# Long-context language understanding on LongBench (Full Metric Suite)

Source: [Long-context language understanding on LongBench (Full Metric Suite)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite>)

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

Data updated: 2026-06-02T07:16:19.688Z

Evaluation of long-context language understanding capabilities across the full range of LongBench subtasks, including specific English and Chinese variants for metrics like Multi-Fact QA and Passage Retrieval, alongside standard QA and summarization tasks.

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) | LongBench |
| Metric | Narrative Score (Nrtv.) |
| 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

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

- [QA Score (Qasp.)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=QA%20Score%20(Qasp.)>): Higher is better; unit: PERCENT.

- [Multi-Fact Score (English) (MF-e)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=Multi-Fact%20Score%20(English)%20(MF-e)>): Higher is better; unit: PERCENT.

- [Multi-Fact Score (MF-z)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=Multi-Fact%20Score%20(MF-z)>): Higher is better; unit: PERCENT.

- [HotpotQA Score (Hotp.)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=HotpotQA%20Score%20(Hotp.)>): Higher is better; unit: PERCENT.

- [2WikiMultiHopQA Score (2Wiki)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=2WikiMultiHopQA%20Score%20(2Wiki)>): Higher is better; unit: PERCENT.

- [MusicQA Score (Musi.)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=MusicQA%20Score%20(Musi.)>): Higher is better; unit: PERCENT.

- [Document Understanding Rate (Dure.)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=Document%20Understanding%20Rate%20(Dure.)>): Higher is better; unit: PERCENT.

- [Government Report QA Score (GovR.)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=Government%20Report%20QA%20Score%20(GovR.)>): Higher is better; unit: PERCENT.

- [Multi-hop QA Score (Micro) (QMSm.)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=Multi-hop%20QA%20Score%20(Micro)%20(QMSm.)>): Higher is better; unit: PERCENT.

- [New Task Score (MNew.)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=New%20Task%20Score%20(MNew.)>): Higher is better; unit: PERCENT.

- [Video Concept Score (Micro) (VCSm.)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=Video%20Concept%20Score%20(Micro)%20(VCSm.)>): Higher is better; unit: PERCENT.

- [TREC QA Score (TREC)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=TREC%20QA%20Score%20(TREC)>): Higher is better; unit: PERCENT.

- [TriviaQA Score (Triv.)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=TriviaQA%20Score%20(Triv.)>): Higher is better; unit: PERCENT.

- [Sample Selection Score (SamS.)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=Sample%20Selection%20Score%20(SamS.)>): Higher is better; unit: PERCENT.

- [Long Summary Score (LSHT)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=Long%20Summary%20Score%20(LSHT)>): Higher is better; unit: PERCENT.

- [Passage Count Metric (PCnt.)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=Passage%20Count%20Metric%20(PCnt.)>): Higher is better; unit: SCALAR.

- [Passage Retrieval Score (English) (PRe-e)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=Passage%20Retrieval%20Score%20(English)%20(PRe-e)>): Higher is better; unit: PERCENT.

- [Passage Retrieval Score (PRe-z)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=Passage%20Retrieval%20Score%20(PRe-z)>): Higher is better; unit: PERCENT.

- [Long Context Comprehension Score (Lcc)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=Long%20Context%20Comprehension%20Score%20(Lcc)>): Higher is better; unit: PERCENT.

- [Reading Comprehension Score (RB-P)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=Reading%20Comprehension%20Score%20(RB-P)>): Higher is better; unit: PERCENT.

- [Average Score (Avg.)](<https://www.sota2.com/research/sota/long-context-language-understanding-on-longbench-full-metric-suite.md?metric=Average%20Score%20(Avg.)>): Higher is better; unit: PERCENT.

## Recorded results

| Method | Narrative Score (Nrtv.) | Result date | Report paper |
| --- | --- | --- | --- |
| PyramidKV | 27.36 | Not recorded | 6a1e635edbd7150506f8502c |
| SnapKV | 25.05 | Not recorded | 6a1e635edbd7150506f8502c |
| SnapKV + ART | 24.57 | Not recorded | 6a1e635edbd7150506f8502c |
| PyramidKV + ART | 24.51 | Not recorded | 6a1e635edbd7150506f8502c |
| Baseline | 24.34 | Not recorded | 6a1e635edbd7150506f8502c |
| ART | 22.04 | Not recorded | 6a1e635edbd7150506f8502c |
| PyramidKV + ART | 19.72 | Not recorded | 6a1e635edbd7150506f8502c |
| StreamingLLM | 18.77 | Not recorded | 6a1e635edbd7150506f8502c |
| StreamingLLM | 18.65 | Not recorded | 6a1e635edbd7150506f8502c |
| PyramidKV | 18.62 | Not recorded | 6a1e635edbd7150506f8502c |
| StreamingLLM + ART | 18.39 | Not recorded | 6a1e635edbd7150506f8502c |
| SnapKV | 17.89 | Not recorded | 6a1e635edbd7150506f8502c |
| StreamingLLM + ART | 17.72 | Not recorded | 6a1e635edbd7150506f8502c |
| SnapKV + ART | 16.19 | Not recorded | 6a1e635edbd7150506f8502c |

### PyramidKV

Narrative Score (Nrtv.): 27.36. Result ID: 6a1e8340fbc373592c182c92.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa78841e209131cea78cba.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Instruct-v0.3","KV cache retention ratio":"0.8","ART integration":"false"}. This record does not assert independent reproduction.

### SnapKV

Narrative Score (Nrtv.): 25.05. Result ID: 6a1e8340fbc373592c182c8e.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa73b21e209131cea5df89.md>)

[Implementation](<https://github.com/FasterDecoding/SnapKV>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Instruct-v0.3","KV cache retention ratio":"0.8","ART integration":"false"}. This record does not assert independent reproduction.

### SnapKV + ART

Narrative Score (Nrtv.): 24.57. Result ID: 6a1e8340fbc373592c182c8f.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Instruct-v0.3","KV cache retention ratio":"0.8","ART integration":"true"}. This record does not assert independent reproduction.

### PyramidKV + ART

Narrative Score (Nrtv.): 24.51. Result ID: 6a1e8340fbc373592c182c93.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Instruct-v0.3","KV cache retention ratio":"0.8","ART integration":"true"}. This record does not assert independent reproduction.

### Baseline

Narrative Score (Nrtv.): 24.34. Result ID: 6a1e8340fbc373592c182c88.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Instruct-v0.3","KV cache retention ratio":"1.0","ART integration":"false"}. This record does not assert independent reproduction.

### ART

Narrative Score (Nrtv.): 22.04. Result ID: 6a1e8340fbc373592c182c89.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Instruct-v0.3","KV cache retention ratio":"1.0","ART integration":"true"}. This record does not assert independent reproduction.

### PyramidKV + ART

Narrative Score (Nrtv.): 19.72. Result ID: 6a1e8340fbc373592c182c95.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Instruct-v0.3","KV cache retention ratio":"0.2","ART integration":"true"}. This record does not assert independent reproduction.

### StreamingLLM

Narrative Score (Nrtv.): 18.77. Result ID: 6a1e8340fbc373592c182c8a.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa72d71e209131cea58d79.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Instruct-v0.3","KV cache retention ratio":"0.8","ART integration":"false"}. This record does not assert independent reproduction.

### StreamingLLM

Narrative Score (Nrtv.): 18.65. Result ID: 6a1e8340fbc373592c182c8c.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa72d71e209131cea58d79.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Instruct-v0.3","KV cache retention ratio":"0.2","ART integration":"false"}. This record does not assert independent reproduction.

### PyramidKV

Narrative Score (Nrtv.): 18.62. Result ID: 6a1e8340fbc373592c182c94.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa78841e209131cea78cba.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Instruct-v0.3","KV cache retention ratio":"0.2","ART integration":"false"}. This record does not assert independent reproduction.

### StreamingLLM + ART

Narrative Score (Nrtv.): 18.39. Result ID: 6a1e8340fbc373592c182c8d.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Instruct-v0.3","KV cache retention ratio":"0.2","ART integration":"true"}. This record does not assert independent reproduction.

### SnapKV

Narrative Score (Nrtv.): 17.89. Result ID: 6a1e8340fbc373592c182c90.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa73b21e209131cea5df89.md>)

[Implementation](<https://github.com/FasterDecoding/SnapKV>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Instruct-v0.3","KV cache retention ratio":"0.2","ART integration":"false"}. This record does not assert independent reproduction.

### StreamingLLM + ART

Narrative Score (Nrtv.): 17.72. Result ID: 6a1e8340fbc373592c182c8b.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Instruct-v0.3","KV cache retention ratio":"0.8","ART integration":"true"}. This record does not assert independent reproduction.

### SnapKV + ART

Narrative Score (Nrtv.): 16.19. Result ID: 6a1e8340fbc373592c182c91.

[Reporting paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

[Method origin paper](<https://www.sota2.com/research/paper/6a1e635edbd7150506f8502c.md>)

Recorded evaluation configuration: {"Backbone":"Mistral-7B-Instruct-v0.3","KV cache retention ratio":"0.2","ART integration":"true"}. This record does not assert independent reproduction.

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