# Long-Context Retrieval on RULER (4K-128K Sweep with Latency) Source: [Long-Context Retrieval on RULER (4K-128K Sweep with Latency)]() Collection index: [llms.txt]() Data updated: 2026-02-26T06:02:10.527Z Evaluation of model performance on the RULER dataset, measuring retrieval accuracy across context window sizes from 4K up to 128K tokens, alongside measured inference latency. 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 retrieval/accuracy across variable lengths | | Dataset (including any stated split) | RULER | | Metric | Retrieval Accuracy (4K) | | Unit | PERCENT | | Direction | Higher is better | | Results loaded | 10 of 10 | | 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 - [Retrieval Accuracy (4K)](): Higher is better; unit: PERCENT. - [Retrieval Accuracy (8K)](): Higher is better; unit: PERCENT. - [Retrieval Accuracy (16K)](): Higher is better; unit: PERCENT. - [Retrieval Accuracy (32K)](): Higher is better; unit: PERCENT. - [Retrieval Accuracy (64K)](): Higher is better; unit: PERCENT. - [Retrieval Accuracy (128K)](): Higher is better; unit: PERCENT. - [Average Retrieval Accuracy](): Higher is better; unit: PERCENT. - [Inference Latency (ms)](): Lower is better; unit: TIME. ## Recorded results | Method | Retrieval Accuracy (4K) | Result date | Report paper | | --- | --- | --- | --- | | LLaMA3.1-8B-Instruct | 96.74 | Not recorded | 693a50246cfd488761a54ba7 | | MInference | 96.54 | Not recorded | 693a50246cfd488761a54ba7 | | TCA-Attention | 96.31 | Not recorded | 693a50246cfd488761a54ba7 | | XAttention | 96.15 | Not recorded | 693a50246cfd488761a54ba7 | | MInference | 96.08 | Not recorded | 693a50246cfd488761a54ba7 | | Qwen2.5-7B-Instruct | 96 | Not recorded | 693a50246cfd488761a54ba7 | | TCA-Attention | 96 | Not recorded | 693a50246cfd488761a54ba7 | | FlexPrefill | 95.99 | Not recorded | 693a50246cfd488761a54ba7 | | FlexPrefill | 95.62 | Not recorded | 693a50246cfd488761a54ba7 | | XAttention | 95.45 | Not recorded | 693a50246cfd488761a54ba7 | ### LLaMA3.1-8B-Instruct Retrieval Accuracy (4K): 96.74. Result ID: 6996c5ae5698665ec4112a2f. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Model":"LLaMA3.1-8B-Instruct","Attention":"Full","Max Context":"128K"}. This record does not assert independent reproduction. ### MInference Retrieval Accuracy (4K): 96.54. Result ID: 6996c5ae5698665ec4112a30. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Model":"LLaMA3.1-8B-Instruct","Max Context":"128K"}. This record does not assert independent reproduction. ### TCA-Attention Retrieval Accuracy (4K): 96.31. Result ID: 6996c5ae5698665ec4112a33. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Model":"LLaMA3.1-8B-Instruct","Max Context":"128K"}. This record does not assert independent reproduction. ### XAttention Retrieval Accuracy (4K): 96.15. Result ID: 6996c5ae5698665ec4112a32. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Model":"LLaMA3.1-8B-Instruct","Max Context":"128K"}. This record does not assert independent reproduction. ### MInference Retrieval Accuracy (4K): 96.08. Result ID: 6996c5ae5698665ec4112a35. [Reporting paper]() [Method origin paper]() [Implementation]() Recorded evaluation configuration: {"Model":"Qwen2.5-7B-Instruct","Max Context":"128K"}. This record does not assert independent reproduction. ### Qwen2.5-7B-Instruct Retrieval Accuracy (4K): 96. Result ID: 6996c5ae5698665ec4112a34. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Model":"Qwen2.5-7B-Instruct","Attention":"Full","Max Context":"128K"}. This record does not assert independent reproduction. ### TCA-Attention Retrieval Accuracy (4K): 96. Result ID: 6996c5ae5698665ec4112a38. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Model":"Qwen2.5-7B-Instruct","Max Context":"128K"}. This record does not assert independent reproduction. ### FlexPrefill Retrieval Accuracy (4K): 95.99. Result ID: 6996c5ae5698665ec4112a31. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Model":"LLaMA3.1-8B-Instruct","Max Context":"128K"}. This record does not assert independent reproduction. ### FlexPrefill Retrieval Accuracy (4K): 95.62. Result ID: 6996c5ae5698665ec4112a36. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Model":"Qwen2.5-7B-Instruct","Max Context":"128K"}. This record does not assert independent reproduction. ### XAttention Retrieval Accuracy (4K): 95.45. Result ID: 6996c5ae5698665ec4112a37. [Reporting paper]() [Method origin paper]() Recorded evaluation configuration: {"Model":"Qwen2.5-7B-Instruct","Max Context":"128K"}. This record does not assert independent reproduction. Page 1 of 1. This is an index page, not the full collection.