Balancing Generality and Locality in Model Editing

22 concepts20 papers

When a large neural network makes a factual error or falls out of date, retraining it from scratch is prohibitively expensive. This has driven the rise of Knowledge-Based Model Editing (KME)—techniques designed to surgically alter specific facts or behaviors within a pre-trained model without degrading its overall capabilities, as surveyed by Song Wang et al. in their comprehensive overview, Knowledge Editing for Large Language Models: A Survey.

However, performing surgery on a network's weights reveals a fundamental friction. If you modify a network so that it successfully updates a fact, how do you ensure it understands that update across different phrasings without accidentally corrupting completely unrelated knowledge?

This tension is known as the Generality-Locality Trade-off, or what Peng Wang et al. in WISE: Rethinking the Knowledge Memory for Lifelong Model Editing of Large Language Models and Ke Wang in MEMOIR: Lifelong Model Editing with Minimal Overwrite and Informed Retention for LLMs frame as an Impossible Triangle or Model Editing Trilemma. The three vertices of this triangle are:

  1. Reliability: The model correctly outputs the new target information when given the exact editing prompt.
  2. Generality: The model applies the update to semantically equivalent queries, paraphrases, or related contexts.
  3. Locality: The model leaves unrelated knowledge and general reasoning capabilities entirely untouched.
                  [ Reliability ]
                  (Exact prompt works)
                         /\
                        /  \
                       /    \
                      /      \
                     /________\
     [ Generality ]              [ Locality ]
(Applies to paraphrases)    (Unrelated facts safe)

Understanding why this trade-off occurs—and why it worsens when scaling up edits—requires looking closely at how knowledge is distributed inside deep networks.

Why the Trade-off Exists: Semantic Bleed vs. Lexical Bias

The core difficulty stems from the fact that neural networks do not store information in isolated, discrete database slots. Instead, knowledge is distributed across highly entangled weight spaces.

The Risk of Over-Generalization

When an editing method forces a weight update to ensure high generality, it risks pushing the model's internal representations too far. As Dongliang Guo et al. observe in BalancEdit: Dynamically Balancing the Generality-Locality Trade-off in Multi-modal Model Editing, aggressive updates cause the edit to leak into neighboring semantic spaces. For instance, changing the model's knowledge about a specific concept might inadvertently alter its behavior for an entire category of distinct, unrelated entities. This problem extends beyond text; Hadas Orgad et al. demonstrate in Editing Implicit Assumptions in Text-to-Image Diffusion Models that editing a single concept in a diffusion model creates a strict boundary conflict: stronger edits improve consistency across related prompts but rapidly degrade the accuracy of distinct semantic groups.

The Trap of Lexical Bias

Conversely, if an editing method prioritizes strict locality by constraining parameter shifts to a tiny neighborhood, it often falls victim to lexical bias. As explained by Hammad Rizwan et al. in Resolving Lexical Bias in Model Editing, weight-preserving editors often mistake surface-level word overlap for semantic meaning. If the edit is tied too closely to the literal words of the training prompt, the model will fail to recognize the updated fact in a rephrased query. Worse, it might erroneously trigger the edit for an entirely different entity simply because the query shares similar-looking words.

The Geometry of Knowledge Storage

To resolve this trade-off, researchers have tried to locate exactly where facts live. In their seminal work Locating and Editing Factual Associations in GPT, Kevin Meng et al. treat feed-forward networks (MLPs) as linear associative memories, showing that factual associations can often be localized to specific middle-layer neurons at the last token of a subject.

However, modifying these localized sites introduces a deeper structural problem. Yuheng Chen et al. challenge the idea of purely static "knowledge neurons" in Knowledge Localization: Mission Not Accomplished? Enter Query Localization!. They argue that knowledge retrieval is dynamic: the attention mechanism routes information to different neurons depending on how a query is phrased (Query Localization).

Because a single fact is accessed via multiple dynamic pathways, attempting to alter all neurons associated with that fact across every possible context causes massive collateral damage. The authors show that forcing these widespread weight changes significantly degrades the model's general reasoning capabilities and linguistic coherence.

The Scaling Collapse in Lifelong Editing

The generality-locality friction becomes acute when moving from a single edit to sequential, "lifelong" editing. When injecting thousands of facts over time, the parameter space quickly becomes crowded.

As highlighted by Chenmien Tan et al. in Massive Editing for Large Language Models via Meta Learning and Wentao Wan et al. in Massive Editing for Large Language Models Based on Dynamic Weight Generation, scaling up the number of simultaneous or sequential edits forces a severe drop in model stability. Data augmentation techniques meant to protect locality often end up choking the model's capacity for semantic generalization. If you increase the learning rate to force the model to internalize an update, it triggers catastrophic forgetting, destroying general task proficiency—a phenomenon documented in code models by Xi Ye et al. in CodeUpdateArena: Benchmarking Knowledge Editing on API Updates.

This problem is not unique to language models. In computer vision, Shengzhuang Chen et al. show in Learning Where to Edit Vision Transformers that surgical updates to fix specific predictive errors in Vision Transformers (ViTs) frequently degrade performance on semantically distant data points.

How Modern Methods Navigate the Trade-off

To break out of this trilemma, recent architectures avoid crude weight updates in favor of more structured, isolated memory mechanisms.

  • Surgical Neuron Addition: Instead of altering existing weights and risking semantic bleed, Zeyu Huang et al. introduce Transformer-Patcher: One Mistake Worth One Neuron, which appends a single new neuron to the final feed-forward layer for each error. This isolates the edit structurally, preserving locality while allowing the model to generalize to paraphrases.
  • Dynamic Influence Radii: To prevent over-generalization, frameworks like the one proposed by Dongliang Guo et al. in BalancEdit compute a dynamic activation boundary using positive and negative reference samples, ensuring that an edit's influence propagates to valid paraphrases but cuts off before bleeding into unrelated concepts.
  • Routing and Mixture-of-Experts (MoE): Rather than modifying the base model, Tai Duc Nguyen et al. develop UniAdapt: A Universal Adapter for Knowledge Calibration, which uses a vector-assisted router to direct queries to specific specialized adapter modules, maintaining high accuracy across thousands of sequential updates.
  • Disentangled Memory Spaces: In Towards Unified Multimodal Editing with Enhanced Knowledge Collaboration, Kaihang Pan et al. point out that traditional weight updates fail because parameter shifts and external contexts operate at clashing semantic levels. By explicitly separating updates into distinct semantic and truthfulness spaces, models can better align generalization with strict locality boundaries.
  • Careful Data Augmentation: Surprisingly, specialized architectures are not always mandatory. Govind Krishnan Gangadhar et al. demonstrate in Model Editing by Standard Fine-Tuning that standard fine-tuning can preserve locality and achieve high generality if the user simply masks the prompt tokens and mixes a small handful of random, unedited facts into the training batch to anchor the surrounding weight space.

Ultimately, solving the generality-locality trade-off means moving away from treating neural networks as uniform blocks of weights. Whether through dynamic routing, structural isolation, or contrastive boundary protection, successful model editing requires respecting the delicate geometry of how networks store what they know.

Key Foundational Work

Transformer-Patcher: One Mistake Worth One Neuron (2023) — contributed 1 concept elements to this direction.
Knowledge Editing for Large Language Models: A Survey (2023) — contributed 1 concept elements to this direction.
Editing Implicit Assumptions in Text-to-Image Diffusion Models (2023) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_generality_locality_trade_off_af8529e3,
  title = {generality-locality trade-off},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generality-locality-trade-off/rdr_af8529e322a5495d851c52b4abfb919e},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_generality_locality_trade_off_af8529e3_v1,
  title = {generality-locality trade-off},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/generality-locality-trade-off/rdr_af8529e322a5495d851c52b4abfb919e/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}

See something wrong? Report a mistake