Code and visualizations for the paper "Inference-Time Reward Hacking in Large Language Models".

This repository contains implementations and demonstrations of inference-time reward hacking in LLMs and methods to mitigate it, including:

• Best-of-n (BoN)
• Soft Best-of-n (SBoN)
• Best-of-Poisson (BoP) [Our Method]
• HedgeTune algorithm for finding optimal inference parameters

• src/: Implementation of methods and experiments
• visuals/: Interactive visualizations and demos
• docs/: Paper and additional documentation

For more details, check out our paper on arXiv.