Reinforced Fast Weights with Next-Sequence Prediction

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Hwang, Hee Seung, Wu, Xindi, Chun, Sanghyuk, Russakovsky, Olga
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912912086401024
author Hwang, Hee Seung
Wu, Xindi
Chun, Sanghyuk
Russakovsky, Olga
author_facet Hwang, Hee Seung
Wu, Xindi
Chun, Sanghyuk
Russakovsky, Olga
contents Fast weight architectures offer a promising alternative to attention-based transformers for long-context modeling by maintaining constant memory overhead regardless of context length. However, their potential is limited by the next-token prediction (NTP) training paradigm. NTP optimizes single-token predictions and ignores semantic coherence across multiple tokens following a prefix. Consequently, fast weight models, which dynamically update their parameters to store contextual information, learn suboptimal representations that fail to capture long-range dependencies. We introduce REFINE (Reinforced Fast weIghts with Next sEquence prediction), a reinforcement learning framework that trains fast weight models under the next-sequence prediction (NSP) objective. REFINE selects informative token positions based on prediction entropy, generates multi-token rollouts, assigns self-supervised sequence-level rewards, and optimizes the model with group relative policy optimization (GRPO). REFINE is applicable throughout the training lifecycle of pre-trained language models: mid-training, post-training, and test-time training. Our experiments on LaCT-760M and DeltaNet-1.3B demonstrate that REFINE consistently outperforms supervised fine-tuning with NTP across needle-in-a-haystack retrieval, long-context question answering, and diverse tasks in LongBench. REFINE provides an effective and versatile framework for improving long-context modeling in fast weight architectures.
format Preprint
id arxiv_https___arxiv_org_abs_2602_16704
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Reinforced Fast Weights with Next-Sequence Prediction
Hwang, Hee Seung
Wu, Xindi
Chun, Sanghyuk
Russakovsky, Olga
Computation and Language
Fast weight architectures offer a promising alternative to attention-based transformers for long-context modeling by maintaining constant memory overhead regardless of context length. However, their potential is limited by the next-token prediction (NTP) training paradigm. NTP optimizes single-token predictions and ignores semantic coherence across multiple tokens following a prefix. Consequently, fast weight models, which dynamically update their parameters to store contextual information, learn suboptimal representations that fail to capture long-range dependencies. We introduce REFINE (Reinforced Fast weIghts with Next sEquence prediction), a reinforcement learning framework that trains fast weight models under the next-sequence prediction (NSP) objective. REFINE selects informative token positions based on prediction entropy, generates multi-token rollouts, assigns self-supervised sequence-level rewards, and optimizes the model with group relative policy optimization (GRPO). REFINE is applicable throughout the training lifecycle of pre-trained language models: mid-training, post-training, and test-time training. Our experiments on LaCT-760M and DeltaNet-1.3B demonstrate that REFINE consistently outperforms supervised fine-tuning with NTP across needle-in-a-haystack retrieval, long-context question answering, and diverse tasks in LongBench. REFINE provides an effective and versatile framework for improving long-context modeling in fast weight architectures.
title Reinforced Fast Weights with Next-Sequence Prediction
topic Computation and Language
url https://arxiv.org/abs/2602.16704