General Knowledge Evaluation on ARC, HellaSwag, LAMBADA, PIQA, SciQ, WinoGrande, TriviaQA, WebQS, MMLU, GSM8K
60.2ARC-CSPLA
Evaluation Results
| Method | Links | ||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| SPLAParameters=14B, Context Length=32k, Pretraining Tokens=640B2026.01 | 60.2 | 85.5 | 61.9 | 75.2 | 80.6 | 95.8 | 75 | 49.6 | 25.6 | 67.7 | 78.6 | 90.5 | |
| DENSEParameters=14B, Context Length=32k, Pretraining Tokens=640B2026.01 | 60.1 | 85.3 | 61.8 | 74.9 | 80.9 | 95.8 | 74 | 49.6 | 24.9 | 67.5 | 77.3 | 90.2 | |
| INF-V2Parameters=14B, Context Length=32k, Pretraining Tokens=640B, Full name=InfLLM-v22026.01 | 59.2 | 85.5 | 61.6 | 75.3 | 80.5 | 96.2 | 74.4 | 49.1 | 23.9 | 67.3 | 77.8 | 90 | |
| SPAParameters=14B, Context Length=32k, Pretraining Tokens=640B, Ablation variant=SPLA without residual linear attention2026.01 | 58.9 | 85.2 | 61.8 | 75.1 | 80.3 | 96.2 | 74.5 | 49.8 | 24.9 | 67.3 | 78.1 | 90.2 | |
| NSAParameters=14B, Context Length=32k, Pretraining Tokens=640B2026.01 | 57 | 85.5 | 60.1 | 73.5 | 79.8 | 95.6 | 73.8 | 49.3 | 24.7 | 66.5 | 77.4 | 90.5 |