Decoupled Reasoning with Implicit Fact Tokens (DRIFT): A Dual-Model Framework for Efficient Long-Context Inference

TL;DR

DRIFT introduces a dual-model framework that decouples knowledge extraction from reasoning in LLMs, enabling more efficient long-context inference by dynamically compressing document chunks into implicit fact tokens.

Contribution

The paper presents DRIFT, a novel dual-model architecture that explicitly separates knowledge extraction from reasoning, improving long-context inference in LLMs.

Findings

Significantly outperforms baselines on long-context tasks.

Enhances the effective context window and reasoning capabilities.

Provides a scalable, efficient inference paradigm.

Abstract

The integration of extensive, dynamic knowledge into Large Language Models (LLMs) remains a significant challenge due to the inherent entanglement of factual data and reasoning patterns. Existing solutions, ranging from non-parametric Retrieval-Augmented Generation (RAG) to parametric knowledge editing, are often constrained in practice by finite context windows, retriever noise, or the risk of catastrophic forgetting. In this paper, we propose DRIFT, a novel dual-model architecture designed to explicitly decouple knowledge extraction from the reasoning process. Unlike static prompt compression, DRIFT employs a lightweight knowledge model to dynamically compress document chunks into implicit fact tokens conditioned on the query. These dense representations are projected into the reasoning model's embedding space, replacing raw, redundant text while maintaining inference accuracy. Extensive experiments show that DRIFT significantly improves performance on long-context tasks, outperforming strong baselines among comparably sized models. Our approach provides a scalable and efficient paradigm for extending the effective context window and reasoning capabilities of LLMs. Our code is available at https://github.com/Lancelot-Xie/DRIFT.