Beyond Hard Writes and Rigid Preservation: Soft Recursive Least-Squares for Lifelong LLM Editing

2026-01-22

AI Summary

This work addresses the plasticity–stability trade-off faced by large language models during continual editing. The authors propose a soft-constrained online quadratic optimization framework that effectively mitigates cumulative interference and forgetting of early edits. By integrating recursive least squares (RLS) with the Woodbury identity, they design an efficient recursive update mechanism that anchors pre-trained knowledge while supporting an unbounded sequence of edits at constant computational cost. Dual regularization terms are introduced to control weight deviations, enabling the model to maintain high editing success rates and strong retention of early edits over tens of thousands of consecutive updates. The method significantly outperforms existing approaches across GLUE, reasoning, and code generation benchmarks.

Technology Category

• Machine Learning: Life-Long and Continual Learning
• Natural Language Processing: (Large) Language Models
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Model editing updates a pre-trained LLM with new facts or rules without re-training, while preserving unrelated behavior. In real deployment, edits arrive as long streams, and existing editors often face a plasticity-stability dilemma: locate-then-edit"hard writes"can accumulate interference over time, while null-space-style"hard preservation"preserves only what is explicitly constrained, so past edits can be overwritten and unconstrained behaviors may deviate, degrading general capabilities in the many-edits regime. We propose RLSEdit, a recursive least-squares editor for long sequential editing. RLSEdit formulates editing as an online quadratic optimization with soft constraints, minimizing a cumulative key-value fitting objective with two regularizers that control for both deviation from the pre-trained weights and from a designated anchor mapping. The resulting update admits an efficient online recursion via the Woodbury identity, with per-edit cost independent of history length and scaling only with the current edit size. We further provide deviation bounds and an asymptotic characterization of the adherence-preservation trade-off in the many-edits regime. Experiments on multiple model families demonstrate stable scaling to 10K edits, outperforming strong baselines in both edit success and holistic stability -- crucially retaining early edits, and preserving general capabilities on GLUE and held-out reasoning/code benchmarks.

Problem

• lifelong editing
• plasticity-stability dilemma
• model editing
• sequential edits
• capability preservation

Innovation

• recursive least-squares
• lifelong model editing
• soft constraints
• online optimization
• plasticity-stability trade-off