Multiple Choice Question Answering on ARC-Challenge 25-shot (test)
58.79AccuracyDeepSeek-V2-Lite-Chat with training-free loop wrapper
Evaluation results
| Method | Links | ||
|---|---|---|---|
| DeepSeek-V2-Lite-Chat with training-free loop wrapperModel=DeepSeek-V2-Lite-Chat, Evaluation Protocol=25-shot, Loop Setting=Loop (Ours), Backbone Architecture=MoE, Iteration Mode=layer-mode, Integration Recipe=3-stage Runge–Kutta2026.05 | 58.79 | 0.85 | |
| DeepSeek-V2-Lite-ChatModel=DeepSeek-V2-Lite-Chat, Evaluation Protocol=25-shot, Loop Setting=Base, Backbone Architecture=MoE2026.05 | 57.94 | — | |
| Qwen1.5-MoE-A2.7B with training-free loop wrapperModel=Qwen1.5-MoE-A2.7B, Evaluation Protocol=25-shot, Loop Setting=Loop (Ours), Backbone Architecture=MoE, Iteration Mode=layer-mode, Integration Recipe=3-stage Runge–Kutta2026.05 | 50.6 | 2.3 | |
| Qwen1.5-MoE-A2.7BModel=Qwen1.5-MoE-A2.7B, Evaluation Protocol=25-shot, Loop Setting=Base, Backbone Architecture=MoE2026.05 | 48.29 | — |