Technology

Notable OpenAI's O1 cot is hidden with only a summary available sometimes

From Deepseek:

Chain of Thought (CoT) in LLMs refers to a prompting technique that guides large language models to articulate intermediate reasoning steps when solving a problem, mimicking human-like logical progression. Here's a concise breakdown:

1. Purpose: Enhances performance on complex tasks (e.g., math, logic, commonsense reasoning) by breaking problems into sequential steps, reducing errors from direct, unstructured answers.

2. Mechanism:

   • Prompting Strategy: Users provide examples (few-shot) or explicit instructions (zero-shot, e.g., "Let's think step by step") to encourage step-by-step explanations.
   • Output Structure: The model generates tokens sequentially, simulating a reasoned pathway to the answer, even though internal processing remains parallel (via transformer architecture).

3. Benefits:

   • Accuracy: Improves results on multi-step tasks by isolating and addressing each component.
   • Transparency: Makes the model's "thinking" visible, aiding debugging and trust.

4. Variants:

   • Few-Shot CoT: Examples with detailed reasoning are included in the prompt.
   • Zero-Shot CoT: Direct instructions trigger step-by-step output without examples.
   • Self-Consistency: Aggregates answers from multiple CoT paths to select the most consistent one.

5. Effectiveness: Particularly impactful for tasks requiring structured reasoning, while less critical for simple queries. Research shows marked accuracy gains in benchmarks like math word problems.

In essence, CoT leverages the model's generative capability to externalize reasoning, bridging the gap between opaque model decisions and interpretable human problem-solving.

Example: