Deep Thinking

What is Deep Thinking?

深度不夠，長度來湊

Question -----> LLM -----> thinking process answer

• <think> ....... </think> -> Verification , Explore , Planning

Example: Alpha Go

• Alpha Go 的思考過程是用 MCTS (Monte Carlo Tree Search)

Test Time Scaling

• 思考的越多結果會越好！
• Paper: Scaling Scaling Laws with Board Games

Build Reasoning LLM Method

你可以混著用

Chain of Thought (CoT)

• Don't need to change the model

1. Few-shot CoT
   • 給Example叫LLM回答
   • Paper: Chain-of-Thought Prompting Elicits Reasoning in Large Language Models
2. Zero-shot CoT
   • 直接叫LLM "Let think step by step" 來回答問題
   • Paper: Large Language Models are Zero-Shot Reasoners
3. Long CoT
   • Paper: Towards Reasoning Era
4. Supervised CoT
   • 用更好的prompt來引導LLM回答問題
   • Paper: Supervised Chain of Thought

給Model reasoning 工作流程

• Don't need to change the model

How to explore?

用同一個問題問ＬＬＭ很多次，他會給出不同的答案

• Paper: Large Language Monkeys

How to choose the right answer?

1. Majority Vote (Self-consistency)
   • Paper: Self-Consistency Improves Chain of Thought Reasoning in Language Models
2. Confidence(used in CoT decoding)
   • Paper: Chain-of-Thought Reasoning Without Prompting
3. 加上 Verification
   • Paper: Training Verifiers to Solve Math Word Problems

Parallel vs. Sequential vs. Parallel + Sequential

How to verify the step?

• create process verifier to predict the correctness of each step
• Paper: Let's Verify Step by Step
• Paper: Math-Shepherd
• Beam Search
• Paper: Self-Evaluation Guided Beam Search for Reasoning
• Paper: Deductive Beam Search
• Monte Carlo Tree Search
• Paper: Monte Carlo Tree Search Boosts Reasoning
• ReST-MCTS
• Paper: ReST-MCTS*: LLM Self-Training via Process Reward Guided Tree Search
• Paper: Mutual Reasoning Makes Smaller LLMs Stronger Problem-Solvers

資訊

Imitation Learning and Reinforcement Learning Post-Training 的一種特例

Imitation Learning

• Need to change the model
• 教 model reasoning

Reasoning process data how to come from?

use LLM to generate reasoning process data 在對的answer 情況下去把readoning process 也視為是對的，並用於訓練

• Paper: rStar-Math
• Stream of Search (SoS)
• 也可以把錯誤的過程放進去
• Paper: Stream of Search (SoS)
• Paper: O1 Replication Journey

Reinforcement Learning

• Need to change the model
• Paper: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning