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 the six non-refusing models reported in the main evaluation, overall accuracy ranged from 18.74% to 41.89%.

Model	Overall	Baseline	Small	Medium (solvable)	Large
GPT-5.4	41.89%	90.00%	68.63%	10.93%	25.50%
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 accuracy is reproduced from Table 5 of the paper, while size-specific accuracy is reproduced from Table 8. The Medium column reports only the solvable Medium instances, matching the paper's size analysis. The benchmark additionally contains 100 unsolvable Medium instances.

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 is substantially lower on the Medium and Large categories than on Baseline and Small. GPT-5.4 is the only evaluated model to solve any Large instances.

Claude Opus-4.7 and Claude Sonnet-4.6 were also attempted but consistently triggered false-positive safety refusals on the benchmark's encoded strings and are therefore not included in the accuracy table.

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.

Additional human calibration evaluated the same fixed evidence on larger grids. Groups of ten participants achieved 100% mean accuracy on 10 × 10 grids and 94% mean accuracy on 15 × 15 grids.

The main model accuracies above cover the full 635-record benchmark, while the primary human calibration subset contains the 50 Baseline problems.

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 other 535 records are solvable.

The Baseline category is the 50-problem subset used for the primary human calibration study. Row lengths may vary, so 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 must infer the latent symbolic system from the demonstrations, reconstruct the query world, determine whether a valid route exists, and return the corresponding encoded action sequence.

A response may look like:

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

This illustrates the action-sequence 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 containing the 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 required 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, install Hugging Face Datasets:

python -m pip install datasets

Load the benchmark:

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 CIPHERGRID rule-inference evaluation, do not add the decoded vocabulary, explicit game rules, or reference solutions beyond the demonstrations already provided in prompt-base.

Evaluation

Use solver-based validation rather than exact string matching.

Many CIPHERGRID worlds admit more than one valid solution. A predicted route therefore does not need to exactly match the stored reference sequence.

A response is counted as correct when either:

• it matches the reference solution, or
• replaying the predicted action sequence under the benchmark transition rules legally reaches the goal.

For solver-certified unsolvable instances, ta is correct. Returning ta for a solvable world is counted as premature resignation.

The accompanying CIPHERGRID code repository includes:

• parse_response_csv.py for response cleaning,
• validate_solutions.py for solver-based validation, and
• validate_by_size.py for grouped evaluation.

Preserve the supplied record IDs throughout evaluation.

When reporting CIPHERGRID results, we recommend including:

• overall accuracy across all 635 records,
• accuracy by size category,
• accuracy by solvability,
• the exact model identifier,
• reasoning or thinking settings, and
• relevant inference limits or budgets.

Because the Medium category contains both solvable and unsolvable records, clearly specify whether a reported Medium accuracy includes all Medium records or only its solvable subset. The paper's Table 8 reports the solvable Medium subset in its M column.

Dataset creation and scope

CIPHERGRID worlds were generated procedurally and checked using the benchmark's reference solver.

Environment sizes range from 5 × 5 to 100 × 100. Generation varies row count and row-length bounds, allowing both square and non-square environments.

Every world contains one start and one goal. Generated worlds are parsed into the benchmark's grid representation and passed to the reference solver, which either produces a valid solution or certifies the instance as unsolvable.

The benchmark 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 across all 635 records. The mapping is not resampled for each problem.

The accompanying paper additionally evaluates alternative encodings on controlled subsets, but those alternative mappings are not separate mappings within the canonical 635-record release.

Intended use

CIPHERGRID is intended for research on:

• multimodal rule inference,
• symbolic grounding,
• abstraction and generalization,
• sequential planning,
• state tracking,
• rule application, and
• interaction between reasoning components.

Its controlled environment supports analysis of how inferred rules are carried into downstream action.

Performance on CIPHERGRID should not be interpreted as a general measure of intelligence or real-world competence. Results depend on model versions, inference settings, reasoning budgets, and the evaluation protocol.

Training, fine-tuning, or otherwise adapting a model using the released evaluation records should be disclosed when reporting benchmark results.

Reproducibility

For reproducible evaluation:

1. Use the supplied eight-example evidence set without revealing the decoded vocabulary or game rules.
2. Evaluate on the canonical test split.
3. Preserve the original record IDs.
4. Keep the answer field hidden from the model.
5. Use solver-based validation rather than exact string matching.
6. Report the exact model version and reasoning configuration.
7. Distinguish solvable from unsolvable performance when relevant.

The paper reports results using provider-hosted model APIs. Because proprietary model behavior can change across versions and deployments, exact reproduction may vary.

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}
}