$\textit{New News}$: System-2 Fine-tuning for Robust Integration of New Knowledge

TL;DR

This paper introduces New News, a dataset for evaluating how well large language models can internalize new information through fine-tuning, and proposes System-2 Fine-tuning (Sys2-FT) with self-play protocols to improve knowledge integration.

Contribution

The paper presents a new dataset and a novel fine-tuning method, Sys2-FT, that enhances models' ability to internalize news information compared to naive fine-tuning.

Findings

Self-QA protocol improves in-weight learning of news.

Contextual shadowing effect degrades learning when rephrasing or QAs are used.

Preliminary scaling law of Sys2-FT performance emerges.

Abstract

Humans and intelligent animals can internalize new information and accurately internalize their implications to perform downstream tasks. While large language models (LLMs) can achieve this through in-context learning (ICL) when the information (news) is explicitly given as context, adequately integrating the information into model weights via fine-tuning remains challenging. In this paper, we introduce New News, a dataset composed of hypothetical yet plausible news spanning multiple domains (mathematics, coding, discoveries, leaderboards, events), accompanied by downstream evaluation questions whose correct answers critically depend on understanding and internalizing the news. First, we demonstrate a substantial gap between naive fine-tuning and in-context learning (FT-ICL gap) on our dataset. To address this gap, we explore a suite of self-play data generation protocols -- paraphrases, implications, and Self-QA -- designed to distill the knowledge processed by the model with context into the weights of the model, which we term System-2 Fine-tuning (Sys2-FT). We systematically evaluate ICL and Sys2-FT performance across data domains and model scales with the Qwen 2.5 family of models. Our results demonstrate that the Self-QA protocol of Sys2-FT significantly improves models' in-weight learning of the news while preserving general capabilities. Furthermore, we discover the contextual shadowing effect, where training with the news in context followed by its rephrases or QAs catastrophically degrades learning of the news. Finally, we show preliminary evidence of an emerging scaling law of Sys2-FT.