AI learns best when it knows where it can go wrong, just like Human. 💡
Why letting AI see its mistakes makes it smarter. i.e. LLMs learn better when they can see and fix their own reasoning mistakes
📚 https://arxiv.org/pdf/2410.16540
Original Problem 🔍:
Few-shot Chain-of-Thought (CoT) prompting boosts LLM reasoning but existing theoretical analyses isolate reasoning steps rather than treating them as an integrated process. This overlooks how real LLMs use all previous context when predicting next tokens.
-----
Solution in this Paper ⚡:
- Introduces "Coherent Chain-of-Thought (CoT)" that integrates earlier reasoning steps vs traditional "Stepwise ICL" that isolates them
- Shows Coherent CoT enables better error correction by considering previous reasoning context
- Proposes incorporating both correct and incorrect reasoning paths in demonstrations
- Validates that models are more sensitive to errors in intermediate steps than final outputs
- Recommends using model-generated incorrect paths vs handcrafted ones
-----
Key Insights from this Paper 💡:
• Chain-of-Thought (CoT) works better when reasoning steps stay connected vs isolated
• Models can self-correct intermediate errors when they maintain full context
• Exposing models to common reasoning mistakes improves performance
• Intermediate reasoning accuracy matters more than final answer accuracy
-----
Results 📊:
• 5-6% accuracy gains on reasoning tasks using error-aware demonstrations
• Better results with model-generated vs handcrafted incorrect paths
• Consistent improvements across GPT-3.5, GPT-4, Gemini Pro, DeepSeek
• Most gains in tracking (6.6%) and disambiguation (6.2%) tasks
Share this post