CIPHERGRID

CIPHERGRID Benchmark: From Multimodal Rule Inference to Sequential Action
Christopher Curtis, Victor Fragoso, and Saiph Savage

Accepted at NeurIPS 2026 — Evaluations and Datasets Track.

Code: GitHub repository
Paper: [Link Posted Upon Publication]

Overview

CIPHERGRID tests whether AI models can discover the unfamiliar relationships connecting familiar reasoning skills. Given eight fixed examples pairing grid-world images with encoded descriptions and solutions, a model must infer the symbolic vocabulary and game rules, reconstruct a new encoded world, plan a valid route to its goal, and express the solution in the same encoded language.

The central challenge is interactional generalization: carrying an inferred symbolic system through multimodal grounding, rule application, and sequential planning. The rules and symbol meanings are inferred from the demonstrations rather than supplied explicitly in the standard evaluation setting.

The benchmark contains 635 procedurally generated problems, including 535 solvable and 100 unsolvable instances. All records share the same eight demonstrations, evidence image, and canonical vocabulary. The mapping is fixed across records; it is not resampled for each problem.

Results reported in the NeurIPS 2026 paper

Across six evaluated models, overall accuracy ranged from 18.74% to 42.20%, with performance declining sharply as grid size increased.

Model	Overall	Baseline	Small	Medium	Large
GPT-5.4	42.20%	90.00%	68.63%	25.50%	10.93%
GPT-5.2	37.48%	90.00%	58.82%	1.64%	0.00%
Gemini-3.1-Pro-Preview	26.61%	62.00%	26.14%	1.64%	0.00%
Gemini-3-Flash-Preview	22.52%	60.00%	8.50%	1.09%	0.00%
Qwen3.5-397B	19.84%	40.00%	6.54%	0.55%	0.00%
Qwen3.5-Plus	18.74%	20.00%	9.15%	0.55%	0.00%

Overall and size-specific accuracies are reproduced as reported in Tables 5 and 8 of the paper. GPT-5.4 used XHigh reasoning; GPT-5.2 and the Gemini models used High reasoning; the Qwen models used their maximum reasoning budget.

Performance declines sharply as grid size increases. GPT-5.4 and GPT-5.2 reach 90% accuracy on the Baseline subset, but most evaluated models solve none of the Large problems.

Human calibration

In the baseline study, 33 participants achieved 94.54% mean accuracy (98% median) on 50 problems with 5 × 5 grids, using the same eight demonstration examples. This supports the sufficiency of the evidence for human rule inference. Human calibration used the smaller Baseline subset; the model accuracies above cover the full benchmark.

Dataset composition

Category	Label	Records	Rows	Columns per row
Baseline	B	50	5	5
Small	S	153	10	3–10
Medium	M	283	40	10–40
Large	L	149	100	10–100
Total		635

All 100 unsolvable instances belong to the Medium category. The Baseline category is the 50-problem subset used for human calibration. Row lengths may vary; the benchmark includes non-square layouts.

The canonical benchmark is provided as a single evaluation split, test. The size categories are values of the size field within that split, not separate training or validation splits.

Task

Each problem provides:

1. A shared image containing eight demonstration worlds.
2. A shared prompt pairing those worlds with encoded descriptions and action sequences.
3. A new encoded query world.

The model returns a hyphen-separated sequence of action tokens, wrapped in --- as requested by the supplied prompt. A correctly formatted response may look like:

---ma-ga-ga-ya-va-ba-ma---

This illustrates the output format, not a solution to a particular benchmark record. For an unsolvable world, the correct response is ---ta---.

The worlds combine navigation with item use, traps, monsters, water, and reversal tiles. Success can require intermediate goals, tracking a single carried item, and applying exceptions to movement rules.

Data format

CIPHERGRID_Benchmark.jsonl contains one JSON object per line.

Field	Type	Description
id	Integer	Unique record identifier. Preserve it when joining predictions and evaluation results; IDs are not necessarily consecutive.
prompt-base	String	Shared instructions and the eight encoded demonstration descriptions and solutions.
query	String	Compact encoded query world, including row-marker and tile tokens.
answer	String	Reference action sequence, or ta for an unsolvable instance. Used for evaluation, not model input.
image	String	Base64-encoded PNG containing the shared visual demonstrations. This is image data, not a file path or URL.
complex	Boolean	Generation metadata. It is true for every record in this release and does not distinguish difficulty categories.
size	String	Size category: B, S, M, or L.

The image and prompt-base values are repeated across records so that each record contains the evidence needed for evaluation. The image depicts the demonstration worlds, not the query world; the query layout must be reconstructed from its encoded text.

Quick start

After downloading CIPHERGRID_Benchmark.jsonl, load the evaluation split with Hugging Face Datasets:

python -m pip install datasets

import base64
from datasets import load_dataset

dataset = load_dataset(
    "json",
    data_files={"test": "CIPHERGRID_Benchmark.jsonl"},
    split="test",
)

example = dataset[0]
prompt = example["prompt-base"] + "\n" + example["query"]
image_bytes = base64.b64decode(example["image"])

print("Record:", example["id"])
print("Category:", example["size"])
print("Total records:", len(dataset))

Supply prompt and the decoded image to the model using its multimodal input interface. Keep answer out of the model input. For the standard rule-inference evaluation, do not add the decoded vocabulary, explicit game rules, or reference solutions beyond the demonstrations already included in the prompt.

Evaluation

Use solver-based validation rather than exact string matching. A model can produce a valid route that differs from the stored reference solution. The evaluation accepts a reference-matching answer or an action sequence that legally reaches the goal under the benchmark transition rules. The action ta is correct only for an unsolvable instance; returning it for a solvable world is premature resignation.

The accompanying CIPHERGRID code repository includes parse_response_csv.py, validate_solutions.py, and validate_by_size.py for response cleaning, validation, and grouped reporting. Preserve the supplied record IDs throughout evaluation and report:

• Overall accuracy across all 635 records.
• Accuracy by size category and by solvability.
• The model identifier, reasoning settings, and inference limits used.

Dataset creation and scope

CIPHERGRID worlds were generated procedurally and checked using the benchmark's reference solver. The dataset contains synthetic puzzle instances and a shared demonstration image; it does not contain participant-level human-study records. The canonical release uses a single fixed substitution mapping and evidence set.

The benchmark is intended for research on multimodal rule inference, symbolic grounding, and sequential planning. Its controlled environment supports analysis of how inferred rules are carried into action, but performance should not be treated as a general measure of intelligence or real-world competence. Results depend on model versions, inference settings, and the evaluation protocol. Training or tuning on the released evaluation records should be disclosed when reporting results.

License

The dataset is released under the Creative Commons Attribution 4.0 International license (CC BY 4.0).

Citation

If you use CIPHERGRID in your research, please cite:

@inproceedings{curtis2026ciphergrid,
  title     = {{CIPHERGRID} Benchmark: From Multimodal Rule Inference to Sequential Action},
  author    = {Curtis, Christopher and Fragoso, Victor and Savage, Saiph},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026},
  note      = {Accepted, Evaluations and Datasets Track}
}