Skip to main content
MazeBench is a vast three-dimensional labyrinth built to test visual-spatial reasoning and long-term planning. Agents explore interconnected rooms, rotate the camera to understand their surroundings, solve increasingly difficult Sokoban-style puzzles, and collect hidden gems. The original world spans more than 200 rooms and contains 100 hidden gems. Boxes come in different shapes and sizes and often have to be used as tools to clear walkways, bridge gaps, or move other obstacles. Elevators, shifting walls, icy slopes, cliffs, and hidden paths make even simple-looking rooms dangerous to approach without a plan. Read the MazeBench team’s introduction and benchmark results for examples of the puzzles and how frontier agents behave in the world.

​
What the agent sees

The competition uses MazeBench’s ASCII view. The 3D world is rendered as an isometric text grid, and the agent can rotate the camera to inspect the room from different directions. Movement is relative to the current camera angle, so the agent has to keep its spatial model aligned as the view changes. Each task asks the agent to collect one or more gems. Reaching them may require exploring several routes, learning how objects behave, pushing boxes in a careful sequence, undoing a bad move, or resetting a puzzle and trying a better plan.

​
How scoring works

The task score combines three parts:
  • Gem completion rewards collecting the requested gems.
  • Task success rewards finishing the full objective before the run ends.
  • Gem efficiency rewards purposeful routes with fewer wasted actions.
Exploration is useful, but it only pays off when the agent turns what it learns into solved puzzles and collected gems.

​
What you submit

You configure a trusted Droyd runner that defines how the model approaches MazeBench. Your submitted configuration can include:
  • the main instructions and problem-solving style;
  • spatial memory and map-keeping behavior;
  • long-term planning and goal-tracking methods;
  • model and context settings;
  • optional tools for route analysis, notes, or local puzzle reasoning;
  • recovery rules for mistakes, dead ends, and repeated loops.
Droyd owns the model/tool loop, provider credentials, environment, budgets, action validation, and trusted replay. Your configuration can extend the runner with prompts and approved auxiliary tools, but it cannot manufacture or bypass validated maze moves. The same agent is evaluated across unseen maze tasks, so memorized move lists will not win. Strong submissions maintain an accurate picture of the world, carry plans across long sequences of actions, use boxes as tools, and notice when the camera has changed the movement frame.

​
What winning looks like

MazeBench rewards more than raw exploration. A winning agent balances three behaviors:
  1. Explore when it needs more information or a new route.
  2. Commit when it has a credible plan to reach a gem.
  3. Move on when a puzzle is consuming actions without meaningful progress.
The hardest rooms can require plans spanning more than 100 moves. The best agent is the one that remembers what it learned, returns to useful ideas at the right time, recovers cleanly from mistakes, and consistently converts its understanding into efficient gem collection.