arXiv:2611.17610 · Parity-Aware Compute Evaluation (PACE)

PACE Studio

Audit multi-agent LLM traces against compute-matched single-agent baselines, inspect debate peer-influence dynamics, explore the 9-study corpus noise band, and size self-consistency pools locally in your browser.

Cite PACE in Your Paper
Sajjadi Mohammadabadi (2026). More Agents, More Samples, or More Thinking? Untangling Coordination, Ensembling, and Test-Time Compute in Multi-Agent LLMs.
Get Python Toolkit
Quick Demos:

Baseline Classification

Waiting

Load a scored trace to calculate its resource vector.

A · Single Pass
B · Sampled
C · Resource-Matched
D · Pareto Frontier

Resource Vector R

No data

Directional Feasibility

No data

Debate Dynamics, Flow Identity & Peer-Influence Lab

Tracks accuracy from first to final round, wrong-to-right (τic) vs. right-to-wrong (τci) flip rates, and the exact flow-identity equilibrium p*.

Load a debate log to inspect round-by-round dynamics.

Corpus & Noise-Band Explorer

Test whether reported MAS-vs-baseline accuracy differences exceed the paired binomial noise band ±1.96 SE(φ) under assumed paired correlation φ.

Load a contrast CSV or click "Load Demo: 9-Study Contrasts CSV".

Prospective Experiment Planner (ktok & Pool Sizer)

Compute the token-equivalent single-agent sample count ktok for a K-agent, T-round multi-agent protocol and size your benchmark sample count n before running.

Configure a budget.
Power calculator available through pace power.

Active Trace: Round Transitions & Peer Correctness

Empirical probability P(correct at t+1) conditioned on own correctness and number of correct peers at round t.

No data

PACE Card Preview & Export

Load a file to render the card.
Run PACE on AutoGen / LangGraph / CrewAI logs: git clone https://github.com/mahmoudsajjadi/pace && pace import-logs --format autogen --input trace.json --output trace.jsonl && pace card trace.jsonl