Summary of Rl on Incorrect Synthetic Data Scales the Efficiency Of Llm Math Reasoning by Eight-fold, By Amrith Setlur et al.
RL on Incorrect Synthetic Data Scales the Efficiency of LLM Math Reasoning by Eight-Fold
by Amrith Setlur, Saurabh Garg, Xinyang Geng, Naman Garg, Virginia Smith, Aviral Kumar
First submitted to arxiv on: 20 Jun 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computation and Language (cs.CL)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary Training large language models (LLMs) by fine-tuning them on synthetic data generated by capable models is a promising approach, but its effectiveness remains unclear. This paper investigates the question for math reasoning through an empirical study and conceptual understanding. The typical approach of finetuning on synthetic correct or positive problem-solution pairs offers modest performance gains, but sampling more correct solutions from the fine-tuned learner itself followed by subsequent fine-tuning doubles efficiency. Training on model-generated positives can amplify spurious correlations, resulting in flat or inverse scaling trends. Using negative responses (model-generated incorrect responses) and constructing them such that training can recover utility can help unlearn spurious correlations and achieve consistent gains over positive data. This approach is equivalent to advantage-weighted reinforcement learning, inheriting robustness benefits. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper explores how fine-tuning large language models on synthetic data affects their performance for math reasoning. Researchers typically use synthetic correct or positive problem-solution pairs to improve model accuracy. However, they found that this approach only gives a small boost in performance. Instead, they tried generating more correct solutions from the fine-tuned model itself and re-training it. This method doubled the efficiency of using synthetic data. They also discovered that training on positive responses can create spurious correlations, which can be fixed by using negative responses (incorrect answers). By doing so, they achieved consistent gains in performance. |
Keywords
» Artificial intelligence » Fine tuning » Reinforcement learning » Synthetic data
