Can LLMs internalize System 2’s complex reasoning into System 1’s intuitive responses through iterative training?
This Paper hypothesize that LLMs, by mimicking human rapid skill acquisition, can generate fast, intuitive answers without additional training data, thus enhancing resource efficiency and reducing dependence on chains of thought (CoT).
Self-training framework allows LLMs to internalize complex reasoning, enhancing performance without explicit CoT prompts.
📚 https://arxiv.org/pdf/2409.03381
Original Problem 🧠:
LLMs show human-like proficiency in cognitive tasks, but it's unclear if they possess a dual-system framework similar to human cognition.
-----
Solution in this Paper 💡:
• Introduces CogniDual Framework for LLMs (CFLLMs)
• Assesses if LLMs can evolve from deliberate to intuitive responses through self-training
• Uses CoT as scaffold to reengineer non-CoT responses
• Applies methodology to Vicuna and Llama2 models of varying sizes
• Evaluates performance on reasoning datasets: GSM8K, ReClor, LogiQA 2.0
-----
Key Insights from this Paper 💡:
• LLMs can simulate dual-system characteristics of human cognition
• Self-training allows LLMs to internalize complex reasoning into intuitive responses
• Larger models require fewer examples to improve performance without CoT
• Framework more effective for tasks with significant accuracy gap between CoT and non-CoT use
-----
Results 📊:
• Marked improvement in non-CoT performance after self-training
• Negligible improvement on GSM8K dataset due to task contamination
• Larger models show greater ability to leverage limited data for improvement
• Performance enhancement correlates with increase in additional examples during self-practice
Share this post