Compositional Generalization on Evaluation Dataset (Fold 2 Seen)
63.63ScoreGemma-7B + COGLM
Evaluation Results
| Method | Links | |
|---|---|---|
| Gemma-7B + COGLMFine-tuning Strategy=LoRA, Model Backbone=Gemma-7B, Method Components=COGLM2026.01 | 63.63 | |
| LLaMA-3-8B + COGLMFine-tuning Strategy=LoRA, Model Backbone=LLaMA-3-8B, Method Components=COGLM2026.01 | 63.46 | |
| Mistral-7B + COGLMFine-tuning Strategy=LoRA, Model Backbone=Mistral-7B, Method Components=COGLM2026.01 | 63.17 | |
| Mistral-7BFine-tuning Strategy=LoRA, Model Backbone=Mistral-7B2026.01 | 62.48 | |
| Qwen2.5-7B + COGLMFine-tuning Strategy=LoRA, Model Backbone=Qwen2.5-7B, Method Components=COGLM2026.01 | 61.85 | |
| LLaMA-3-8BFine-tuning Strategy=LoRA, Model Backbone=LLaMA-3-8B2026.01 | 61.81 | |
| Qwen2.5-7BFine-tuning Strategy=LoRA, Model Backbone=Qwen2.5-7B2026.01 | 61.61 | |
| Gemma-7BFine-tuning Strategy=LoRA, Model Backbone=Gemma-7B2026.01 | 57.37 | |
| DeepSeek-7B + COGLMFine-tuning Strategy=LoRA, Model Backbone=DeepSeek-7B, Method Components=COGLM2026.01 | 52.04 | |
| DeepSeek-7BFine-tuning Strategy=LoRA, Model Backbone=DeepSeek-7B2026.01 | 49.52 | |
| DeepSeek V3Evaluation Protocol=Zero-shot, Model Backbone=DeepSeek V32026.01 | 45.27 | |
| DeepSeek V3 + FS*Evaluation Protocol=Few-shot, Model Backbone=DeepSeek V32026.01 | 44.18 | |
| Claude 3.5 Sonnet + FS*Evaluation Protocol=Few-shot, Model Backbone=Claude 3.5 Sonnet2026.01 | 43.18 | |
| Claude 3.5 SonnetEvaluation Protocol=Zero-shot, Model Backbone=Claude 3.5 Sonnet2026.01 | 41.11 | |
| GPT-4oEvaluation Protocol=Zero-shot, Model Backbone=GPT-4o2026.01 | 40 | |
| GPT-4o + FS*Evaluation Protocol=Few-shot, Model Backbone=GPT-4o2026.01 | 39.12 | |
| Llama 3 (70B) + FS*Evaluation Protocol=Few-shot, Model Backbone=Llama 3 (70B)2026.01 | 33.49 | |
| Llama 3 (70B)Evaluation Protocol=Zero-shot, Model Backbone=Llama 3 (70B)2026.01 | 23.66 |