# Long-context understanding and generation on LongBench (test)

Source: [Long-context understanding and generation on LongBench (test)](<https://www.sota2.com/research/sota/long-context-understanding-and-generation-on-longbench-test>)

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

Data updated: 2026-03-24T14:17:06.940Z

Evaluation of language models on the LongBench test set, focusing on performance across key tasks (Single-Doc QA, Multi-Doc QA, Summarization, Few-shot, Synthetic, Code) and reporting overall scores segmented by English (EN) and Chinese (ZH) performance.

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 and generation |
| Dataset (including any stated split) | LongBench (test) |
| Metric | Single-Doc QA Score |
| Unit | SCALAR |
| Direction | Higher is better |
| Results loaded | 7 of 7 |
| 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

- [Single-Doc QA Score](<https://www.sota2.com/research/sota/long-context-understanding-and-generation-on-longbench-test.md?metric=Single-Doc%20QA%20Score>): Higher is better; unit: SCALAR.

- [Multi-Doc QA Score](<https://www.sota2.com/research/sota/long-context-understanding-and-generation-on-longbench-test.md?metric=Multi-Doc%20QA%20Score>): Higher is better; unit: SCALAR.

- [Summarization Score](<https://www.sota2.com/research/sota/long-context-understanding-and-generation-on-longbench-test.md?metric=Summarization%20Score>): Higher is better; unit: SCALAR.

- [Few-shot Score](<https://www.sota2.com/research/sota/long-context-understanding-and-generation-on-longbench-test.md?metric=Few-shot%20Score>): Higher is better; unit: SCALAR.

- [Synthetic Task Score](<https://www.sota2.com/research/sota/long-context-understanding-and-generation-on-longbench-test.md?metric=Synthetic%20Task%20Score>): Higher is better; unit: SCALAR.

- [Code Task Score](<https://www.sota2.com/research/sota/long-context-understanding-and-generation-on-longbench-test.md?metric=Code%20Task%20Score>): Higher is better; unit: SCALAR.

- [Overall Score (EN)](<https://www.sota2.com/research/sota/long-context-understanding-and-generation-on-longbench-test.md?metric=Overall%20Score%20(EN)>): Higher is better; unit: SCALAR.

- [Overall Score (ZH)](<https://www.sota2.com/research/sota/long-context-understanding-and-generation-on-longbench-test.md?metric=Overall%20Score%20(ZH)>): Higher is better; unit: SCALAR.

- [Overall Score (All Languages)](<https://www.sota2.com/research/sota/long-context-understanding-and-generation-on-longbench-test.md?metric=Overall%20Score%20(All%20Languages)>): Higher is better; unit: SCALAR.

## Recorded results

| Method | Single-Doc QA Score | Result date | Report paper |
| --- | --- | --- | --- |
| GPT-3.5-Turbo-16k | 45.1 | Not recorded | 69c21a6e0636e5bf4afc46ef |
| Vicuna-v1.5-7B-16k | 31.8 | Not recorded | 69c21a6e0636e5bf4afc46ef |
| HiCI-7B-16k | 31.1 | Not recorded | 69c21a6e0636e5bf4afc46ef |
| HiCI-7B-16k† | 29.9 | Not recorded | 69c21a6e0636e5bf4afc46ef |
| LongChat-7B-32k | 28.8 | Not recorded | 69c21a6e0636e5bf4afc46ef |
| LongLoRA-7B-16k | 23.7 | Not recorded | 69c21a6e0636e5bf4afc46ef |
| Llama2-7B-chat-4k | 21.7 | Not recorded | 69c21a6e0636e5bf4afc46ef |

### GPT-3.5-Turbo-16k

Single-Doc QA Score: 45.1. Result ID: 69c29cd747f34ce924ab32c8.

[Reporting paper](<https://www.sota2.com/research/paper/69c21a6e0636e5bf4afc46ef.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa727a1e209131cea56a1e.md>)

[Implementation](<https://github.com/openai/evals>)

Recorded evaluation configuration: {"Context window size":"16k"}. This record does not assert independent reproduction.

### Vicuna-v1.5-7B-16k

Single-Doc QA Score: 31.8. Result ID: 69c29cd747f34ce924ab32cb.

[Reporting paper](<https://www.sota2.com/research/paper/69c21a6e0636e5bf4afc46ef.md>)

Recorded evaluation configuration: {"Context window size":"16k"}. This record does not assert independent reproduction.

### HiCI-7B-16k

Single-Doc QA Score: 31.1. Result ID: 69c29cd747f34ce924ab32cd.

[Reporting paper](<https://www.sota2.com/research/paper/69c21a6e0636e5bf4afc46ef.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69c21a6e0636e5bf4afc46ef.md>)

Recorded evaluation configuration: {"Context window size":"16k"}. This record does not assert independent reproduction.

### HiCI-7B-16k†

Single-Doc QA Score: 29.9. Result ID: 69c29cd747f34ce924ab32ce.

[Reporting paper](<https://www.sota2.com/research/paper/69c21a6e0636e5bf4afc46ef.md>)

[Method origin paper](<https://www.sota2.com/research/paper/69c21a6e0636e5bf4afc46ef.md>)

Recorded evaluation configuration: {"Context window size":"16k","Inference attention mechanism":"training-consistent HiCI attention during inference prefill"}. This record does not assert independent reproduction.

### LongChat-7B-32k

Single-Doc QA Score: 28.8. Result ID: 69c29cd747f34ce924ab32ca.

[Reporting paper](<https://www.sota2.com/research/paper/69c21a6e0636e5bf4afc46ef.md>)

Recorded evaluation configuration: {"Context window size":"32k"}. This record does not assert independent reproduction.

### LongLoRA-7B-16k

Single-Doc QA Score: 23.7. Result ID: 69c29cd747f34ce924ab32cc.

[Reporting paper](<https://www.sota2.com/research/paper/69c21a6e0636e5bf4afc46ef.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa727a1e209131cea56a34.md>)

Recorded evaluation configuration: {"Context window size":"16k"}. This record does not assert independent reproduction.

### Llama2-7B-chat-4k

Single-Doc QA Score: 21.7. Result ID: 69c29cd747f34ce924ab32c9.

[Reporting paper](<https://www.sota2.com/research/paper/69c21a6e0636e5bf4afc46ef.md>)

[Method origin paper](<https://www.sota2.com/research/paper/68fa701f1e209131cea492fe.md>)

[Implementation](<https://github.com/facebookresearch/llama>)

Recorded evaluation configuration: {"Context window size":"4k"}. This record does not assert independent reproduction.

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