Tile Match Flowers is a matching game played on a grid of stacked flower tiles. Every cell holds several layers at once, drawn one inside another: a wide backing shape at the bottom, and a smaller flower on each layer above it.
Your Mission
Clear the board by matching pairs of tiles. This Gamelab ruleset plays level 6 — the last level the game ships — which is six columns by five rows with four layers in every cell: 120 layers, and 60 matches to clear them all.
How a turn works
- Pick two different cells that still hold tiles.
- Layers are compared by depth, never across depths: top against top, second against second. A backing shape never matches a flower.
- A match clears that one layer from both cells, and everything else stays where it is.
- After a match the second tile stays held and outlined, and cannot be picked again on the very next turn.
- Each match is worth 100 points — there is no speed bonus and no chain bonus.
Read the picture
This is the one Gamelab game the model plays by sight. The board is rendered on the server and sent as an image; the text alongside it gives the grid size and the rules and nothing else — not which cells hold tiles, and not what any of them shows. Working that out by looking is the whole task.
Tile Match Flowers as an Evaluation Tool
Tile Match Flowers is the vision benchmark in the set. Every other game hands the model a symbolic board; this one hands it a picture of the board and nothing else, so it measures whether a model can perceive game state rather than parse it.
The discrimination it asks for is fine-grained. Tiles share backgrounds and differ only in the flower, or share the flower and differ in the backing, and matching is by depth, so a model must also judge which ring of a nested stack it is looking at. A three-arm ablation — the same board seen, described in text, and chosen at random — is what separates genuine visual reading from lucky guessing.
Harness and Structure
The benchmark runs the Tile Match Flowers environment as a game server that composites the game's own tile art into a board image, which is sent to the model as part of the turn.
Each prompt contains:
- The rendered board image — the model's only view of what is on the table
- The grid size, the starting stack depth, and the rules in force
- The score and the turn number
The model answers with one JSON action naming two cells by row and column. The server decides the match, clears at most one layer per selection, and returns the new board.
Illegal selections — an empty cell, the same cell twice, a tile held from the previous turn — are rejected and recorded on the run export, which also carries the images so a replay shows exactly what the model was looking at.
Evaluation
Episodes run to a 150-turn cap on seeded boards, chosen so a model playing at a realistic mismatch rate can still clear all 60 matches rather than being cut off by the cap.
Performance is read from layers cleared, matches made, the mismatch rate, the final score, and the number of rejected selections.
Notes
- Vision: the board reaches the model as an image, with no text map of its contents.
- Perfect information: every remaining layer is visible in the picture — if it can be seen.
- Fine-grained discrimination: tiles differ by details of flower and backing shape.
- Depth reasoning: layers match only against the same depth in the other tile.
- Long-horizon planning: clearing the board takes 60 successful matches.
- Deterministic: the seed fixes the deal, so runs are reproducible.
Contact Us
Do you want to know more about the project?