MultiNet v2.0-Gridworld
Agentic Behavior in 2D Mazes: Previewing a Multi-Domain Benchmark
We put 3 highly capable models in 50 simple 2-dimensional mazes each. Claude Opus 4.8, Kimi K2.6, and Qwen 3.6-27B together solved just 6 out of 150 maze runs, and 45 out of the 50 mazes were not solved by any model.
A few examples of models failing the mazes
An open science initiative with contributions from research teams at:
MultiNet milestones over the past 2 years:
Frontier Model Performance
How far did the models get in each of the mazes?
Each column is a maze and each row a model, with the mazes ranked left to right by how many actions the shortest solution takes. Colour is how far toward the goal the model got, and the stars mark the six solves - all of them among the shortest mazes.
Do you think you can beat frontier models on 2D mazes?
Try out some of the mazes the models were evaluated on, and see how you fare!
MultiNet v2.0-Gridworld
Can you solve what frontier models cannot?
| Player | Result |
|---|
The Evaluation Environment: Why Mazes?
With MultiNet v2.0 we aim to build interactive environments that represent real-world scenarios. At the same time we are keen to keep the setup controllable, which will allow us to deterministically vary parameters in our environment in order to make it easier or more difficult for models to succeed in.
The capabilities we aim to benchmark are long-horizon action taking and causal reasoning, which involves various sub-capabilities such as planning, action execution, error recovery, visual object association, and so much more.
A simple underlying substrate that brings all these aspects together for an environment and benchmarking task is a maze with mechanisms. Moreover, this setup helps us identify and separate different failure modes in agents, such as not working out how to operate a mechanism, messing up the order of operation of mechanisms, or losing its way and not recovering - each of which requires a very different fix. Four properties that mazes as an underlying representation possess, which are highly valuable to us as we build towards this vision:
Knowledge-free
No prior knowledge required - no API familiarity, no jargon, and no domain-specific rules. Failures here cannot be blamed on unfamiliarity of interface.
Generative
No fixed answer key to contaminate, and difficulty scales with the models instead of saturating.
Projectable
The same maze can be re-rendered in language or 3D or many other domains, quantifying cross-domain generalization. For the full version of the benchmark, we will be expanding to multiple other domains.
Verifiable
BFS gives the exact optimal path from any state - an objective difficulty scale, partial credit, and the ability to call one move strictly wrong.
How does your agent do on cross-domain, multimodal, long-horizon tasks?
If you build models or agents, or work on benchmarking and evaluation, we would love to hear from you - whether that means getting your model on our MultiNet v2.0 benchmark, contributing to the environments we are building, or working with us on what comes after!
Citation
Please cite this work as:
Guruprasad, P., Rivera, S., Lu, H., Jain, A., Ren, H. and Sikka, H. (2026) Agentic behavior in 2D mazes: previewing a multi-domain benchmark. Available at: https://www.fig.inc/agentic-behavior-in-2d-mazes-previewing-a-multi-domain-benchmarkOr use the BibTeX citation
@online{multinet_v2_gridworld_technical_report_2026,
title = {Agentic behavior in 2D mazes: previewing a multi-domain benchmark},
author = {Pranav Guruprasad and Sean Rivera and Helen Lu and Arushi Jain and
Hangliang Ren and Harshvardhan Sikka},
year = {2026},
url = {www.fig.inc/agentic-behavior-in-2d-mazes-previewing-a-multi-domain-benchmark},
note = {MultiNet v2.0, GridWorld domain}
}