# Long-context understanding on RULER 32K-HARD

Source: [Long-context understanding on RULER 32K-HARD](<https://www.sota2.com/research/sota/long-context-understanding-on-ruler-32k-hard>)

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

Data updated: 2026-05-26T12:36:45.113Z

Evaluation of model performance on the HARD subset of the RULER dataset configured for a 32,000 token context length, reporting Average 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 | Long-context understanding |
| Dataset (including any stated split) | RULER 32K-HARD |
| Metric | Average Accuracy |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 15 of 15 |
| 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

- [Average Accuracy](<https://www.sota2.com/research/sota/long-context-understanding-on-ruler-32k-hard.md?metric=Average%20Accuracy>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Average Accuracy | Result date | Report paper |
| --- | --- | --- | --- |
| SDPA | 88.74 | Not recorded | 68fa7bf01e209131cea8b723 |
| vAttention(oracle-top-k) | 88.61 | Not recorded | 68fa7bf01e209131cea8b723 |
| oracle-top-k | 87.18 | Not recorded | 68fa7bf01e209131cea8b723 |
| vAttention(HAT) | 86.56 | Not recorded | 68fa7bf01e209131cea8b723 |
| HAT | 81.94 | Not recorded | 68fa7bf01e209131cea8b723 |
| SDPA | 65.41 | Not recorded | 68fa7bf01e209131cea8b723 |
| vAttention(oracle-top-k) | 65.15 | Not recorded | 68fa7bf01e209131cea8b723 |
| vAttention(HAT) | 65.06 | Not recorded | 68fa7bf01e209131cea8b723 |
| oracle-top-k | 64.87 | Not recorded | 68fa7bf01e209131cea8b723 |
| oracle-top-k | 64.37 | Not recorded | 68fa7bf01e209131cea8b723 |
| vAttention(oracle-top-k) | 64.12 | Not recorded | 68fa7bf01e209131cea8b723 |
| SDPA | 64.05 | Not recorded | 68fa7bf01e209131cea8b723 |
| HAT | 60.7 | Not recorded | 68fa7bf01e209131cea8b723 |
| vAttention(HAT) | 56.9 | Not recorded | 68fa7bf01e209131cea8b723 |
| HAT | 54.66 | Not recorded | 68fa7bf01e209131cea8b723 |

### SDPA

Average Accuracy: 88.74. Result ID: 6a1593d3fbc373592c0e5abb.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa703d1e209131cea49db1.md>)

[Implementation](<https://github.com/tensorflow/tensor2tensor>)

Recorded evaluation configuration: {"Backbone":"Llama-3.1-8B-Inst","Sparsity":"0%"}. This record does not assert independent reproduction.

### vAttention(oracle-top-k)

Average Accuracy: 88.61. Result ID: 6a1593d3fbc373592c0e5ac1.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

Recorded evaluation configuration: {"Backbone":"Llama-3.1-8B-Inst","Sparsity":"10%"}. This record does not assert independent reproduction.

### oracle-top-k

Average Accuracy: 87.18. Result ID: 6a1593d3fbc373592c0e5abe.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

Recorded evaluation configuration: {"Backbone":"Llama-3.1-8B-Inst","Sparsity":"10%"}. This record does not assert independent reproduction.

### vAttention(HAT)

Average Accuracy: 86.56. Result ID: 6a1593d3fbc373592c0e5ac7.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

Recorded evaluation configuration: {"Backbone":"Llama-3.1-8B-Inst","Sparsity":"10%"}. This record does not assert independent reproduction.

### HAT

Average Accuracy: 81.94. Result ID: 6a1593d3fbc373592c0e5ac4.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa790e1e209131cea7be39.md>)

Recorded evaluation configuration: {"Backbone":"Llama-3.1-8B-Inst","Sparsity":"10%"}. This record does not assert independent reproduction.

### SDPA

Average Accuracy: 65.41. Result ID: 6a1593d3fbc373592c0e5abc.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa703d1e209131cea49db1.md>)

[Implementation](<https://github.com/tensorflow/tensor2tensor>)

Recorded evaluation configuration: {"Backbone":"Dpsk-R1-Distill-Llama-8B","Sparsity":"0%"}. This record does not assert independent reproduction.

### vAttention(oracle-top-k)

Average Accuracy: 65.15. Result ID: 6a1593d3fbc373592c0e5ac2.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

Recorded evaluation configuration: {"Backbone":"Dpsk-R1-Distill-Llama-8B","Sparsity":"10%"}. This record does not assert independent reproduction.

### vAttention(HAT)

Average Accuracy: 65.06. Result ID: 6a1593d3fbc373592c0e5ac8.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

Recorded evaluation configuration: {"Backbone":"Dpsk-R1-Distill-Llama-8B","Sparsity":"10%"}. This record does not assert independent reproduction.

### oracle-top-k

Average Accuracy: 64.87. Result ID: 6a1593d3fbc373592c0e5abf.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

Recorded evaluation configuration: {"Backbone":"Dpsk-R1-Distill-Llama-8B","Sparsity":"10%"}. This record does not assert independent reproduction.

### oracle-top-k

Average Accuracy: 64.37. Result ID: 6a1593d3fbc373592c0e5ac0.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Inst-v0.3","Sparsity":"10%"}. This record does not assert independent reproduction.

### vAttention(oracle-top-k)

Average Accuracy: 64.12. Result ID: 6a1593d3fbc373592c0e5ac3.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Inst-v0.3","Sparsity":"10%"}. This record does not assert independent reproduction.

### SDPA

Average Accuracy: 64.05. Result ID: 6a1593d3fbc373592c0e5abd.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa703d1e209131cea49db1.md>)

[Implementation](<https://github.com/tensorflow/tensor2tensor>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Inst-v0.3","Sparsity":"0%"}. This record does not assert independent reproduction.

### HAT

Average Accuracy: 60.7. Result ID: 6a1593d3fbc373592c0e5ac5.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa790e1e209131cea7be39.md>)

Recorded evaluation configuration: {"Backbone":"Dpsk-R1-Distill-Llama-8B","Sparsity":"10%"}. This record does not assert independent reproduction.

### vAttention(HAT)

Average Accuracy: 56.9. Result ID: 6a1593d3fbc373592c0e5ac9.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Inst-v0.3","Sparsity":"10%"}. This record does not assert independent reproduction.

### HAT

Average Accuracy: 54.66. Result ID: 6a1593d3fbc373592c0e5ac6.

[Reporting paper](<https://www.sota2.com/research/paper/68fa7bf01e209131cea8b723.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa790e1e209131cea7be39.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Inst-v0.3","Sparsity":"10%"}. This record does not assert independent reproduction.

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