Abstract2Appendix: Academic Reviews Enhance LLM Long-Context Capabilities

TL;DR

The zero-shot benchmark results indicate that aggregated high-quality human reviews are overwhelmingly preferred over LLM-generated responses, even for the most capable models like GPT-4o, which suggests that high-quality human reviews are extremely rich in information, reasoning, and long-context retrieval, capabilities that even the most advanced models have not fully captured.

Abstract

Large language models (LLMs) have shown remarkable performance across various tasks, yet their ability to handle long-context reading remains challenging. This study explores the effectiveness of leveraging high-quality academic peer review data for fine-tuning LLMs to enhance their long-context capabilities. We compare the Direct Preference Optimization (DPO) method with the Supervised Fine-Tuning (SFT) method, demonstrating DPO's superiority and data efficiency. Our experiments show that the fine-tuned model achieves a 4.04-point improvement over phi-3 and a 2.6\% increase on the Qasper benchmark using only 2000 samples. Despite facing limitations in data scale and processing costs, this study underscores the potential of DPO and high-quality data in advancing LLM performance. Additionally, the zero-shot benchmark results indicate that aggregated high-quality human reviews are overwhelmingly preferred over LLM-generated responses, even for the most capable models like GPT-4o. This suggests that high-quality human reviews are extremely rich in information, reasoning, and long-context retrieval, capabilities that even the most advanced models have not fully captured. These findings highlight the high utility of leveraging human reviews to further advance the field.

Figures (1)

• Figure 1: Token count distribution for combinations of paper and review (n=2005), with an average of 26353 tokens and standard deviations of 11774. In downstream experiments, we filter out samples longer than 33000