Runtime controls Which gMAS controls actually change cost or observed accuracy? Six paired ablations on BBH Results
GAIA tool use What happens under a fixed five-tool budget? GAIA validation, levels 1–3, 165 tasks Results
AgentPrune Does learned graph pruning improve the cost-penalized research-idea score? MultiAgentBench research-idea track; 3 train and 97 held-out tasks Results
Terminal-Bench How does a conditional gMAS team compare with a single gMAS agent on terminal tasks? Terminal-Bench 2.0, 89 tasks Harness available; no complete matched result is reported yet

Result snapshot

The values below were transcribed from the manuscript generated on 21 July 2026. They describe those runs only; model, provider, task, and graph choices can change the outcome.

  • In the matched topology aggregate, gMAS latency was 8.0% lower for fan-in and 9.3% lower for fan-out. Chain-3 latency was 1.0% higher. Aggregate token changes were small and mixed.
  • Early stopping was the strongest direct cost control in the ablation: 52.6% fewer tokens, 45.0% lower latency, and 3.0 fewer executed agents on average, with a non-significant observed accuracy change.
  • On GAIA, gMAS used fewer tokens and achieved higher accuracy in both reported search settings, while taking longer. This is a correctness/budget result, not a latency win.
  • AgentPrune increased the cost-penalized mean score from 0.735 to 0.822 on the fixed held-out set. The mean change across 14 pruned realizations was +0.087, with a reported 95% confidence interval of [+0.023, +0.150].

How to read the numbers

Negative latency or token deltas mean lower operational cost for gMAS. Accuracy deltas are percentage points, not percentages. Unless a statistical test is shown, an accuracy difference is an observed effect only.

The runs do not support a blanket claim that gMAS is faster. They show that:

  1. topology controls can avoid work when a graph contains optional or unreachable roles;
  2. the largest savings come from explicit stopping and filtering controls;
  3. aggregate token use can increase in some topologies;
  4. tool-enabled runs may trade latency for accuracy and lower token consumption;
  5. learned pruning remains task-, reward-, and configuration-dependent.

See reproducibility for the run-artifact contract and benchmarks/ for executable entry points.