Step More: Going Beyond Single Backpropagation in Meta Learning Based Model Editing

2025-08-05

AI Summary

To address the poor editing performance of meta-learning-based model editing (MLBME) under low-data regimes and its training inefficiency—primarily constrained by the computational overhead of KL-divergence gradient computation—this paper proposes Multi-Step Backpropagation for Meta-Learning Editing (MBPS). MBPS enhances knowledge adaptation through multi-step inner-loop parameter updates and introduces a lightweight weight-norm regularization to substantially alleviate the gradient-computation burden imposed by KL divergence. It is the first work to systematically integrate multi-step backpropagation into the meta-learning model editing framework, achieving a favorable trade-off between editing accuracy and training efficiency. Experiments on two benchmark datasets and two large language models demonstrate that MBPS consistently outperforms existing meta-learning editing methods in few-shot editing tasks. Moreover, its modular design enables plug-and-play integration, effectively boosting the performance of other editing approaches—thereby validating MBPS’s effectiveness, efficiency, and generalizability.

Technology Category

• Machine Learning: Learning on the Edge & Model Compression
• Multiagent Systems: Multiagent Learning
• Search and Optimization: Metareasoning and Metaheuristics

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Large Language Models (LLMs) underpin many AI applications, but their static nature makes updating knowledge costly. Model editing offers an efficient alternative by injecting new information through targeted parameter modifications. In particular, meta-learning-based model editing (MLBME) methods have demonstrated notable advantages in both editing effectiveness and efficiency. Despite this, we find that MLBME exhibits suboptimal performance in low-data scenarios, and its training efficiency is bottlenecked by the computation of KL divergence. To address these, we propose $ extbf{S}$tep $ extbf{M}$ore $ extbf{Edit}$ ($ extbf{SMEdit}$), a novel MLBME method that adopts $ extbf{M}$ultiple $ extbf{B}$ackpro$ extbf{P}$agation $ extbf{S}$teps ($ extbf{MBPS}$) to improve editing performance under limited supervision and a norm regularization on weight updates to improve training efficiency. Experimental results on two datasets and two LLMs demonstrate that SMEdit outperforms prior MLBME baselines and the MBPS strategy can be seamlessly integrated into existing methods to further boost their performance. Our code will be released soon.

Problem

• Improving low-data performance in meta-learning model editing
• Reducing KL divergence computation bottleneck in training
• Enhancing editing efficiency with norm regularization on updates

Innovation

• Multiple Backpropagation Steps for better editing
• Norm regularization to enhance training efficiency
• Seamless integration into existing MLBME methods