In-context retrieval on DROP
88.6AccuracyBERT-Judge
Evaluation Results
| Method | Links | |
|---|---|---|
| BERT-Judgeclean_name=BERT-as-a-Judge2026.04 | 88.6 | |
| Regex2026.04 | 77 | |
| LLM-Judgebackbone=Qwen-3 0.6B2026.04 | 69.3 | |
| Transformer++Parameter scale=410M2026.03 | 28.2 | |
| Hybrid Gated DeltaNet + M2RNN-3Parameter scale=410M2026.03 | 27.2 | |
| Hybrid M2RNNParameter scale=410M2026.03 | 26.8 | |
| Hybrid Mamba-2Parameter scale=410M2026.03 | 26.4 | |
| Hybrid Mamba-2 + M2RNN-3Parameter scale=410M2026.03 | 26.4 | |
| Hybrid Mamba-2 + M2RNN-1Parameter scale=410M2026.03 | 26.3 | |
| Hybrid Gated DeltaNetParameter scale=410M2026.03 | 26 | |
| Hybrid Gated DeltaNet + M2RNN-1Parameter scale=410M2026.03 | 25.6 | |
| Gated DeltaNetParameter scale=410M2026.03 | 24.9 | |
| M2RNNParameter scale=410M2026.03 | 22.9 | |
| Mamba-2Parameter scale=410M2026.03 | 19.5 | |
| GRUParameter scale=410M2026.03 | 17.2 | |
| RNNParameter scale=410M2026.03 | 14.6 |