Zero-shot text reconstruction on Chinese official news text n=200 chunks (test)
99.1BERTScore F1Entropy-in-FreqBuckets
Evaluation Results
| Method | Links | |
|---|---|---|
| Entropy-in-FreqBucketsRetention rate (rkeep)=0.9, Decoder=Gemini 2.0 Flash2026.05 | 99.1 | |
| StepRetention rate (rkeep)=0.9, Decoder=Gemini 2.0 Flash2026.05 | 98.43 | |
| Word LengthRetention rate (rkeep)=0.9, Decoder=Gemini 2.0 Flash2026.05 | 98.36 | |
| Word FrequencyRetention rate (rkeep)=0.9, Decoder=Gemini 2.0 Flash2026.05 | 98.31 | |
| Qwen SFTRetention rate (rkeep)=0.9, Decoder=Qwen SFT2026.05 | 98.28 | |
| Entropy-LPRetention rate (rkeep)=0.9, Decoder=Gemini 2.0 Flash2026.05 | 98.23 | |
| EntropyRetention rate (rkeep)=0.9, Decoder=Gemini 2.0 Flash2026.05 | 98.18 | |
| Opt (w/ polish)Retention rate (rkeep)=0.9, Decoder=Gemini 2.0 Flash2026.05 | 98.03 | |
| Opt (w/o polish)Retention rate (rkeep)=0.9, Decoder=Gemini 2.0 Flash2026.05 | 97.67 | |
| Hybrid-0.3Retention rate (rkeep)=0.9, Decoder=Gemini 2.0 Flash2026.05 | 97.6 | |
| Hybrid-0.7Retention rate (rkeep)=0.9, Decoder=Gemini 2.0 Flash2026.05 | 97.47 | |
| Hybrid-0.5Retention rate (rkeep)=0.9, Decoder=Gemini 2.0 Flash2026.05 | 97.42 | |
| Word FrequencyRetention rate (rkeep)=0.7, Decoder=Gemini 2.0 Flash2026.05 | 95.14 | |
| Word LengthRetention rate (rkeep)=0.7, Decoder=Gemini 2.0 Flash2026.05 | 95.02 | |
| Qwen baseRetention rate (rkeep)=0.9, Decoder=Qwen base2026.05 | 95.02 | |
| StepRetention rate (rkeep)=0.7, Decoder=Gemini 2.0 Flash2026.05 | 94.68 | |
| Entropy-LPRetention rate (rkeep)=0.7, Decoder=Gemini 2.0 Flash2026.05 | 93.95 | |
| Opt (w/ polish)Retention rate (rkeep)=0.7, Decoder=Gemini 2.0 Flash2026.05 | 93.82 | |
| EntropyRetention rate (rkeep)=0.7, Decoder=Gemini 2.0 Flash2026.05 | 93.71 | |
| Opt (w/o polish)Retention rate (rkeep)=0.7, Decoder=Gemini 2.0 Flash2026.05 | 93.6 | |
| Qwen SFTRetention rate (rkeep)=0.7, Decoder=Qwen SFT2026.05 | 93.53 | |
| Entropy-in-FreqBucketsRetention rate (rkeep)=0.7, Decoder=Gemini 2.0 Flash2026.05 | 92.35 | |
| Hybrid-0.3Retention rate (rkeep)=0.7, Decoder=Gemini 2.0 Flash2026.05 | 92.18 | |
| Hybrid-0.7Retention rate (rkeep)=0.7, Decoder=Gemini 2.0 Flash2026.05 | 90.87 | |
| Hybrid-0.5Retention rate (rkeep)=0.7, Decoder=Gemini 2.0 Flash2026.05 | 90.68 | |
| Word FrequencyRetention rate (rkeep)=0.5, Decoder=Gemini 2.0 Flash2026.05 | 87.03 | |
| Qwen SFTRetention rate (rkeep)=0.5, Decoder=Qwen SFT2026.05 | 86.52 | |
| Word LengthRetention rate (rkeep)=0.5, Decoder=Gemini 2.0 Flash2026.05 | 86.48 | |
| StepRetention rate (rkeep)=0.5, Decoder=Gemini 2.0 Flash2026.05 | 86.36 | |
| Opt (w/ polish)Retention rate (rkeep)=0.5, Decoder=Gemini 2.0 Flash2026.05 | 85.07 | |
| EntropyRetention rate (rkeep)=0.5, Decoder=Gemini 2.0 Flash2026.05 | 84.98 | |
| Entropy-LPRetention rate (rkeep)=0.5, Decoder=Gemini 2.0 Flash2026.05 | 84.96 | |
| Hybrid-0.3Retention rate (rkeep)=0.5, Decoder=Gemini 2.0 Flash2026.05 | 84.81 | |
| Opt (w/o polish)Retention rate (rkeep)=0.5, Decoder=Gemini 2.0 Flash2026.05 | 84.6 | |
| Qwen baseRetention rate (rkeep)=0.7, Decoder=Qwen base2026.05 | 83.47 | |
| Hybrid-0.5Retention rate (rkeep)=0.5, Decoder=Gemini 2.0 Flash2026.05 | 83.37 | |
| Entropy-in-FreqBucketsRetention rate (rkeep)=0.5, Decoder=Gemini 2.0 Flash2026.05 | 82.53 | |
| Hybrid-0.7Retention rate (rkeep)=0.5, Decoder=Gemini 2.0 Flash2026.05 | 81.91 | |
| Hybrid-0.3Retention rate (rkeep)=0.3, Decoder=Gemini 2.0 Flash2026.05 | 76.59 | |
| Hybrid-0.5Retention rate (rkeep)=0.3, Decoder=Gemini 2.0 Flash2026.05 | 76.21 | |
| Entropy-LPRetention rate (rkeep)=0.3, Decoder=Gemini 2.0 Flash2026.05 | 75.94 | |
| Opt (w/ polish)Retention rate (rkeep)=0.3, Decoder=Gemini 2.0 Flash2026.05 | 75.83 | |
| EntropyRetention rate (rkeep)=0.3, Decoder=Gemini 2.0 Flash2026.05 | 75.59 | |
| Qwen SFTRetention rate (rkeep)=0.3, Decoder=Qwen SFT2026.05 | 74.51 | |
| Opt (w/o polish)Retention rate (rkeep)=0.3, Decoder=Gemini 2.0 Flash2026.05 | 74.36 | |
| Entropy-in-FreqBucketsRetention rate (rkeep)=0.3, Decoder=Gemini 2.0 Flash2026.05 | 74.32 | |
| Hybrid-0.7Retention rate (rkeep)=0.3, Decoder=Gemini 2.0 Flash2026.05 | 73.94 | |
| StepRetention rate (rkeep)=0.3, Decoder=Gemini 2.0 Flash2026.05 | 73.89 | |
| Word FrequencyRetention rate (rkeep)=0.3, Decoder=Gemini 2.0 Flash2026.05 | 73.84 | |
| Word LengthRetention rate (rkeep)=0.3, Decoder=Gemini 2.0 Flash2026.05 | 73.51 | |
| Qwen baseRetention rate (rkeep)=0.3, Decoder=Qwen base2026.05 | 68.31 | |
| Hybrid-0.3Retention rate (rkeep)=0.1, Decoder=Gemini 2.0 Flash2026.05 | 66.99 | |
| Hybrid-0.5Retention rate (rkeep)=0.1, Decoder=Gemini 2.0 Flash2026.05 | 66.65 | |
| Qwen baseRetention rate (rkeep)=0.5, Decoder=Qwen base2026.05 | 66.32 | |
| Opt (w/ polish)Retention rate (rkeep)=0.1, Decoder=Gemini 2.0 Flash2026.05 | 66.26 | |
| Qwen SFTRetention rate (rkeep)=0.1, Decoder=Qwen SFT2026.05 | 66.1 | |
| Hybrid-0.7Retention rate (rkeep)=0.1, Decoder=Gemini 2.0 Flash2026.05 | 65.94 | |
| Entropy-LPRetention rate (rkeep)=0.1, Decoder=Gemini 2.0 Flash2026.05 | 65.23 | |
| StepRetention rate (rkeep)=0.1, Decoder=Gemini 2.0 Flash2026.05 | 65.12 | |
| EntropyRetention rate (rkeep)=0.1, Decoder=Gemini 2.0 Flash2026.05 | 64.55 | |
| Word LengthRetention rate (rkeep)=0.1, Decoder=Gemini 2.0 Flash2026.05 | 63.71 | |
| Word FrequencyRetention rate (rkeep)=0.1, Decoder=Gemini 2.0 Flash2026.05 | 63.68 | |
| Entropy-in-FreqBucketsRetention rate (rkeep)=0.1, Decoder=Gemini 2.0 Flash2026.05 | 62.71 | |
| Opt (w/o polish)Retention rate (rkeep)=0.1, Decoder=Gemini 2.0 Flash2026.05 | 58.17 | |
| Qwen baseRetention rate (rkeep)=0.1, Decoder=Qwen base2026.05 | 57.12 |