# Modular arithmetic prediction on MNIST modular arithmetic Unseen operations (test)

Source: [Modular arithmetic prediction on MNIST modular arithmetic Unseen operations (test)](<https://www.sota2.com/research/sota/modular-arithmetic-prediction-on-mnist-modular-arithmetic-unseen-operations-test>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-06-02T10:56:53.650Z

Evaluation of model performance on the test split of the MNIST modular arithmetic dataset, specifically measuring generalization capabilities to unseen arithmetic operations using metrics like Zero-Shot Op Accuracy and Rollout Accuracy.

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

| Scope | Recorded value |
| --- | --- |
| Task | Modular arithmetic prediction |
| Dataset (including any stated split) | MNIST modular arithmetic Unseen operations (test) |
| Metric | Zero-Shot Op Accuracy |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 14 of 14 |
| Coverage | All results currently returned for this benchmark |

Evaluation settings may differ between results. The original papers and recorded configurations are needed to establish comparability. A shared metric name alone does not establish an identical protocol.

## Metrics

- [Zero-Shot Op Accuracy](<https://www.sota2.com/research/sota/modular-arithmetic-prediction-on-mnist-modular-arithmetic-unseen-operations-test.md?metric=Zero-Shot%20Op%20Accuracy>): Higher is better; unit: PERCENT.

- [Unseen Op Rollout Accuracy](<https://www.sota2.com/research/sota/modular-arithmetic-prediction-on-mnist-modular-arithmetic-unseen-operations-test.md?metric=Unseen%20Op%20Rollout%20Accuracy>): Higher is better; unit: PERCENT.

- [Unseen Op KNN Accuracy](<https://www.sota2.com/research/sota/modular-arithmetic-prediction-on-mnist-modular-arithmetic-unseen-operations-test.md?metric=Unseen%20Op%20KNN%20Accuracy>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Zero-Shot Op Accuracy | Result date | Report paper |
| --- | --- | --- | --- |
| ResNet w/ multi-frequency rotation (MFR) | 99.46 | Not recorded | 6a1e6410dbd7150506f856a7 |
| ResNet18 w/ compositional consistency loss | 99.31 | Not recorded | 6a1e6410dbd7150506f856a7 |
| ResNet w/ single-frequency rotation (SFR) | 98.93 | Not recorded | 6a1e6410dbd7150506f856a7 |
| MLP w/ compositional consistency loss | 97.77 | Not recorded | 6a1e6410dbd7150506f856a7 |
| MLP w/ multi-frequency rotation (MFR) | 97.24 | Not recorded | 6a1e6410dbd7150506f856a7 |
| MLP w/ single-frequency rotation (SFR) | 94.26 | Not recorded | 6a1e6410dbd7150506f856a7 |
| MLP w/ multi-frequency rotation (MFR) | 56.5 | Not recorded | 6a1e6410dbd7150506f856a7 |
| ResNet18 w/ multi-frequency rotation (MFR) | 54.92 | Not recorded | 6a1e6410dbd7150506f856a7 |
| MLP w/ single-frequency rotation (SFR) | 13.92 | Not recorded | 6a1e6410dbd7150506f856a7 |
| ResNet18 w/ single-frequency rotation (SFR) | 10.1 | Not recorded | 6a1e6410dbd7150506f856a7 |
| MLP | 9.95 | Not recorded | 6a1e6410dbd7150506f856a7 |
| ResNet18 | 9.74 | Not recorded | 6a1e6410dbd7150506f856a7 |
| MLP | 5.93 | Not recorded | 6a1e6410dbd7150506f856a7 |
| ResNet18 | 2.42 | Not recorded | 6a1e6410dbd7150506f856a7 |

### ResNet w/ multi-frequency rotation (MFR)

Zero-Shot Op Accuracy: 99.46. Result ID: 6a1eb69ffbc373592c19b73f.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

Recorded evaluation configuration: {"Backbone":"ResNet","Training Objective":"BRo-JEPA","Rotation type (SFR/MFR)":"MFR","Embedding type":"block rotation"}. This record does not assert independent reproduction.

### ResNet18 w/ compositional consistency loss

Zero-Shot Op Accuracy: 99.31. Result ID: 6a1eb69ffbc373592c19b737.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f061e209131cea43581.md>)

Recorded evaluation configuration: {"Backbone":"ResNet18","Training Objective":"JEPA","Compositional consistency loss":"true","Embedding type":"additive"}. This record does not assert independent reproduction.

### ResNet w/ single-frequency rotation (SFR)

Zero-Shot Op Accuracy: 98.93. Result ID: 6a1eb69ffbc373592c19b73e.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

Recorded evaluation configuration: {"Backbone":"ResNet","Training Objective":"BRo-JEPA","Rotation type (SFR/MFR)":"SFR","Embedding type":"block rotation"}. This record does not assert independent reproduction.

### MLP w/ compositional consistency loss

Zero-Shot Op Accuracy: 97.77. Result ID: 6a1eb69ffbc373592c19b735.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6f061e209131cea43581.md>)

Recorded evaluation configuration: {"Backbone":"MLP","Training Objective":"JEPA","Compositional consistency loss":"true","Embedding type":"additive"}. This record does not assert independent reproduction.

### MLP w/ multi-frequency rotation (MFR)

Zero-Shot Op Accuracy: 97.24. Result ID: 6a1eb69ffbc373592c19b73d.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

Recorded evaluation configuration: {"Backbone":"MLP","Training Objective":"BRo-JEPA","Rotation type (SFR/MFR)":"MFR","Embedding type":"block rotation"}. This record does not assert independent reproduction.

### MLP w/ single-frequency rotation (SFR)

Zero-Shot Op Accuracy: 94.26. Result ID: 6a1eb69ffbc373592c19b73c.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

Recorded evaluation configuration: {"Backbone":"MLP","Training Objective":"BRo-JEPA","Rotation type (SFR/MFR)":"SFR","Embedding type":"block rotation"}. This record does not assert independent reproduction.

### MLP w/ multi-frequency rotation (MFR)

Zero-Shot Op Accuracy: 56.5. Result ID: 6a1eb69ffbc373592c19b739.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

Recorded evaluation configuration: {"Backbone":"MLP","Training Objective":"Supervised","Rotation type (SFR/MFR)":"MFR","Embedding type":"block rotation"}. This record does not assert independent reproduction.

### ResNet18 w/ multi-frequency rotation (MFR)

Zero-Shot Op Accuracy: 54.92. Result ID: 6a1eb69ffbc373592c19b73b.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa71271e209131cea4f8df.md>)

Recorded evaluation configuration: {"Backbone":"ResNet18","Training Objective":"Supervised","Rotation type (SFR/MFR)":"MFR","Embedding type":"block rotation"}. This record does not assert independent reproduction.

### MLP w/ single-frequency rotation (SFR)

Zero-Shot Op Accuracy: 13.92. Result ID: 6a1eb69ffbc373592c19b738.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

Recorded evaluation configuration: {"Backbone":"MLP","Training Objective":"Supervised","Rotation type (SFR/MFR)":"SFR","Embedding type":"block rotation"}. This record does not assert independent reproduction.

### ResNet18 w/ single-frequency rotation (SFR)

Zero-Shot Op Accuracy: 10.1. Result ID: 6a1eb69ffbc373592c19b73a.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa71271e209131cea4f8df.md>)

Recorded evaluation configuration: {"Backbone":"ResNet18","Training Objective":"Supervised","Rotation type (SFR/MFR)":"SFR","Embedding type":"block rotation"}. This record does not assert independent reproduction.

### MLP

Zero-Shot Op Accuracy: 9.95. Result ID: 6a1eb69ffbc373592c19b734.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

Recorded evaluation configuration: {"Backbone":"MLP","Training Objective":"JEPA","Embedding type":"additive"}. This record does not assert independent reproduction.

### ResNet18

Zero-Shot Op Accuracy: 9.74. Result ID: 6a1eb69ffbc373592c19b736.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa62321e209131ce9f8328.md>)

Recorded evaluation configuration: {"Backbone":"ResNet18","Training Objective":"JEPA","Embedding type":"additive"}. This record does not assert independent reproduction.

### MLP

Zero-Shot Op Accuracy: 5.93. Result ID: 6a1eb69ffbc373592c19b732.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

Recorded evaluation configuration: {"Backbone":"MLP","Training Objective":"Supervised","Embedding type":"additive"}. This record does not assert independent reproduction.

### ResNet18

Zero-Shot Op Accuracy: 2.42. Result ID: 6a1eb69ffbc373592c19b733.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e6410dbd7150506f856a7.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa62321e209131ce9f8328.md>)

Recorded evaluation configuration: {"Backbone":"ResNet18","Training Objective":"Supervised","Embedding type":"additive"}. This record does not assert independent reproduction.

Page 1 of 1. This is an index page, not the full collection.
