Reasoning trace (generated by Gemini)
Reasoning Trace Generation: According to the paper, they used the Google Gemini Flash Thinking API to generate reasoning traces. The process worked like this:
# Simplified version of their process
async def generate_reasoning_trace(question):
# Using Google Gemini Flash Thinking API
response = await gemini.flash_thinking(
prompt=question,
mode="experimental-1219" # As mentioned in paper
)
# The response contains:
reasoning_trace = response.thinking # Step-by-step reasoning
final_answer = response.answer
return {
'question': question,
'reasoning_trace': reasoning_trace,
'solution': final_answer
}
Key aspects of the reasoning trace generation:
- They used Gemini's Flash Thinking mode which is specifically designed for step-by-step reasoning
- They extracted both the reasoning process and the final answer
- Each trace follows a structured format:
- Breaking down the problem
- Showing intermediate steps
- Explaining key insights
- Reaching a conclusion
The paper mentions that they got high-quality traces because Gemini's Flash Thinking mode is specifically optimized for showing detailed reasoning steps. They then used these traces for training their model to learn this reasoning pattern.