# From Frontier to Point-in-Time Models

> Methods for adapting open-source frontier models into point-in-time models that substantially reduce look-ahead bias, with benchmarks, usage guidelines and production-ready inference code.

Authors: Humzah Merchant, Bradford Levy (University of Chicago, Booth School of Business).

## Contents

- [Overview](https://frontiertopit.com/#overview): Why look-ahead bias matters, and what this work contributes.
- [Benchmarking](https://frontiertopit.com/#benchmarking): Utility and look-ahead-bias results against frontier baselines and point-in-time models.
- [Look-ahead bias](https://frontiertopit.com/#lookahead-intro): What look-ahead bias is and why LLMs exhibit it.
- [Long-context retrieval](https://frontiertopit.com/#long-context-intro): Needle-in-a-haystack retrieval from 4K to 128K tokens, by needle depth.
- [Application to Common Social Science Research Tasks](https://frontiertopit.com/#tasks): Extraction, sentiment classification, and summarization results for the point-in-time setups.
- [Inference Costs](https://frontiertopit.com/#production): Steady-state decode throughput and cost per million output tokens.
- [Usage Guidelines](https://frontiertopit.com/#guidelines): How to prompt these models, how to apply the method, and its known weaknesses.

## Output Explorer

- [Output Explorer](https://frontiertopit.com/explorer): every benchmark prompt (10,025 across
  look-ahead bias, extraction, summarization and long-context) with each model’s first response,
  laid side by side. Not included in llms-full.txt.

## Full text

- [llms-full.txt](https://frontiertopit.com/llms-full.txt): the complete page as Markdown, including all
  example transcripts, the LaTeX source of every equation, and every figure’s underlying numbers as a table.

## Code and models

- [Center for Applied AI](https://www.chicagobooth.edu/research/center-for-applied-artificial-intelligence)
- [GitHub repository](https://github.com/fin-ai-lab/frontier-to-pit/)
- [Hugging Face models](https://huggingface.co/fin-ai-lab)