Reproduce “Solving LLM Repetition Problem in Production: A Comprehensive Study of Multiple Solutions” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Weiwei Wang, Weijie Zou, Jiyong Min
Researchers from Shenzhen Sunline Tech Co., Ltd. addressed the LLM repetition problem in production financial batch code interpretation by evaluating multiple solutions. Their study found that Beam Search with `early_stopping=True` effectively eliminated repetition, achieving a 0% repetition rate, and that Direct Preference Optimization (DPO) reduced repetition rates from 13-60% to 0-2% across various scenarios, restoring normal processing performance.
Key insights
- Theoretical analysis using Markov models confirmed that context repetition, self-reinforcement, and high initial probabilities are root causes trapping greedy decoding in infinite repetition loops.
- The parameter `early_stopping=True` is decisive and non-optional for Beam Search to effectively prevent repetition and restore normal processing, a critical finding for practical deployment.
- DPO fine-tuning, when trained with systematic 'power-of-2' repetition patterns, provides a universal, model-level solution applicable across different repetition types.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.