ActionCipher

ActionCipher tests whether a vision-language model can infer an episode-specific symbol-to-action mapping from five visual transition demonstrations and then emit the shortest symbol sequence that transforms a query start pose into its goal.

Dataset

• 3 grid sizes: 3 by 3, 4 by 4, and 5 by 5
• shortest plan length h: 1, 2, or 3
• 50 independent problems for every grid_size and h combination
• 450 independent problems total
• 2 symbol mappings per problem, giving 900 rows
• 12 separate model-input images plus one overview audit image per row
• one unique shortest answer per row
• 26 columns; internal renderer, pose, and duplicated audit columns are omitted

Split	Independent problems	Rows
discovery	225	450
validation	90	180
heldout_test	135	270
Total	450	900

Row structure

Each row contains five demonstrations. A demonstration consists of separate before and after images plus its symbol label. The query consists of separate start and goal images. The model should return answer and nothing else.

codebook, primitive_plan, and answer are gold fields for scoring and analysis. The overview image also contains the gold answer in its title. These gold fields and the overview must not be inserted into the normal model prompt.

Prompt

Infer what each symbol means from the demonstrations.
Apply symbols from left to right.
For the query, return exactly one line containing only the shortest valid symbol sequence.
Put one space between symbols. Stop immediately after the last required symbol;
do not explain and do not repeat symbols after the answer is complete.

Present demo_1_before, demo_1_after, then demo_1_label, and repeat through demo 5. Finally present query_start, query_goal, and ask for the answer.

Load

from datasets import load_dataset

dataset = load_dataset("Hanoi0126/ActionCipher")

See COLUMN_REFERENCE.md for the exact 26-column schema.