Question Answering on ARC-Easy (val)
96.2AccuracyLoRA
Evaluation results
| Method | Links | |
|---|---|---|
| LoRABackbone=Qwen2.5-7B-Instruct, Training Epochs=1, Scoring Method=CLL2026.02 | 96.2 | |
| DoRABackbone=Qwen2.5-7B-Instruct, Training Epochs=1, Scoring Method=CLL2026.02 | 96.2 | |
| D2-LoRABackbone=Qwen2.5-7B-Instruct, Training Epochs=1, Scoring Method=CLL2026.02 | 96.2 | |
| DoRABackbone=Llama-3.2-3B-Instruct, Training Epochs=2, Scoring Method=CLL2026.02 | 87.6 | |
| D2-LoRABackbone=Llama-3.2-3B-Instruct, Training Epochs=2, Scoring Method=CLL2026.02 | 87.4 | |
| LoRABackbone=Llama-3.2-3B-Instruct, Training Epochs=2, Scoring Method=CLL2026.02 | 87.2 | |
| Bank of Values (BoV)Output=Ev[i], Layers=last 1/3, Coeff.=γv, Model Scale=780M, Zero-shot=true2026.06 | 71.6 | |
| x0WV variantOutput=x0WV, Layers=last 1/3, Coeff.=γv, Model Scale=780M, Zero-shot=true2026.06 | 71.4 | |
| V1 variant (last 1/3)Output=V1, Layers=last 1/3, Coeff.=γv, Model Scale=780M, Zero-shot=true2026.06 | 70.2 | |
| Standard AttentionOutput=V, Model Scale=780M, Zero-shot=true2026.06 | 69.1 | |
| V1 variant (every layer)Output=V1, Layers=every, Coeff.=1, Model Scale=780M, Zero-shot=true2026.06 | 67.7 | |
| SmolLM Ceiling (O(L^2))Parameters=135M, Training Tokens=∼600 Billion, Evaluation Mode=Zero-shot2026.04 | 61.74 | |
| CAWN (O(1))Parameters=150M, Training Tokens=∼5 Billion, Evaluation Mode=Zero-shot2026.04 | 45.45 | |
| TMMFormerBackbone=12-layer Transformer (d=768)2026.05 | 43.43 | |
| AdamFormerBackbone=12-layer Transformer (d=768)2026.05 | 43.39 | |
| YuriiFormerBackbone=12-layer Transformer (d=768)2026.05 | 43.06 | |
| AdamWFormerBackbone=12-layer Transformer (d=768)2026.05 | 41.88 | |
| VanillaTransformerBackbone=12-layer Transformer (d=768)2026.05 | 41.67 | |
| Pythia Baseline (O(L^2))Parameters=160M, Training Tokens=∼2.1 Billion, Evaluation Mode=Zero-shot2026.04 | 30.64 |