MASC / README.md
jwyjohnab's picture
Update README.md
0644d03 verified
metadata
license: unlicense
task_categories:
  - question-answering
language:
  - en
tags:
  - agent
pretty_name: Million Agent Sorting Challenge
size_categories:
  - 1M<n<10M

Million Agent Sorting Challenge (MASC)

Benchmarking Large Scale Multi-Agent Systems Through Coordinated Sorting


Overview

The Million Agent Sorting Challenge (MASC) is a benchmark designed to study scalability, coordination, and communication in large-scale multi-agent systems (MAS). Each agent receives a shuffled subset of integers and must cooperate with others through communication channels (e.g., shared blackboard, messaging, or tool calls) to collectively produce a globally sorted list.

MASC evaluates how efficiently an MAS can perform distributed computation, state synchronization, and coordinated decision-making as the number of agents grows from 10¹ to 10⁶.


Task Definition

Each instance defines:

  • n: number of agents
  • k: number of integers per agent
  • xs[i]: input list for agent i
  • ys[i]: output list expected from agent i

Agents collectively must satisfy:

def check_results(n, k, xs, ys):
    xl = sorted(sum(xs, []))
    yl = sum(ys, [])
    return xl == sorted(yl)

Dataset Structure

Each JSON file (task_n{n}_k{k}.json) contains:

{
  "n": 1000000,
  "k": 10,
  "agents": [
    {
      "agent_id": 0,
      "prompt": "... task description ...",
      "input_data": [...],
      "expected_output": [...]
    },
    ...
  ]
}

Data sizes scale from n = 10¹ up to n = 10⁶, allowing controlled experiments on MAS scalability and emergent coordination.


Intended Use

This dataset can be used for:

  • Studying scaling laws of cooperation
  • Designing communication protocols and distributed reasoning architectures
  • Benchmarking multi-agent emergent behavior at increasing system sizes

Citation

Preprint coming soon.