"ProgCo: Program Helps Self-Correction of Large Language Models"

LLMs become their own code reviewers by writing verification programs to catch mistakes

ProgCo enables LLMs to verify and fix their own mistakes through self-generated programs, significantly improving accuracy across complex tasks.

-----

https://arxiv.org/abs/2501.01264

🤔 Original Problem:

LLMs struggle with self-correction, especially in complex reasoning tasks. Current methods fail to effectively detect errors and often provide misleading feedback, leading to incorrect revisions.

-----

💡 Solution in this Paper:

→ ProgCo introduces program-driven verification (ProgVe) that generates and executes verification pseudo-programs to validate model outputs

→ These programs can express complex verification logic beyond simple checklists

→ Program-driven refinement (ProgRe) performs dual optimization of both responses and verification programs

→ The system uses contrast-based refinement to avoid misleading feedback

→ Integration with Python tools enhances verification capabilities for numerical operations

-----

🎯 Key Insights:

→ Programs are more effective than natural language for expressing verification logic

→ LLMs can act as program executors while incorporating their knowledge

→ Dual refinement of responses and programs improves accuracy

→ Contrast-based refinement helps avoid error propagation

-----

📊 Results:

→ Improved GPT-3.5 performance by 4.62% on IFEval(Pr), 3.23% on IFEval(Ins)

→ Enhanced mathematical reasoning with 5.84% gain on GSM8K, 5.8% on MATH

→ Consistently outperformed baseline methods across all benchmarks

→ Further improvements when combined with Python executor tools