Summary of Bond: Aligning Llms with Best-of-n Distillation, by Pier Giuseppe Sessa et al.

BOND: Aligning LLMs with Best-of-N Distillation

by Pier Giuseppe Sessa, Robert Dadashi, Léonard Hussenot, Johan Ferret, Nino Vieillard, Alexandre Ramé, Bobak Shariari, Sarah Perrin, Abe Friesen, Geoffrey Cideron, Sertan Girgin, Piotr Stanczyk, Andrea Michi, Danila Sinopalnikov, Sabela Ramos, Amélie Héliou, Aliaksei Severyn, Matt Hoffman, Nikola Momchev, Olivier Bachem

First submitted to arxiv on: 19 Jul 2024

GrooveSquid.com Paper Summaries

GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!

Summary difficulty	Written by	Summary
High	Paper authors	High Difficulty Summary Read the original abstract here
Medium	GrooveSquid.com (original content)	Medium Difficulty Summary The proposed Best-of-N Distillation (BOND) algorithm aims to emulate the effectiveness of Best-of-N sampling in reinforcement learning from human feedback (RLHF), while reducing computational overhead at inference time. By forcing the distribution of generations from the policy to match the Best-of-N distribution, BOND seeks to improve quality and safety in large language models. The algorithm uses Jeffreys divergence to balance between mode-covering and mode-seeking behavior, and is demonstrated through experiments on abstractive summarization and Gemma models.
Low	GrooveSquid.com (original content)	Low Difficulty Summary Reinforcement learning from human feedback (RLHF) helps make large language models better and safer. A simple but powerful way to do this is called Best-of-N sampling. This paper proposes a new method called Best-of-N Distillation (BOND) that tries to get the same results as Best-of-N, but without using up too much computer power. BOND makes the policy’s generation distribution match the Best-of-N distribution. It uses something called Jeffreys divergence to make sure it’s doing a good job. The paper shows how well this method works by testing it on summarizing text and training Gemma models.