# Long-context language understanding on SCROLLS (test)

Source: [Long-context language understanding on SCROLLS (test)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test>)

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

Data updated: 2026-02-26T05:35:48.422Z

Evaluation of long-context language understanding capabilities on the SCROLLS test set, reporting performance across various sub-tasks using ROUGE, F1, and Exact Match metrics.

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 language understanding |
| Dataset (including any stated split) | SCROLLS (test) |
| Metric | Average Score |
| Unit | PERCENT |
| Direction | Higher is better |
| Results loaded | 18 of 18 |
| 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 Score](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=Average%20Score>): Higher is better; unit: PERCENT.

- [GovRep ROUGE-1](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=GovRep%20ROUGE-1>): Higher is better; unit: PERCENT.

- [GovRep ROUGE-2](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=GovRep%20ROUGE-2>): Higher is better; unit: PERCENT.

- [GovRep ROUGE-L](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=GovRep%20ROUGE-L>): Higher is better; unit: PERCENT.

- [SumScr ROUGE-1](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=SumScr%20ROUGE-1>): Higher is better; unit: PERCENT.

- [SumScr ROUGE-2](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=SumScr%20ROUGE-2>): Higher is better; unit: PERCENT.

- [SumScr ROUGE-L](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=SumScr%20ROUGE-L>): Higher is better; unit: PERCENT.

- [QMSum ROUGE-1](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=QMSum%20ROUGE-1>): Higher is better; unit: PERCENT.

- [QMSum ROUGE-2](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=QMSum%20ROUGE-2>): Higher is better; unit: PERCENT.

- [QMSum ROUGE-L](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=QMSum%20ROUGE-L>): Higher is better; unit: PERCENT.

- [Qspr F1](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=Qspr%20F1>): Higher is better; unit: PERCENT.

- [Nrtv F1](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=Nrtv%20F1>): Higher is better; unit: PERCENT.

- [QALT EM (T)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=QALT%20EM%20(T)>): Higher is better; unit: PERCENT.

- [QALT EM (H)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=QALT%20EM%20(H)>): Higher is better; unit: PERCENT.

- [CNLI EM](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=CNLI%20EM>): Higher is better; unit: PERCENT.

- [TQA F1](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=TQA%20F1>): Higher is better; unit: PERCENT.

- [QAS F1](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=QAS%20F1>): Higher is better; unit: PERCENT.

- [QuAL EM](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=QuAL%20EM>): Higher is better; unit: PERCENT.

- [arXiv ROUGE](<https://www.sota2.com/research/sota/long-context-language-understanding-on-scrolls-test.md?metric=arXiv%20ROUGE>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Average Score | Result date | Report paper |
| --- | --- | --- | --- |
| COLT5-XL | 47.4 | Not recorded | 68fa70f71e209131cea4e70a |
| LONGT5-XL | 46.6 | Not recorded | 68fa70f71e209131cea4e70a |
| LONGT5-L | 45.3 | Not recorded | 68fa70f71e209131cea4e70a |
| COLT5-L | 45.3 | Not recorded | 68fa70f71e209131cea4e70a |
| LONGT5-B | 43.1 | Not recorded | 68fa70f71e209131cea4e70a |
| COLT5-B | 42.4 | Not recorded | 68fa70f71e209131cea4e70a |
| LongT5XL | 41.9 | Not recorded | 68fa6e481e209131cea3ef77 |
| LongT5large | 40.5 | Not recorded | 68fa6e481e209131cea3ef77 |
| LongT5base | 38.2 | Not recorded | 68fa6e481e209131cea3ef77 |
| BARTlarge-SLED | 38 | Not recorded | 68fa6e481e209131cea3ef77 |
| UL2 | 37.9 | Not recorded | 68fa6e481e209131cea3ef77 |
| BARTbase-SLED | 35.4 | Not recorded | 68fa6e481e209131cea3ef77 |
| LEDbase | 33.6 | Not recorded | 68fa6e481e209131cea3ef77 |
| T5base-SLED | 33.3 | Not recorded | 68fa6e481e209131cea3ef77 |
| BARTlarge | 32.1 | Not recorded | 68fa6e481e209131cea3ef77 |
| BARTbase | 30.6 | Not recorded | 68fa6e481e209131cea3ef77 |
| LED\_SCROLLS | 29.2 | Not recorded | 68fa6e481e209131cea3ef77 |
| T5base | 26.3 | Not recorded | 68fa6e481e209131cea3ef77 |

### COLT5-XL

Average Score: 47.4. Result ID: 699183ec3a6c31484908abcc.

[Reporting paper](<https://www.sota2.com/research/paper/68fa70f71e209131cea4e70a.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa70f71e209131cea4e70a.md>)

[Implementation](<https://github.com/google/flaxformer>)

Recorded evaluation configuration: {"Model Scale":"XL","Inference Speed (inf)":"2.3 samples/s","Fine-tuning Speed (fn)":"0.5 samples/s"}. This record does not assert independent reproduction.

### LONGT5-XL

Average Score: 46.6. Result ID: 699183ec3a6c31484908abcb.

[Reporting paper](<https://www.sota2.com/research/paper/68fa70f71e209131cea4e70a.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6c7c1e209131cea34428.md>)

[Implementation](<https://github.com/google-research/longt5>)

Recorded evaluation configuration: {"Model Scale":"XL","Inference Speed (no MQA)":"0.2 samples/s","Inference Speed (with MQA)":"1.2 samples/s","Fine-tuning Speed (fn)":"0.4 samples/s"}. This record does not assert independent reproduction.

### LONGT5-L

Average Score: 45.3. Result ID: 699183ec3a6c31484908abc9.

[Reporting paper](<https://www.sota2.com/research/paper/68fa70f71e209131cea4e70a.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6c7c1e209131cea34428.md>)

[Implementation](<https://github.com/google-research/longt5>)

Recorded evaluation configuration: {"Model Scale":"Large","Inference Speed (no MQA)":"0.3 samples/s","Inference Speed (with MQA)":"3.0 samples/s","Fine-tuning Speed (fn)":"1.3 samples/s"}. This record does not assert independent reproduction.

### COLT5-L

Average Score: 45.3. Result ID: 699183ec3a6c31484908abca.

[Reporting paper](<https://www.sota2.com/research/paper/68fa70f71e209131cea4e70a.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa70f71e209131cea4e70a.md>)

[Implementation](<https://github.com/google/flaxformer>)

Recorded evaluation configuration: {"Model Scale":"Large","Inference Speed (inf)":"5.0 samples/s","Fine-tuning Speed (fn)":"2.0 samples/s"}. This record does not assert independent reproduction.

### LONGT5-B

Average Score: 43.1. Result ID: 699183ec3a6c31484908abc7.

[Reporting paper](<https://www.sota2.com/research/paper/68fa70f71e209131cea4e70a.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6c7c1e209131cea34428.md>)

[Implementation](<https://github.com/google-research/longt5>)

Recorded evaluation configuration: {"Model Scale":"Base","Inference Speed (no MQA)":"0.6 samples/s","Inference Speed (with MQA)":"7.4 samples/s","Fine-tuning Speed (fn)":"3.7 samples/s"}. This record does not assert independent reproduction.

### COLT5-B

Average Score: 42.4. Result ID: 699183ec3a6c31484908abc8.

[Reporting paper](<https://www.sota2.com/research/paper/68fa70f71e209131cea4e70a.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa70f71e209131cea4e70a.md>)

[Implementation](<https://github.com/google/flaxformer>)

Recorded evaluation configuration: {"Model Scale":"Base","Inference Speed (inf)":"11.2 samples/s","Fine-tuning Speed (fn)":"6.5 samples/s"}. This record does not assert independent reproduction.

### LongT5XL

Average Score: 41.9. Result ID: 69900c743a6c314849fdbf0c.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6c7c1e209131cea34428.md>)

[Implementation](<https://github.com/google-research/longt5>)

Recorded evaluation configuration: {"Chunk/Input":"255/16K","#Params":"3B"}. This record does not assert independent reproduction.

### LongT5large

Average Score: 40.5. Result ID: 69900c743a6c314849fdbf0b.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6c7c1e209131cea34428.md>)

[Implementation](<https://github.com/google-research/longt5>)

Recorded evaluation configuration: {"Chunk/Input":"255/16K","#Params":"770M"}. This record does not assert independent reproduction.

### LongT5base

Average Score: 38.2. Result ID: 69900c743a6c314849fdbf0a.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6c7c1e209131cea34428.md>)

[Implementation](<https://github.com/google-research/longt5>)

Recorded evaluation configuration: {"Chunk/Input":"255/16K","#Params":"220M"}. This record does not assert independent reproduction.

### BARTlarge-SLED

Average Score: 38. Result ID: 69900c743a6c314849fdbf08.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Implementation](<https://github.com/Mivg/SLED>)

Recorded evaluation configuration: {"Chunk/Input":"256/16K","#Params":"406M"}. This record does not assert independent reproduction.

### UL2

Average Score: 37.9. Result ID: 69900c743a6c314849fdbf0d.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6eaf1e209131cea416bb.md>)

[Implementation](<https://github.com/google-research/google-research/tree/master/ul2>)

Recorded evaluation configuration: {"Chunk/Input":"2K/2K","#Params":"20B"}. This record does not assert independent reproduction.

### BARTbase-SLED

Average Score: 35.4. Result ID: 69900c743a6c314849fdbf06.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Implementation](<https://github.com/Mivg/SLED>)

Recorded evaluation configuration: {"Chunk/Input":"256/16K","#Params":"139M"}. This record does not assert independent reproduction.

### LEDbase

Average Score: 33.6. Result ID: 69900c743a6c314849fdbf02.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa68d41e209131cea1e701.md>)

[Implementation](<https://github.com/allenai/longformer>)

Recorded evaluation configuration: {"Chunk/Input":"256/16K","#Params":"162M"}. This record does not assert independent reproduction.

### T5base-SLED

Average Score: 33.3. Result ID: 69900c743a6c314849fdbf04.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Implementation](<https://github.com/Mivg/SLED>)

Recorded evaluation configuration: {"Chunk/Input":"256/16K","#Params":"220M"}. This record does not assert independent reproduction.

### BARTlarge

Average Score: 32.1. Result ID: 69900c743a6c314849fdbf07.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa66921e209131cea10fff.md>)

Recorded evaluation configuration: {"Chunk/Input":"1K/1K","#Params":"406M"}. This record does not assert independent reproduction.

### BARTbase

Average Score: 30.6. Result ID: 69900c743a6c314849fdbf05.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa66921e209131cea10fff.md>)

Recorded evaluation configuration: {"Chunk/Input":"1K/1K","#Params":"139M"}. This record does not assert independent reproduction.

### LED\_SCROLLS

Average Score: 29.2. Result ID: 69900c743a6c314849fdbf09.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa6d9d1e209131cea3ad2e.md>)

[Implementation](<https://github.com/tau-nlp/scrolls>)

Recorded evaluation configuration: {"Chunk/Input":"1K/16K","#Params":"162M"}. This record does not assert independent reproduction.

### T5base

Average Score: 26.3. Result ID: 69900c743a6c314849fdbf03.

[Reporting paper](<https://www.sota2.com/research/paper/68fa6e481e209131cea3ef77.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa709e1e209131cea4c470.md>)

[Implementation](<https://github.com/google-research/text-to-text-transfer-transformer>)

Recorded evaluation configuration: {"Chunk/Input":"1K/1K","#Params":"220M"}. This record does not assert independent reproduction.

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