Text reconstruction on Chinese Wikipedia n = 200 (test)
98.49BERTScore F1Entropy-in-FreqBuckets
Evaluation Results
| Method | Links | |
|---|---|---|
| Entropy-in-FreqBucketsr_keep=0.9, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 98.49 | |
| Entropy-LPr_keep=0.9, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 97.66 | |
| Qwen SFTr_keep=0.9, Decoder=Qwen SFT2026.05 | 97.63 | |
| Word Frequencyr_keep=0.9, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 97.61 | |
| Entropyr_keep=0.9, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 97.54 | |
| Word Lengthr_keep=0.9, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 97.49 | |
| Stepr_keep=0.9, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 96.86 | |
| Opt (w/ polish)r_keep=0.9, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 96.63 | |
| Hybrid-0.3r_keep=0.9, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 96.3 | |
| Opt (w/o polish)r_keep=0.9, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 96.04 | |
| Hybrid-0.7r_keep=0.9, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 95.79 | |
| Hybrid-0.5r_keep=0.9, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 95.68 | |
| Qwen baser_keep=0.9, Decoder=Qwen base2026.05 | 95.27 | |
| Entropy-in-FreqBucketsr_keep=0.7, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 93.07 | |
| Word Frequencyr_keep=0.7, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 92.42 | |
| Word Lengthr_keep=0.7, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 92.41 | |
| Stepr_keep=0.7, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 92.13 | |
| Entropy-LPr_keep=0.7, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 91.36 | |
| Entropyr_keep=0.7, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 91.27 | |
| Qwen SFTr_keep=0.7, Decoder=Qwen SFT2026.05 | 90.78 | |
| Opt (w/ polish)r_keep=0.7, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 90.61 | |
| Opt (w/o polish)r_keep=0.7, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 90.51 | |
| Hybrid-0.3r_keep=0.7, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 89.34 | |
| Hybrid-0.5r_keep=0.7, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 87.58 | |
| Hybrid-0.7r_keep=0.7, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 86.96 | |
| Word Lengthr_keep=0.5, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 84.3 | |
| Word Frequencyr_keep=0.5, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 84.05 | |
| Entropy-in-FreqBucketsr_keep=0.5, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 83.52 | |
| Qwen SFTr_keep=0.5, Decoder=Qwen SFT2026.05 | 83.32 | |
| Stepr_keep=0.5, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 83.03 | |
| Entropy-LPr_keep=0.5, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 82.85 | |
| Entropyr_keep=0.5, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 82.73 | |
| Opt (w/o polish)r_keep=0.5, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 82.5 | |
| Opt (w/ polish)r_keep=0.5, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 82.11 | |
| Hybrid-0.3r_keep=0.5, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 82.04 | |
| Hybrid-0.5r_keep=0.5, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 80.75 | |
| Qwen baser_keep=0.7, Decoder=Qwen base2026.05 | 80.55 | |
| Hybrid-0.7r_keep=0.5, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 78.5 | |
| Entropy-LPr_keep=0.3, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 75.6 | |
| Entropyr_keep=0.3, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 75.43 | |
| Hybrid-0.3r_keep=0.3, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 75.18 | |
| Entropy-in-FreqBucketsr_keep=0.3, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 74.93 | |
| Hybrid-0.5r_keep=0.3, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 74.7 | |
| Opt (w/ polish)r_keep=0.3, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 74.02 | |
| Opt (w/o polish)r_keep=0.3, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 72.77 | |
| Word Frequencyr_keep=0.3, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 72.52 | |
| Word Lengthr_keep=0.3, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 72.36 | |
| Stepr_keep=0.3, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 72.14 | |
| Hybrid-0.7r_keep=0.3, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 71.89 | |
| Qwen SFTr_keep=0.3, Decoder=Qwen SFT2026.05 | 70.58 | |
| Qwen baser_keep=0.3, Decoder=Qwen base2026.05 | 67.47 | |
| Qwen baser_keep=0.5, Decoder=Qwen base2026.05 | 66.38 | |
| Entropy-LPr_keep=0.1, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 65.63 | |
| Entropyr_keep=0.1, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 65.15 | |
| Hybrid-0.3r_keep=0.1, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 65.1 | |
| Entropy-in-FreqBucketsr_keep=0.1, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 64.81 | |
| Hybrid-0.5r_keep=0.1, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 64.13 | |
| Opt (w/ polish)r_keep=0.1, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 63.43 | |
| Stepr_keep=0.1, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 62.86 | |
| Qwen SFTr_keep=0.1, Decoder=Qwen SFT2026.05 | 62.86 | |
| Word Frequencyr_keep=0.1, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 62.68 | |
| Word Lengthr_keep=0.1, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 62.65 | |
| Hybrid-0.7r_keep=0.1, Decoder=Gemini 2.0 Flash zero-shot, surprisal_pipeline=Chinese GPT-22026.05 | 61.88 | |
| Qwen baser_keep=0.1, Decoder=Qwen base2026.05 | 56.8 | |
| Opt (w/o polish)r_keep=0.1, Decoder=Gemini 2.0 Flash zero-shot2026.05 | 56.35 |