vanyacohen/MET-Bench-Chess

MET-Bench: Multimodal Entity Tracking for Evaluating the Limitations of Vision-Language and Reasoning Models

Vanya Cohen and Raymond Mooney · ICML 2026

Paper · Publication page · Load the dataset · Citation

Domains: Chess · Shell Game · Minecraft

MET-Bench evaluates entity state tracking across text and image modalities. This repository contains the Chess domain.

Chess

Chess is an entity state tracking task in which a model follows the positions of pieces through a sequence of legal moves. Given the initial board state and the actions, the goal is to recover the board state after the final move. The trajectories are drawn from MillionBase games.

Parallel text and image modalities

Each trajectory contains two aligned representations of its actions: text-encoded actions in actions and image-encoded actions in image_actions. The entries actions[i] and image_actions[i] represent the same move at the same step.

Representation	Field	Contents
Text	actions	Moves in Universal Chess Interface (UCI) notation, such as e2e4.
Image	image_actions	Board images marking the move's source square in green and destination square in red.

Both representations are stored together in each row and share the same initial_state, states, and final_state. States are strings in Forsyth–Edwards Notation (FEN), including piece placement, side to move, castling rights, en passant target, and move counters.

The action images show the highlighted squares on an empty board. The piece positions are tracked from the initial state and preceding actions. For promotions, the UCI action includes a suffix identifying the promoted piece, such as q for a queen; the image encodes the source and destination squares.

Dataset splits

The full release contains 99,770 training trajectories, 3,019 validation trajectories, and 3,048 test trajectories. Each trajectory contains 100 actions and the board state after every action. One action is one player's move (one ply).

Configuration	Actions per example	Train	Validation	Test
full (default)	100	99,770	3,019	3,048

The full 100-action trajectories are disjoint across splits. The training split contains 99,763 distinct trajectories across 99,770 rows; validation and test trajectories are unique within each split. Source row order and repeated training examples are preserved.

This release contains the full trajectories. The paper's experiments use deduplicated subsets at selected action-sequence lengths.

Evaluation subset

The evaluation configuration contains a single 500-example test split for benchmark evaluation. Examples are the first 500 unique (initial_state, actions[:10]) inputs in the full test split, in source order. Each example contains ten actions, ten aligned action images, and eleven states; final_state is the state after the tenth action.

The evaluation_text_only configuration contains the same examples in the same order, with image fields omitted for text-only evaluation. Both configurations preserve the source example_id values and all text fields. The full configuration remains the default.

from datasets import load_dataset

test = load_dataset("vanyacohen/MET-Bench-Chess", "evaluation", split="test")
text_test = load_dataset("vanyacohen/MET-Bench-Chess", "evaluation_text_only", split="test")

Usage

from datasets import load_dataset

dataset = load_dataset("vanyacohen/MET-Bench-Chess", "full", streaming=True)
example = next(iter(dataset["test"]))

initial_state = example["initial_state"]
text_actions = example["actions"]
image_actions = example["image_actions"]  # decoded PIL images
target = example["final_state"]

This example streams one trajectory at a time. final_state contains the target FEN. states contains the complete sequence of board states, including the initial state.

Data fields

The dataset is stored in Parquet shards with text actions and embedded images in each row.

Field	Type	Description
example_id	string	Stable identifier for the source split and row.
initial_state	string	Initial board state in FEN.
actions	list of strings	Ordered moves in UCI notation.
image_actions	list of Image	One embedded 400×400 PNG per action, aligned with actions.
states	list of strings	Initial FEN followed by the FEN after each action.
final_state	string	Board state in FEN after the final action.

Rows in full contain 100 text actions, 100 images, and 101 states. Rows in evaluation contain 10 text actions, 10 images, and 11 states. Action i, represented by both actions[i] and image_actions[i], takes the board from states[i] to states[i + 1]. The first state equals initial_state, and the last state equals final_state.

For illustration, two opening moves have:

initial_state = "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1"
actions = ["e2e4", "e7e5"]
states = [
    initial_state,
    "rnbqkbnr/pppppppp/8/8/4P3/8/PPPP1PPP/RNBQKBNR b KQkq - 0 1",
    "rnbqkbnr/pppp1ppp/8/4p3/4P3/8/PPPP1PPP/RNBQKBNR w KQkq - 0 2",
]
final_state = states[-1]

The two corresponding images highlight e2 → e4 and e7 → e5. The released trajectories follow this structure with 100 actions in full and 10 actions in evaluation.

Citation

@inproceedings{cohen2026metbench,
  title={MET-Bench: Multimodal Entity Tracking for Evaluating the Limitations of Vision-Language and Reasoning Models},
  author={Cohen, Vanya and Mooney, Raymond},
  booktitle={International Conference on Machine Learning},
  year={2026},
  url={https://arxiv.org/abs/2502.10886}
}

License

MIT.