Challenges to update LLM knowledge
LLM shows excellent performance for a variety of tasks by performing extensive pre-training on a wide range of datasets. However, these models often produce outdated or inaccurate information and can reflect bias during deployment, so their knowledge needs to be continuously updated. Traditional fine-tuning methods are expensive and susceptible to catastrophic forgetting. This inspires lifelong model editing, which effectively updates model knowledge. To generate the correct predictions, each edit requires reliability, generalization, and localization. Methods such as nonparametrics enable precise local editing but have poor generalizations, while parameter methods provide better generalizations but suffer from catastrophic forgetting.
Limitations of previous model editing techniques
Earlier works explore sparse neural activation in continuous learning, and methods such as packnet and supermasks in-superposition allocate subsets of disjoint parameters for each task. Gradient-based approaches such as GPM and SPARCL improve efficiency through orthogonal updates, but are limited to continuous learning environments. Parameter methods, such as Roman, magnetic fields and wise methods, modify weights by locateing to edit policies or auxiliary modules, but suffer losses due to forgetting the extended edit sequence. Nonparametric methods such as Grace and Loka store knowledge retain original weights externally, enabling precise local editing. However, these methods rely on exact input matching, limiting their ability to generalize.
Introduction to Memoirs: A Structured Model Editing Method
Researchers from EPFL, Lausanne, Switzerland, proposed a memoir (model editing, minimal coverage and informed retention) that strikes the best balance between reliability, generalization and large-scale editing. It introduces a memory module consisting of a fully connected layer in a single transformer block where all edits occur. Memoirs resolve catastrophic forgetting by assigning different subsets of parameters to each edit and retrieving them during inference so that only relevant knowledge can be activated to activate relevant knowledge in a specific prompt. Furthermore, the method utilizes structured sparsity with sample dependency masks during editing, only timely specific subsets of parameters are activated. It distributes new knowledge on the parameter space, reducing coverage and minimizing catastrophic forgetting.
Evaluation and experimental results
Memoirs are run through residual storage frameworks during inference, where edited output integrates the original layer output with residual memory output. It is an evaluation for baselines, such as grace periods for external knowledge storage, latency for inference time routing, causal tracking methods such as Roman, MEMIT and ALPHAEDIT, and memory-based methods such as Wise. Direct fine-tuning can be used as an additional baseline comparison. The experiments were conducted on four autoregressive language models: Llama-3-8b-Instruct, Mistral-7b, Llama-2-7b and GPT-J-6B, which can be fully evaluated on different models and scales to show the effectiveness and universality of Momoir.
On the dataset asked by ZSRE, the average metric of memoirs on Llama-3 was 0.95 and reached 0.95 in 1000 edits, outperforming all previous methods with a profit margin of 0.16. Mistral can see similar results, again this approach reaches the highest average score, highlighting its robustness and effectiveness on various LLMs. In addition, memoirs maintain optimal balance performance by using hallucination correction from SelfCheckGPT dataset to increase editing volume. Memoirs scored in the most challenging situations of 600 edits, while achieving the second best performance method on Llama-3 and Mistral, respectively, reaching 57% and 77% over WISE.
Conclusion and future direction
In summary, memoirs are an extensible framework for lifelong model editing that effectively balances reliability, generalization and locality using innovative sparse techniques. This method searches for related updates by comparing sparse activation modes, so that edits can be summarized as rewriting queries while maintaining model behavior on irrelevant prompts. However, there are certain limitations, such as modifying only a single linear layer, which may limit the handling of long-match edits or knowledge that requires broader model changes. Future directions include extending the method to multiple layers, layered editing strategies, and applying multimodal or encoder models beyond the current decoder-only transformer focus.
Check Paper. All credits for this study are to the researchers on the project. Also, please stay tuned for us twitter And don’t forget to join us 100K+ ml reddit And subscribe Our newsletter.
Sajjad Ansari is a final year undergraduate student from IIT Kharagpur. As a technology enthusiast, he delves into the practical application of AI, focusing on understanding AI technology and its real-world impact. He aims to express complex AI concepts in a clear and easy way.
