Don't Think Too Much

Reasoning length and answer accuracy

• Paper: Revisiting the Test-Time Scaling of o1-like Models
• reasoning length ⬆️, answer accuracy ⬇️ => Question difficulty ⬆️ , reasoning length ⬆️, answer accuracy ⬇️ ????

資訊

The four methods are compared with Deep Thinking

Avoid reasoning length too long

1-[1] Chain of Draft

• paper: Chain of Draft: Thinking Faster by Writing Less

2

並沒有多講什麼，就自行降低Bean Search 的數目.....

3-[1]

• paper: Towards Thinking-Optimal Scaling of Test-Time Compute for LLM Reasoning
• 選擇最短的 Reasoning Processing 作為Training Data

3-[2] From Explicit CoT to Implicit CoT

• paper: From Explicit CoT to Implicit CoT:

4-[1]

• 即使answer 是對的 Reasoning process length 還要低於平均才會給出正面的評價
• paper: O1-Pruner
• paper: Kimi k1.5
• paper: Training Language Models to Reason Efficiently

4-[2]

• paper: L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning