Test1

🧩 Background

Chain-of-Thought (CoT) prompting enables powerful multi-step reasoning in LLMs. But transferring this ability into small language models (SLMs) remains a major challenge.

---

🧪 Experiment Design

We study three axes:

1. Granularity of intermediate steps
2. Supervision format: explanation-only vs answer+reasoning
3. Teacher model: GPT-3.5 vs GPT-4

7 datasets are used: GSM8K, SVAMP, DROP, etc.

---

📈 Key Findings

• Fine-grained step supervision leads to more generalizable reasoning
• Answer-only distillation fails to transfer reasoning skills
• GPT-4 outputs yield more structured logic chains than GPT-3.5

---

💡 Takeaway

Distillation is not copying — it’s translation.

“The way we teach reasoning determines what reasoning emerges.”