Language Modeling on LAMBADA (dev)
12.34PerplexityFusionGated-FIRNet
Evaluation Results
| Method | Links | ||
|---|---|---|---|
| FusionGated-FIRNetDiscovery Source=AI Discovered, Abbreviation=FG2026.03 | 12.34 | 47.53 | |
| PathGate-FusionNetDiscovery Source=AI Discovered, Abbreviation=PG2026.03 | 12.62 | 47.6 | |
| Hier-GateNetDiscovery Source=AI Discovered, Abbreviation=H2026.03 | 13.06 | 46.56 | |
| Gated DeltaNetDiscovery Source=Human Discovered2026.03 | 13.31 | 46.26 | |
| Mamba2Discovery Source=Human Discovered2026.03 | 13.33 | 46.24 | |
| Content-SharpRouterDiscovery Source=AI Discovered, Abbreviation=C2026.03 | 13.45 | 46.13 | |
| DeltaNetDiscovery Source=Human Discovered2026.03 | 13.63 | 45.47 | |
| AdaMulti-PathGateNetDiscovery Source=AI Discovered, Abbreviation=AM2026.03 | 13.75 | 45.04 | |
| Universal Transformerhalting=dynamic2018.07 | 134 | 0.22 | |
| Neural cache model2016.12 | 138 | — | |
| Neural Cache2018.07 | 139 | — | |
| Universal Transformersteps=8, halting=fixed2018.07 | 192 | 0.21 | |
| Universal Transformersteps=9, halting=fixed2018.07 | 214 | 0.21 | |
| Universal Transformersteps=6, halting=fixed2018.07 | 279 | 0.18 | |
| WB5+cacheDescription=5-gram language model with cache2016.12 | 768 | — | |
| WB5Description=5-gram language model with Witten-Bell smoothing2016.12 | 3,125 | — | |
| LSTM-1024Hidden units=1024, Implementation=Current paper reproduction2016.12 | 4,088 | — | |
| LSTM2018.07 | 4,966 | 0 | |
| Transformer2018.07 | 5,122 | 0 | |
| LSTM-512Hidden units=5122016.12 | 5,357 | — |