Can LLMs Help Uncover Insights about LLMs? A Large-Scale, Evolving Literature Analysis of Frontier LLMs

Jungsoo Park, Junmo Kang, Gabriel Stanovsky, Alan Ritter

Abstract

The surge of LLM studies makes synthesizing their findings challenging. Analysis of experimental results from literature can uncover important trends across studies, but the time-consuming nature of manual data extraction limits its use. Our study presents a semi-automated approach for literature analysis that accelerates data extraction using LLMs. It automatically identifies relevant arXiv papers, extracts experimental results and related attributes, and organizes them into a structured dataset, LLMEvalDB. We then conduct an automated literature analysis of frontier LLMs, reducing the effort of paper surveying and data extraction by more than 93% compared to manual approaches. We validate LLMEvalDB by showing that it reproduces key findings from a recent manual analysis of Chain-of-Thought (CoT) reasoning and also uncovers new insights that go beyond it, showing, for example, that in-context examples benefit coding & multimodal tasks but offer limited gains in math reasoning tasks compared to zero-shot CoT. Our automatically updatable dataset enables continuous tracking of target models by extracting evaluation studies as new data becomes available. Through LLMEvalDB and empirical analysis, we provide insights into LLMs while facilitating ongoing literature analyses of their behavior.

Introduction

The pace of AI research is currently overwhelming. Thousands of papers are uploaded to preprint servers like arXiv every month, each testing different models on various benchmarks with diverse prompting configurations. For researchers, performing a "meta-analysis"—synthesizing results across many studies to find broad trends—has become nearly impossible to do manually without the data becoming obsolete by the time the analysis is finished.

This study addresses this "information overload" by asking: Can we use frontier LLMs to automatically extract, structure, and analyze the experimental results buried in the very papers written about them? By creating a semi-automated pipeline, we have developed LLMEvalDB, a structured dataset that reduces the manual effort of literature review by over 93%.

Methodology: The LLMEvalDB Pipeline

To turn unstructured LaTeX source files from arXiv into a structured database, we built a multi-stage pipeline:

  1. Preprocessing & Filtering: The system scans arXiv for papers in specific CS categories (AI, CL, LG, CV) and uses a fast model (Llama-3.1-70B) to identify "leaderboard-style" tables containing results for target models.
  2. Schema-Driven Extraction: Using GPT-4o, the system extracts specific attributes from these tables, such as the dataset name, metric (e.g., Accuracy, F1), prompting method (e.g., CoT), and the number of few-shot examples.
  3. Context Augmentation: Because tables often lack detail, the LLM reads the surrounding text of the paper to fill in missing information and clarify experimental setups.
  4. Dataset Description Generation: The system generates summaries of the benchmarks used, either from the LLM’s internal knowledge or by following citations to the original dataset papers.
The semi-automated pipeline for continuous data extraction and literature analysis of LLM research.

Key Results and Insights

The primary output, LLMEvalDB, contains 18,127 experimental records extracted from 1,737 papers, focusing on frontier models like GPT-4, GPT-4o, Claude 3 Opus, and Gemini 1.0 Pro.

Efficiency and Accuracy

The automated pipeline is 93% faster than manual data extraction. While a single researcher might take 350 hours to annotate the tables in this study, the pipeline completes the task in a day for under $500 in API costs. Human evaluation showed the system is highly reliable, achieving 100% accuracy on model names and 95% on dataset names and few-shot counts.

Scientific Insights

  • The "Math vs. Coding" Split: While CoT is essential for math and symbolic reasoning, adding few-shot examples (ICL) provides massive boosts for coding and multimodal tasks but offers surprisingly limited gains for math compared to a simple zero-shot CoT setup.
  • Where Prompting Fails: "Expert Knowledge" tasks (like specialized medical or legal benchmarks) are the most resistant to prompting improvements. If the model lacks the underlying facts, "thinking step-by-step" cannot compensate.
  • CoT is a "Lift," Not a "Rescue": Data shows that while few-shot examples improve CoT performance, they don't actually increase CoT's relative advantage over standard prompting. If CoT wasn't working well for a task in a zero-shot setting, adding examples usually won't suddenly make it the superior choice.

Conclusion and Future Work

LLMEvalDB demonstrates the feasibility of using LLMs to monitor the progress of the field itself. Future work will expand the model scope to include open-source models and newer "reasoning" models like GPT-4o1. Additionally, we aim to improve canonicalization challenges—matching different names for the same dataset—and explore the categorization of fine-tuning methods to allow for more granular comparative analyses.

Summary updated

Abstract

The surge of LLM studies makes synthesizing their findings challenging. Analysis of experimental results from literature can uncover important trends across studies, but the time-consuming nature of manual data extraction limits its use. Our study presents a semi-automated approach for literature analysis that accelerates data extraction using LLMs. It automatically identifies relevant arXiv papers, extracts experimental results and related attributes, and organizes them into a structured dataset, LLMEvalDB. We then conduct an automated literature analysis of frontier LLMs, reducing the effort of paper surveying and data extraction by more than 93% compared to manual approaches. We validate LLMEvalDB by showing that it reproduces key findings from a recent manual analysis of Chain-of-Thought (CoT) reasoning and also uncovers new insights that go beyond it, showing, for example, that in-context examples benefit coding & multimodal tasks but offer limited gains in math reasoning tasks compared to zero-shot CoT. Our automatically updatable dataset enables continuous tracking of target models by extracting evaluation studies as new data becomes available. Through LLMEvalDB and empirical analysis, we provide insights into LLMs while facilitating ongoing literature analyses of their behavior.
BibTeX
@inproceedings{park2025can,
  author = {Jungsoo Park and Junmo Kang and Gabriel Stanovsky and Alan Ritter},
  title = {{Can LLMs Help Uncover Insights about LLMs? A Large-Scale, Evolving Literature Analysis of Frontier LLMs}},
  booktitle = {ACL},
  year = {2025},
  url = {https://openreview.net/forum?id=gm7BdqMiR3},
}
Figure