Scientific Reasoning on SciQ
97.08AccuracySelective Parameter Optimization
Evaluation results
| Method | Links | ||
|---|---|---|---|
| Selective Parameter OptimizationBackbone=LLaMA-32026.04 | 97.08 | 4.74 | |
| Selective Parameter OptimizationBackbone=GPT-J2026.04 | 96.92 | 3.34 | |
| RSLoRABackbone=LLaMA-32026.04 | 96.5 | 5.28 | |
| LoRA(μ)Backbone=GPT-J2026.04 | 96.5 | 3.43 | |
| LoRA(ν + μ)Backbone=GPT-J2026.04 | 96.5 | 3.48 | |
| EWCLoRABackbone=LLaMA-32026.04 | 96.3 | 4.88 | |
| LoRA(μ)Backbone=LLaMA-32026.04 | 96.2 | 5.05 | |
| RSLoRABackbone=GPT-J2026.04 | 96.2 | 3.5 | |
| LoRA(ν + μ)Backbone=LLaMA-32026.04 | 96.1 | 5.31 | |
| EWCLoRABackbone=GPT-J2026.04 | 96.1 | 3.38 | |
| BaseBackbone=LLaMA-32026.04 | 95.1 | 4.94 | |
| Qwen3-4B-ThinkingSteering Mechanism=Self-Reading Steering (+ CAA)2026.04 | 94.6 | — | |
| Qwen3-4B-ThinkingSteering Mechanism=CAA2026.04 | 94.2 | — | |
| Qwen3-4B-ThinkingSteering Mechanism=Base2026.04 | 93.9 | — | |
| BaseBackbone=GPT-J2026.04 | 91.6 | 3.28 | |
| R1-Distill-Llama-8BSteering Mechanism=Self-Reading Steering (+ CAA)2026.04 | 88.3 | — | |
| R1-Distill-Llama-8BSteering Mechanism=CAA2026.04 | 87.7 | — | |
| R1-Distill-Llama-8BSteering Mechanism=Base2026.04 | 87.3 | — | |
| R1-Distill-Qwen-7BSteering Mechanism=Self-Reading Steering (+ CAA)2026.04 | 86.7 | — | |
| R1-Distill-Qwen-7BSteering Mechanism=CAA2026.04 | 86.4 | — | |
| R1-Distill-Qwen-7BSteering Mechanism=Base2026.04 | 85.9 | — | |
| MoEModel size=7B, Evaluation Protocol=Zero-shot2026.05 | 80 | — | |
| FoEModel size=7B, Evaluation Protocol=Zero-shot2026.05 | 80 | — | |
| MoEModel size=1B, Evaluation Protocol=Zero-shot2026.05 | 77.7 | — | |
| FoEModel size=1B, Evaluation Protocol=Zero-shot2026.05 | 74.6 | — |