Journey to the Center of the Knowledge Neurons: Discoveries of Language-Independent Knowledge Neurons and Degenerate Knowledge Neurons

Problem setting

• Knowledge localization seeks the parameter locations of specific factual knowledge, whose relevant parameters are termed knowledge neurons.The paper frames this task against uncertainty about how PLMs store factual knowledge.Source: §1 · p. 1

Knowledge localization

• AMIG uses architecture adaptation and multilingual threshold adjustment to localize knowledge neurons more precisely across PLM architectures and languages.It addresses traditional methods’ lack of universality across auto-encoding and auto-regressive models and their limited multilingual exploration.Source: §1 · pp. 1–2

Language-Independent Knowledge Neurons

• Language-independent knowledge neurons store factual knowledge across languages and enable a single edit to affect corresponding knowledge in multiple languages.The neurons are obtained by intersecting knowledge neurons derived from different languages, and experiments report superior cross-lingual editing performance over existing models.Source: §1 · p. 2

Degenerate Knowledge Neurons

• Degenerate knowledge neurons comprise partially overlapping neurons that store the same fact, so suppressing some preserves expression while suppressing all prevents it.They cannot simply be deleted because one neuron may participate in multiple facts.Source: §1 · p. 2
• Degenerate knowledge neurons help PLMs detect wrong facts, supporting more stable mastery of factual knowledge.For a fact such as Tanzania’s capital, the model must activate at least one corresponding degenerate neuron to predict the correct answer.Source: §1 · p. 2

2.1 Knowledge Neuron Localization

• Knowledge neuron localization evaluates a PLM’s probability for the correct answer in a fill-in-the-blank cloze task.The method computes neuron-level attribution scores with integrated gradients and identifies neurons above a threshold.Source: §2.1 · pp. 2–3
• AMIG adapts integrated-gradients baselines to different PLM architectures and aggregates multilingual attribution scores to localize knowledge neurons.It uses architecture-specific baseline vectors and selects neurons whose normalized attribution exceeds threshold τ.Source: §2.1 · p. 3

2.2 Language-Independent Knowledge Neuron Detection

• Language-independent knowledge neurons are obtained by adjusting attribution thresholds per language and intersecting the resulting language-specific neuron sets.The intersection represents factual knowledge encoded consistently across all considered languages.Source: §2.2 · pp. 3–4

2.3 Degenerate Knowledge Neuron Detection

• Degenerate knowledge neurons are distinct neuron subsets whose individual suppression has little effect, while suppressing the complete subset substantially lowers prediction probability.The detection algorithm first filters individually insensitive neurons, then tests pairs using low and high probability-difference thresholds.Source: §2.3 · p. 4
• The general search for degenerate subsets has O(2^n) complexity, so the method assumes each degenerate subset contains two neurons, reducing complexity to O(n^2).This assumption makes exhaustive subset evaluation more tractable but constrains the detected subset structure.Source: §2.3 · p. 4

3.2 Localization of Knowledge Neurons

• 84.34% and 44.49% higher success rates for AMIG than B-KN on Chinese m-BERT and m-GPT, respectively, with gains in all settings.The evaluation uses average success rate SR, where higher values indicate more precise localization.Source: §3.2 · p. 5, Table 1 · p. 5
• m-BERT knowledge neurons concentrate mainly in final layers, whereas m-GPT knowledge neurons span early, middle, and final layers.Figure 3 plots layer position against the percentage of knowledge neurons; English and Chinese distributions are relatively similar but not identical.Source: §3.2 · p. 5, Figure 3 · p. 5

3.3 Language-Independence Neurons and Cross-Lingual Knowledge Editing

• Language-independent knowledge neurons concentrate in the final one or two layers of both m-BERT and m-GPT.The authors interpret language-independent facts as high-level features encoded late in the models.Source: §3.3 · p. 5, Figure 3 · p. 5
• 213.05% and 277.36% higher Chinese-dataset success rates result from editing language-independent neurons rather than single-language neurons in m-BERT and m-GPT.The experiments compare editing one language, sequentially editing two languages, and editing language-independent neurons.Source: §3.3 · p. 6, Table 2 · p. 6
• Sequentially editing two languages produces 42.97% and 58.80% lower success rates than language-independent editing in m-BERT and m-GPT.The authors suggest that multiple edits may introduce confusion while also requiring more computational resources.Source: §3.3 · p. 6

3.4 Degenerate Knowledge Neurons and Fact-Checking Experiment

• Degenerate knowledge neurons in m-BERT and m-GPT show distributions similar to knowledge neurons, linking degeneracy to factual knowledge and the models’ fact mastery.The multilingual distributions are reported in Figure 4.Source: §3.4 · p. 6, Figure 4 · p. 6
• 167150% higher F1 score is reported for with-DKN than wo-DKN in m-GPT on Chinese datasets, with improvements across various settings.The fact-checking experiment uses Precision, Recall, and F1-score and evaluates correct versus randomly constructed incorrect facts.Source: §3.4 · p. 7, Table 3 · p. 7
• Degenerate-neuron fact-checking also works in monolingual BERT and GPT-2, indicating that its existence is unrelated to multilingual support.Figure 5 presents the monolingual DKN distributions, while the authors report fact-checking utility in these models.Source: §3.4 · pp. 6–7, Figure 5 · p. 6

• Knowledge localization methods primarily use gradient-based attribution or causal-inspired tracing to identify neurons associated with factual knowledge.Source: §4 · p. 7
• Existing methods remain limited by incomplete architectural universality and multilingual coverage, especially beyond English datasets.Source: §4 · p. 7
• Integrated Gradients and related attribution methods improve neuron-contribution estimation through interpolation, sequential evaluation, or Shapley-based efficiency enhancements.Source: §4 · p. 7

• The study introduces architecture-adapted multilingual integrated gradients and discovers language-independent and degenerate knowledge neurons in multilingual PLMs.Language-independent neurons support cross-lingual knowledge editing, while degenerate neurons can detect incorrect facts.Source: §5 · p. 7