Smart agents extract and combine task rules, helping LLMs reason better without examples.
Feedback, Guideline, and Tree-gather agents (FGT)
AI learns its own rulebook: no more spoon-feeding examples to language models
Original Problem 🔍:
Can we use only guidelines instead of few-shot examples in prompts for LLMs?
Solution in this Paper 🧠:
• Feedback, Guideline, and Tree-gather agents (FGT) framework: Feedback, Guideline, Tree-gather agents
• Feedback agent: Analyzes Q&A pairs for performance insights
• Guideline agent: Extracts guidelines from feedback
• Tree-Gather agent: Hierarchically aggregates guidelines
• Process prompt: Encourages LLM to show reasoning steps
Key Insights from this Paper 💡:
• Automatically learned guidelines can replace few-shot examples
• Tree-gather approach improves guideline aggregation
• Process prompt enhances guideline adherence and accuracy
• Effective for structured reasoning tasks (math, logic)
Results 📊:
• Outperforms baselines on Big-Bench Hard dataset
• Math calculating: 89.5% accuracy (vs 88.3% Leap)
• Logic reasoning: 93.9% accuracy (vs 89.1% Leap)
• Context understanding: 88.1% accuracy (vs 87.3% Leap)
• Surpasses few-shot and many-shot methods in most tasks
🧠 The paper investigates whether guidelines alone can be used effectively in prompts for LLMs, without relying on few-shot examples.
It explores if automatically learned task-specific guidelines can achieve comparable or better performance than traditional few-shot and chain-of-thought prompting methods.
