Fast Weight Attention for Continual Learning

Hugging Face Daily Papers Papers

Summary

This paper analyzes recurrent fast-weight memories and selective state-space models as online learning rules, deriving normalized update families that improve length extrapolation and remain competitive in language modeling.

Recurrent fast-weight memories and selective state-space models compress an expanding context into a fixed-size recurrent state, making the state transition an online learning rule. We study this rule under read-after-write autoregressive semantics. For the prefix-prediction objective considered here, the local fast-memory example revealed at step t is the prefix-aligned pair (x_t,y_t)=(ϕ(k_{t-1}),v_t). The common same-step association (ϕ(k_t),v_t) remains causal, but optimizes a different internal objective. We derive normalized first-order updates for squared-error regression and negative inner-product objectives. The regression family comprises Falcon-1 (a scalar NLMS update), Falcon-2 (its per-column extension), and Falcon-3 (a sliding-window mini-batch update); Falcon-1A/Falcon-2A/Falcon-3A are the corresponding inner-product variants. We provide recurrent, masked-parallel, and chunk-parallel forms, together with numerically stable positive-decay renormalization. Representative variants remain competitive in language modeling and improve length extrapolation on variable-digit addition. This framework separates temporal alignment, plasticity, forgetting, and bounded rehearsal in recurrent sequence models.
Original Article
View Cached Full Text

Cached at: 09/01/26, 12:08 PM

Paper page - Fast Weight Attention for Continual Learning

Source: https://huggingface.co/papers/2608.27763 Authors:

,

,

,

,

,

,

,

,

,

Abstract

Recurrent fast-weight memories and selective state-space models are analyzed as online learning rules under autoregressive semantics, yielding normalized update families with stable renormalization that improve length extrapolation.

Recurrentfast-weight memoriesandselective state-space modelscompress an expanding context into a fixed-sizerecurrent state, making the state transition anonline learning rule. We study this rule underread-after-write autoregressive semantics. For theprefix-predictionobjective considered here, the local fast-memory example revealed at step t is the prefix-aligned pair (x_t,y_t)=(ϕ(k_{t-1}),v_t). The common same-step association (ϕ(k_t),v_t) remains causal, but optimizes a different internal objective. We derivenormalized first-order updatesforsquared-error regressionandnegative inner-productobjectives. The regression family comprisesFalcon-1(a scalar NLMS update),Falcon-2(its per-column extension), andFalcon-3(a sliding-window mini-batch update);Falcon-1A/Falcon-2A/Falcon-3A are the corresponding inner-product variants. We provide recurrent,masked-parallel, andchunk-parallelforms, together with numerically stablepositive-decay renormalization. Representative variants remain competitive in language modeling and improvelength extrapolationon variable-digit addition. This framework separates temporal alignment, plasticity, forgetting, and bounded rehearsal in recurrent sequence models.

View arXiv pageView PDFProject pageGitHub15Add to collection

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.27763 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.27763 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.27763 in a Space README.md to link it from this page.

Collections including this paper6

Browse 6 collections that include this paper

Similar Articles

Can a Language Model Learn Facts Continually in Its Weights?

arXiv cs.CL

This paper investigates whether language models can learn new facts in their weights through continual learning. Using invented facts and sequential writes into Qwen3 models, it finds that training data breadth determines knowledge type and retention: bare-statement facts are quickly forgotten (1% accuracy after 20 writes), while facts learned from diverse restatements retain 46% accuracy. Forgotten facts are not erased but become behaviorally inaccessible due to later writes redirecting questions, and context remains the reliable channel for fact composition and survival.

Δ-Mem: Efficient Online Memory for Large Language Models

Hacker News Top

Proposes delta-Mem, a lightweight online memory mechanism that uses a compact state matrix updated by delta-rule learning to improve long-context performance of frozen LLMs without full fine-tuning or context extension.

Dynamic Linear Attention

arXiv cs.CL

This paper proposes DLA, a dynamic memory modeling framework for multi-state linear attention that adaptively merges states based on token information variation and maintains a fixed-size state cache, enabling better long-context representation without the quadratic complexity of standard attention.

Context Memorization for Efficient Long Context Generation

Hugging Face Daily Papers

Proposes attention-state memory, a training-free approach that stores precomputed attention states in lightweight memory to improve accuracy and reduce latency for long prefix inference, outperforming traditional methods on benchmarks.