PaperlayerView exact revision on arXiv

Source-linked AI summary

AlphaEdit: Null-Space Constrained Knowledge Editing for Language Models

Junfeng Fang, Houcheng Jiang, Kun Wang, Yunshan Ma, Shi Jie, Xiang Wang, Xiangnan He, Tat-seng Chua

arXiv:2410.02355v4cs.CLcs.AI

TL;DR

Locate-then-edit methods can disrupt preserved knowledge, especially during sequential editing. AlphaEdit projects update perturbations into the preserved knowledge’s null space, and experiments report a 36.7% average improvement across multiple LLMs and editing methods with a single additional code line.

  • Problem

    Locate-then-edit perturbations can disrupt preserved knowledge, particularly in sequential editing scenarios, creating a need to balance knowledge updates with preservation.

  • Method

    AlphaEdit removes the preservation-error term from the editing objective and projects parameter perturbations onto the null space of preserved knowledge before applying them.

  • Results

    36.7% average improvement in editing capabilities is reported across multiple base LLMs and existing model editing methods.

  • Takeaways & Limitations

    AlphaEdit provides a lightweight, plug-and-play modification that improves sequential knowledge editing while preserving general capabilities after extensive editing.

  • Takeaways & Limitations

    Applicability to multi-modal LLMs and large reasoning models remains unexplored.

Abstract

from arXiv · show

Large language models (LLMs) often exhibit hallucinations due to incorrect or outdated knowledge. Hence, model editing methods have emerged to enable targeted knowledge updates. To achieve this, a prevailing paradigm is the locating-then-editing approach, which first locates influential parameters and then edits them by introducing a perturbation. While effective, current studies have demonstrated that this perturbation inevitably disrupt the originally preserved knowledge within LLMs, especially in sequential editing scenarios. To address this, we introduce AlphaEdit, a novel solution that projects perturbation onto the null space of the preserved knowledge before applying it to the parameters. We theoretically prove that this projection ensures the output of post-edited LLMs remains unchanged when queried about the preserved knowledge, thereby mitigating the issue of disruption. Extensive experiments on various LLMs, including LLaMA3, GPT2-XL, and GPT-J, show that AlphaEdit boosts the performance of most locating-then-editing methods by an average of 36.7% with a single line of additional code for projection solely. Our code is available at: https://github.com/jianghoucheng/AlphaEdit.

1 INTRODUCTION

AlphaEdit addresses the trade-off between updating target knowledge and preserving existing knowledge in locate-then-edit model editing. It projects parameter perturbations into the preserved knowledge’s null space to reduce disruption, with experiments showing substantial gains across models and editing methods.

  • LLM model editing updates incorrect or outdated knowledge more efficiently than full fine-tuning by modifying targeted parameters.Full fine-tuning is described as often prohibitively time-consuming.
  • Current locate-then-edit methods locate influential parameters and optimize a perturbation using update error e1 alongside preservation error e0.The balance between these errors is central to the conventional objective.
  • Sequential editing can shift hidden-representation distributions, degrade preservation and fluency, and eventually cause model forgetting and model collapse.The paper attributes these effects to accumulated overfitting toward updated knowledge.
  • AlphaEdit removes e0 from the objective and projects the resulting perturbation onto the null space of preserved knowledge before applying it to model parameters.The projection is intended to keep hidden-representation distributions invariant while reducing e1.
  • 36.7% average performance improvement over best-performing baselines is reported across multiple representative LLMs by adding one line of code to MEMIT.The paper also describes the projection as a plug-and-play enhancement for most existing editing methods.

2 PRELIMINARY

The preliminary section models LLM feed-forward layers as key-value associative memory and formulates model editing as modifying output weights to map keys for knowledge subjects and relations to desired values.

  • 2.1 AUTOREGRESSIVE LANGUAGE MODEL: An autoregressive LLM predicts the next token from preceding tokens, with hidden state h_l combining attention and feed-forward outputs through layer normalization.The formulation represents attention and FFN modules in parallel.
  • 2.1 AUTOREGRESSIVE LANGUAGE MODEL: The FFN output weight matrix W^l_out is interpreted as linear associative memory that stores information by associating input keys with output values.Keys encode subject-relation pairs, while values encode objects.
  • 2.1 AUTOREGRESSIVE LANGUAGE MODEL: Most model editing methods modify FFN layers because their output weights are viewed as key-value storage for retrieved knowledge.The paper subsequently denotes W^l_out simply as W.
  • 2.2 MODEL EDITING IN LLMS: Each edit adds a perturbation Δ to W so stacked keys K1 map to stacked desired values V1 for u pieces of updated knowledge.W has dimensions d1×d0, while K1 and V1 collect the corresponding key and value vectors.
  • 2.2 MODEL EDITING IN LLMS: Preserved knowledge is represented by matrices K0 and V0, and existing methods include an error term involving them to limit disruption.The formulation assumes W K0 = V0 for preserved knowledge and estimates K0 from abundant text input.

3 METHOD

AlphaEdit projects editing perturbations into the null space of preserved knowledge, protecting existing key-value associations while focusing optimization on updated knowledge. The resulting method requires only a minor modification to standard editing solutions and is designed for efficient integration into existing algorithms.

  • 3.2 NULL SPACE PROJECTING: The method projects perturbation ∆ onto the null space of K0, where K0 encodes preserved knowledge, before adding it to parameters W.The projection is implemented through the null space of K0K0^T, which has the same null space as K0.
  • 3.1 NULL SPACE: (W + ∆′)K0 = W K0 = V0, showing that projected perturbations preserve the key-value associations of preserved knowledge.This property ensures that the storage of preserved knowledge remains intact after editing.
  • 3.2 NULL SPACE PROJECTING: The projection matrix P is constructed from eigenvectors associated with zero eigenvalues after SVD, yielding ∆P · K0 = 0.In practice, eigenvectors corresponding to eigenvalues above 10^-2 are removed because eigenvalues are rarely exactly zero.
  • 3.3 NULL-SPACE CONSTRAINED MODEL EDITING: AlphaEdit replaces ∆ with ∆P, removes the preserved-knowledge error term, and adds ||∆P||2 regularization for stable convergence.For sequential editing, an additional term protects associations from previous edits; K1 and V1 denote the current edit’s key and value matrices.
  • 3.3 NULL-SPACE CONSTRAINED MODEL EDITING: The resulting ∆AlphaEdit stores current updates while keeping preserved knowledge and previously updated knowledge unaffected.The final perturbation ∆AlphaEdit = ∆P is added to the model parameters W.
  • 3.3 NULL-SPACE CONSTRAINED MODEL EDITING: AlphaEdit requires only a minor modification to standard solutions, can reuse a projection matrix independent of updated knowledge, and adds negligible time consumption compared with baselines.The paper describes this integration as a single-line code modification for existing editing methods.

4 EXPERIMENT

The experiments evaluate AlphaEdit across sequential editing, general capability retention, hidden-representation stability, and compatibility with existing editing methods. Results show strong preservation and broad improvements across models and tasks.

  • Experimental Setup: Experiments cover GPT2-XL, GPT-J, and LLaMA3 using Counterfact and ZsRE with Efficacy, Generalization, Specificity, Fluency, and Consistency metrics.Additional evaluations include six GLUE-based general capability tasks and baseline-integration studies.
  • Sequential Editing: 12.54% and 16.78% average improvements in Efficacy and Generalization, respectively, are achieved over the best baseline.On LLaMA3, the corresponding gains reach 32.85% and 30.60%.
  • Sequential Editing: 18.33% improvement in Fluency over the strongest baseline is achieved on GPT2-XL, alongside improved text-generation coherence.The result indicates preservation of fluent generation in addition to editing performance.
  • General Capability Tests: After editing 3,000 samples, AlphaEdit maintains the original model performance across all general-capability metrics.Baseline methods show rapidly declining metrics after 2,000 edited samples.
  • Hidden Representations Analysis: AlphaEdit keeps hidden representations consistent with the original distribution across all three base models after editing.Baseline editing produces a significant distribution shift, whereas AlphaEdit exhibits minimal shift.

5 RELATED WORK

Related work divides model editing into parameter-modifying approaches, parameter-preserving approaches, and benchmarks for evaluating knowledge edits. These lines differ in how they store updates and assess editing behavior.

  • Parameter-modifying Model Editing: Parameter-modifying methods use meta-learning or locate-then-edit strategies to alter model parameters for knowledge updates.Examples include KE, MEND, InstructEdit, ROME, and MEMIT.
  • Parameter-preserving Model Editing: Parameter-preserving methods store updated knowledge in additional modules or incorporate it into prompts rather than directly modifying the base parameters.Examples include SERAC, T-Patcher, GRACE, MELO, MemPrompt, and IKE.
  • Evaluating Knowledge Editing: Knowledge-editing benchmarks evaluate insertion, modification, erasure, long-form, multilingual, and multi-hop knowledge settings.Examples include KnowEdit, LEME, CKnowEdit, and MQuAKE.

6 LIMITATIONS & FUTURE DISCUSSION

AlphaEdit’s applicability to multimodal LLMs and large reasoning models remains unexplored. Future work therefore includes extending evaluation to broader classes of base models.

  • Limitations: AlphaEdit has not yet been evaluated on multimodal LLMs or large reasoning models.The authors identify broader base-model coverage as a future research direction.

7 CONCLUSION

AlphaEdit addresses the trade-off between updating and preserving knowledge by projecting parameter perturbations onto the null space of key matrices. Across several base LLMs, it improves existing editing methods with a single-line modification.

  • Conclusion: AlphaEdit projects parameter perturbations onto the null space of key matrices so the model can focus on knowledge updates while minimizing disruption to preserved knowledge.The method is introduced as a single-line modification to current editing approaches.
  • Conclusion: 36.7% average improvement in editing capabilities is demonstrated across LLaMA3, GPT-2 XL, and GPT-J.The conclusion reports enhancement of existing model editing methods across multiple base LLMs.

ETHICS STATEMENT

AlphaEdit’s performance benefits are accompanied by risks from directly modifying stored knowledge. The authors call for strict validation, oversight, and responsible use.

  • Directly modifying stored knowledge can introduce false or harmful information.

REPRODUCIBILITY

The reproducibility materials describe the datasets and metrics used to evaluate model editing. They cover efficacy, generalization, specificity, fluency, and consistency across ZsRE and Counterfact.

  • Datasets: Counterfact contrasts counterfactual and factual statements and evaluates efficacy, generalization, and specificity.
  • ZsRE metrics: ZsRE efficacy measures average top-1 accuracy on edit samples.
  • ZsRE metrics: ZsRE generalization measures average top-1 accuracy on rephrased versions of edit prompts.
  • ZsRE metrics: Specificity evaluates whether predictions for unrelated samples remain unchanged.
  • Counterfact metrics: Counterfact additionally measures fluency through generation entropy and consistency through cosine similarity between generated and reference-text TF-IDF vectors.

A.3 IMPLEMENTATION DETAILS

The implementation follows MEMIT configurations for GPT-2 XL and GPT-J, with specified target layers and optimization settings. Experiments use a single A40 GPU and reproduce several baseline methods.

  • Model configurations: GPT-2 XL edits layers [13, 14, 15, 16, 17] with λ = 20,000, 20 optimization steps, and learning rate 0.5.
  • Model configurations: GPT-J edits layers [3, 4, 5, 6, 7, 8] with λ = 15,000 and 25 optimization steps.
  • Hardware and software: All experiments run on a single A40 GPU using Hugging-Face Transformers.
  • Baselines: The study reproduces MEND, InstructEdit, ROME, and MEMIT, while implementing PRUNE from its paper because its code is unavailable.

B IMPLEMENTATION DETAILS OF CURRENT MODEL EDITING & RELATED PROOFS

Locate-then-edit methods identify influential layers, optimize updated knowledge, and modify selected parameters; AlphaEdit instead constrains updates to the null space of preserved knowledge. Experiments report stronger sequential-editing behavior, reduced representation shifts, and broad gains in efficacy and generalization.

  • Current editing pipeline: Causal tracing locates influential FFN layers by injecting noise, restoring hidden states, and measuring output recovery.
  • Current editing pipeline: The editing pipeline computes desired layer outputs and uses gradient descent to increase the probability of the updated object.
  • Sequential editing: Sequential-editing case studies show that baselines can produce incoherent outputs, whereas AlphaEdit performs edits while maintaining coherent generations.
  • Representation shifts: Across all metrics and base LLMs, AlphaEdit-optimized methods exhibit minimal hidden-representation distribution shifts.
  • Performance: Across models and tasks, AlphaEdit achieves the highest efficacy and generalization scores; on GPT-J Counterfact, efficacy reaches 99.75.
  • Additional base LLMs: On Gemma and phi-1.5, AlphaEdit outperforms MEMIT and RECT across key metrics, including Gemma fluency 398.96 and phi-1.5 efficacy 70.79.

C.7 EVALUATION ON EXPANDING BENCHMARK: KNOWEDIT, LEME AND MQUAKE

AlphaEdit is evaluated on expanded benchmarks covering factual editing, multi-hop reasoning, long-form generation, and reduced-data sequential editing. Results indicate strong editing performance, stable efficacy and generalization under dataset reduction, and no additional runtime overhead versus MEMIT.

  • KnowEdit: 96.10% editing success on wiki_recent exceeds RECT’s 82.47%, while AlphaEdit also reaches 95.34% on wikibio.These evaluations cover editing success, portability, locality, and fluency.
  • MQUAKE: AlphaEdit achieves Multi-hop and Multi-hop (CoT) scores of 9.14 and 9.75, respectively, on MQUAKE.The benchmark evaluates factual-update ripple effects through multi-hop reasoning questions.
  • LEME: On LEME, AlphaEdit performs strongly across edit consistency, factual consistency, and internal consistency for long-form editing.LEME emphasizes consistency, factual correctness, and lexical cohesion.
  • Dataset-size robustness: At 10% of the original dataset size, Efficacy and Generalization decline by less than 5%, whereas Specificity drops by 11.76%.The results suggest neighborhood-knowledge storage depends more heavily on sufficient dataset size.
  • Scalability: AlphaEdit incurs no additional runtime overhead compared with MEMIT across LLaMA3, GPT-J, and GPT2-XL as model size or knowledge-base size grows.Runtime was measured for 100 edits across different models and knowledge bases.

D VISUALIZING THE COUNTERFACT AND ZSRE DATASETS THROUGH EXAMPLES

The paper provides examples from the Counterfact and ZSRE datasets to clarify the factual modifications and updates used in model editing. These examples are presented in Figures 12 and 13.

  • Counterfact: Figure 12 presents a sample from the Counterfact dataset.The example illustrates the type of factual modification applied during editing.
  • ZSRE: Figure 13 presents samples from the ZSRE dataset.The examples illustrate factual updates applied to models during editing.
Loading 2410.02355v4…