Text Generation on Wikitext-103 (Perplexity & Entropy)
32.88PerplexityRefined by Gemma3 27B
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| Refined by Gemma3 27Bbackbone=Gemma3 27B2026.03 | 32.88 | 7.14 | — | |
| WS-DFMstarting time t0=0.52026.03 | 64.68 | 7.16 | 4.2 | |
| WS-DFMstarting time t0=0.82026.03 | 67.86 | 7.19 | 1.7 | |
| Original DFM2026.03 | 69.06 | 7.42 | 8.33 | |
| FS-DFMSize (B)=1.70, Tokens (B)=65.5 + 4.1‡2025.09 | 73.65 | 7.61 | — | |
| FS-DFMSize (B)=1.30, Tokens (B)=65.5 + 8.2‡2025.09 | 77.21 | 7.4 | — | |
| FS-DFMSize (B)=0.17, Tokens (B)=65.5 + 24.6‡2025.09 | 90.49 | 7.29 | — | |
| SDTT (7 rounds)Size (B)=0.862025.09 | 131.71 | 7.78 | — | |
| LSTM2026.03 | 171.23 | 7.56 | — | |
| HDLM (γ = 0.05)Size (B)=0.16, Tokens (B)=5242025.09 | 192.74 | 8.13 | — | |
| SDTT (5 rounds)Size (B)=0.862025.09 | 220.81 | 8.09 | — | |
| HDLM (ϵ = 0.01)Size (B)=0.16, Tokens (B)=5242025.09 | 279.6 | 7.54 | — | |
| DFMSize (B)=1.30, Tokens (B)=65.52025.09 | 331.36 | 8.11 | — | |
| DFMSize (B)=0.17, Tokens (B)=65.52025.09 | 335.64 | 8.06 | — | |
| SDTT (3 rounds)Size (B)=0.862025.09 | 376.19 | 8.19 | — | |
| DFMSize (B)=1.70, Tokens (B)=65.52025.09 | 444.71 | 8.18 | — | |
| ReMDM–confSize (B)=0.17, Tokens (B)=2622025.09 | 549.86 | 8.35 | — | |
| ReMDM–capSize (B)=0.17, Tokens (B)=2622025.09 | 559.74 | 8.34 | — | |
| ReMDM–rescaleSize (B)=0.17, Tokens (B)=2622025.09 | 560.06 | 8.35 | — | |
| SEDDSize (B)=0.322025.09 | 602.09 | 8.42 | — | |
| SDTT (1 round)Size (B)=0.862025.09 | 612.38 | 8.38 | — | |
| MDLM–OWTSize (B)=0.17, Tokens (B)=2622025.09 | 837.58 | 8.47 | — | |
| ReMDM–loopSize (B)=0.17, Tokens (B)=2622025.09 | 1,358.84 | 8.41 | — |