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

TL;DR

This work tackles how to robustly embed newly encountered information into the weights of large language models, addressing the gap between fine-tuning and in-context learning (FT-ICL). It introduces New News, a dataset of 75 plausible news items across five domains with 375 downstream questions, designed to test deep internalization, and System-2 Fine-tuning (Sys2-FT), a family of self-play data-generation protocols (Paraphrase, Implication, Self-QA) that distill context-derived knowledge into model weights. The Self-QA protocol stands out, delivering significant improvements across model scales and domains, and sometimes approaching ICL performance while preserving general capabilities. The study also uncovers the contextual shadowing effect and the curse of overexposure, and reports preliminary evidence for a scaling law in Sys2-FT, suggesting larger models can be more data-efficient learners. Overall, Sys2-FT offers a principled, architecture-agnostic pathway to integrate dynamic knowledge into models, with implications for continual learning and deployment in changing information environments.

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.

Figures (43)

• Figure 1: Overview. We introduce New News, a dataset consisting of hypothetical but not counterfactual news which has rich downstream implications in order to test the ability to integrate new information. To update model weights, we explore a suite of methods we dub System-2 Fine-tuning (Sys2-FT). Sys2-FT involves generating synthetic datasuch as paraphrases, implications and QA pairs from the news using models' native in-context learning abilities. We find that our specific Self-QA protocol performs significantly better than naive FT.
• Figure 2: The fine-tuning to in-context learning gap (FT-ICL gap).New News clearly demonstrates the FT-ICL gap of downstream QA accuracy for all splits. Small models (0.5B, 1.5B) struggle to find the right answer even given the news demonstrating their inability to reason over hypothetical scenarios. Larger models (3B$\sim$32B) shows high ICL accuracy, but naively fine-tuning the model with the news shows poor performance.
• Figure 3: System-2 Fine-tuning Protocols. Given a topic (usually just the main entity of the news) and news, we set up three System-2 Fine-tuning protocols: Paraphrase protocol prompts the model to generate paraphrases of the news in a sequential manner to enhance diversity; Implication protocol prompts the model to reason about implications/consequences of the given news; Self-QA protocol first prompts the model to generate questions that is related to the news, then generates answers using another conversation with the news in context. All protocols result in replay elements as data that is further arranged in a conversation format to fine-tune the model. The fire emoji denotes tokens where the loss is computed, which we take the usual supervised fine-tuning format from the assistant tokens. See App. \ref{['app:generation']} for further methodological details.
• Figure 4: System-2 Fine-Tuning (Sys2-FT). a) Sys2-FT results on Qwen2.5-14B-Instruct. Sys2-FT bridges the gap between naive fine tuning and in-context learning. We find Self-QA as the best Sys2-FT method among the ones we explored. We also notice that quantitative domains (math and coding) benefit the most from Sys2-FT. b) Scaling properties of Sys2-FT on the Qwen2.5 family of models. We find that bigger models achieve higher performance: Sys2-FT is a scalable method for new knowledge integration.
• Figure 5: Context Prefix Format. The replay element (here Self-QA) is prefixed by a small conversation containing the news. The original FT data is denoted in dotted lines.