PaperlayerView exact revision on arXiv

Source-linked AI summary

The Transfer Neurons Hypothesis: An Underlying Mechanism for Language Latent Space Transitions in Multilingual LLMs

Hinata Tezuka, Naoya Inoue

arXiv:2509.17030v1cs.CLcs.AIcs.LG

TL;DR

The dynamics and mechanism of multilingual latent-space transitions remain underexplored. This paper studies representation changes across layers and proposes that MLP transfer neurons mediate movement between language-specific and shared semantic spaces. It reports evidence for these spaces and neurons, and concludes that transfer neurons are essential for reasoning.

  • Problem

    The dynamics of representation transformation and its underlying mechanism remain underexplored, leaving the latent-space transition process without quantitative demonstration.

    Source: §1 · p. 1, §2.1 · p. 2

  • Method

    The paper studies representations across layers and proposes that specific MLP neurons transfer representations between language-specific and shared semantic latent spaces.

    Source: §1 · pp. 1–2

  • Results

    The study reports evidence for language-specific and shared semantic latent spaces, identifies transfer neurons with language and language-family specificity, and finds them essential for reasoning.

    Source: §7 · p. 9

  • Takeaways & Limitations

    The authors suggest selectively fine-tuning transfer neurons after multilingual pretraining may align low-resource language states with the English-centric space where reasoning occurs.

    Source: §7 · p. 9

    The analysis considers MLP-induced linear transformations and does not consider other modules that might induce alternative vector movements.

    Source: Limitations · p. 10

Abstract

from arXiv · show

Recent studies have suggested a processing framework for multilingual inputs in decoder-based LLMs: early layers convert inputs into English-centric and language-agnostic representations; middle layers perform reasoning within an English-centric latent space; and final layers generate outputs by transforming these representations back into language-specific latent spaces. However, the internal dynamics of such transformation and the underlying mechanism remain underexplored. Towards a deeper understanding of this framework, we propose and empirically validate The Transfer Neurons Hypothesis: certain neurons in the MLP module are responsible for transferring representations between language-specific latent spaces and a shared semantic latent space. Furthermore, we show that one function of language-specific neurons, as identified in recent studies, is to facilitate movement between latent spaces. Finally, we show that transfer neurons are critical for reasoning in multilingual LLMs.

1 Introduction

The introduction frames multilingual processing as movement between language-specific and shared semantic representations, then proposes MLP transfer neurons as a mechanism. It reports that representations converge into a shared space and later diverge, and that transfer neurons are necessary for transfer and important for downstream tasks.

1 Introduction

  • The internal dynamics of representation transformation and its underlying mechanism remain underexplored.The introduction identifies this gap as limiting comprehensive understanding of semantic processing in multilingual LLMs.Source: §1 · p. 1
  • The Transfer Neurons Hypothesis proposes that specific MLP neurons mediate transitions between language-specific and shared semantic latent spaces.These neurons are hypothesized to be sensitive to input language; their firing adds corresponding value vectors to the residual stream, facilitating transfer.Source: §1 · pp. 1–2
  • Representations begin language-specific, converge into a shared latent space across languages, and finally diverge again.The introduction presents this as the observed dynamics of multilingual LLM representations.Source: §1 · p. 2
  • The paper reports empirical evidence for transfer neurons and intervention-based evidence that they are necessary for representation transfer.The introduction describes these neurons as mediating transitions between language-specific and shared semantic spaces.Source: §1 · p. 2
  • The authors suggest that language-specific neurons facilitate movement between latent spaces and report that transfer neurons are essential for downstream tasks.They identify reasoning as a downstream task for which transfer neurons are essential.Source: §1 · p. 2

2 Background

Prior work proposes shared or English-centric processing and identifies language-specific regions, but the latent-space transition process and the role of those regions remain insufficiently explained. The section also introduces an MLP key-value view and a PCA figure of same-meaning representations across languages.

2.1 Related Works

  • Cross-lingual concept neurons have been reported (Chen et al., 2024; Zhao et al., 2024a; Zhang et al., 2025), and LangBridge connects multilingual encoders to LLMs and improves low-resource reasoning (Yoon et al., 2024).Prior encoder studies report a degree of language-agnosticism (Libovický et al., 2020).Source: §2.1 · p. 2
  • Prior work describes English-centric or shared-semantic processing frameworks (Zhao et al., 2024b; Zeng et al., 2025), while language-specific regions are reported mainly in initial and final layers (Kojima et al., 2024; Tang et al., 2024; Zhao et al., 2024b; Zeng et al., 2025; Duan et al., 2025).The overall latent-space transition process lacks quantitative demonstration, and the role of language-specific regions remains insufficiently revealed.Source: §2.1 · p. 2

2.2 Neuronic View of MLPs in Transformers

  • The MLP key-value view treats up-projection columns as keys, down-projection rows as values, and the input as the query (Geva et al., 2021).The section describes activation scores as relevance scores between keys and the query, and notes that experiments use a gated MLP form (Shazeer, 2020; Liu et al., 2021).Source: §2.2 · p. 2
  • Figure 2 applies PCA to hidden representations of same-meaning Japanese, Italian, Dutch, English, and Korean sentences from LLaMA3-8B.The caption states that results for all layers and models appear in Appendix F.Source: Figure 2 · p. 3

3 The Dynamics of Hidden States

The paper traces representations from language-specific spaces toward a shared, English-centric space in middle layers, then back toward language-specific spaces. The degree of convergence varies across languages, while middle-layer representations show greater dimensionality and cross-language similarity.

3.1 Spatial Transition Phenomenon

  • Language-specific representations converge toward a shared space in middle layers and diverge again in final layers, but Japanese and Korean converge less fully than Dutch and Italian.Dutch and Italian remain close to English even in initial and final layers; the pattern is consistent across models.Source: §3.1 · p. 3

3.2 Geometry of Language-Specific Latent Spaces

  • In initial and final layers, 95% of representations lie in relatively low-dimensional spaces; middle-layer representations form a higher-dimensional space.The low-dimensional tendency is especially pronounced for Japanese and Korean, and centroid-distance transitions align with the observed spatial transitions.Source: §3.2 · p. 4, Figure 3 · p. 4
  • Hidden states for each language occupy relatively low-dimensional latent spaces, while the study treats English as the shared semantic space based on prior work.The dimensionality analysis uses SVD and cumulative explained variance ratio (CEVR).Source: §3.2 · pp. 3–4, §3 · p. 3

3.3 The Existence of a Shared Semantic Latent Space in Middle Layers

  • English–other-language latent-space similarity peaks in the middle layers, supporting shared semantic processing across models.The study measures mutual k-nearest-neighbor alignment for parallel sentence representations; additional hidden-state and activation-pattern analyses also support this pattern.Source: Figure 4 · p. 4, §3.3 · p. 4

4 Identifying Transfer Neurons

The authors hypothesize that MLP activations and their value vectors shift representations between language-specific and shared semantic spaces. They identify two transfer-neuron types and score candidate neurons by how they affect distance to the target space.

4.1 Hypothesis: Specific Activations and Value Vectors Facilitate a Parallel Shift of Representations to the Target Latent Space

  • Transfer Neurons are activations and corresponding value vectors hypothesized to shift representations between language-specific and shared semantic latent spaces.The proposed mechanism is a parallel shift toward the target latent space.Source: §4.1 · p. 4

4.2 Preparation: Two Types of Transfer Neurons

  • Type-1 neurons are sought in initial layers for input-to-shared-space transfer, while Type-2 neurons are sought in final layers for shared-to-language-specific output transfer.The study targets layers 1–20 for Type-1 and layers 21–32 for Type-2, based on where parallel and non-parallel sentence similarity differences widen.Source: §4.2 · p. 5

4.3 Scoring the Candidate Neurons

  • The scoring procedure estimates latent-space centroids from layerwise hidden states for parallel English–L2 sentence pairs, then ranks candidate neurons by score.English is assumed to serve as the shared semantic space; cosine similarity is used as the distance function, with Euclidean distance reported to yield similar results.Source: §4.3 · p. 5
  • Candidate neurons are ranked by whether their activation and value vector move representations closer to the target-space centroid than the pre-MLP hidden state does.Positive scores indicate movement closer for most samples; negative scores indicate movement farther away. The target centroid is shared for Type-1 and language-specific for Type-2.Source: §4.3 · p. 5

5 Detecting and Controlling Transfer Neurons

Transfer neurons appear at different layer locations by type, and deactivating Type-1 neurons disrupts the mapping of representations to the shared semantic latent space. The effect is distinct from random-neuron deactivation, though English-proximal languages retain some similarity separation.

5.1 Distribution

  • Type-1 neurons occur noticeably in early and middle layers, whereas Type-2 neurons are concentrated in the final layer; these distributions resemble language-specific neuron distributions and are consistent across languages and models.Figure 6 depicts the two types’ distributions.Source: §5.1 · p. 6, Figure 6 · p. 6

5.2 Representation Similarity Measurement while Deactivating Transfer Neurons

  • Deactivating only 0.2% of neurons sharply reduces the similarity gap between parallel and non-parallel sentence pairs; deactivating random neurons has almost no effect.The pattern is consistent across other language pairs and models, and the authors interpret it as Type-1 neurons being critical for mapping inputs into the shared semantic space.Source: §5.2 · p. 6
  • English-proximal Dutch and Italian retain a relatively larger similarity difference after Type-1 deactivation, which the authors attribute to their latent spaces already being close to the English-centric shared space.This remaining difference occurs across models.Source: §5.2 · p. 7
  • Deactivating Type-2 neurons also significantly inhibits latent-space transitions, according to the appendix results.Appendix ablations include different numbers of deactivated neurons.Source: §5.2 · p. 7

6 The Nature of Transfer Neurons

Deactivating a small set of Type-1 neurons degrades multilingual QA performance, while language specificity and overlap vary with transfer direction and linguistic distance. Similar language pairs share more transfer neurons, especially Type-2 neurons.

6.1 Language and Language-Family Specificity

  • Type-1 neurons generally show no language correlation except for Japanese-targeted transfers, while Type-2 language correlation strengthens as neurons move representations toward the target language-specific space.For both types, correlation strength tracks linguistic distance from English; Dutch and Italian tend to have sparser early-layer language-specific neurons and lower correlation scores.Source: §6.1 · p. 7
  • Linguistically similar language pairs tend to share more transfer neurons, particularly Type-2 neurons, consistent with their language-specific latent spaces occupying similar locations in the model’s hidden-state space.The authors identify facilitating movement between language-specific and shared semantic spaces as one role of language-specific neurons.Source: §6.1 · p. 8

6.2 Assessing the Importance of Transfer Neurons in Reasoning

  • For questions whose original generated answers had token-based F1 above 0.5, deactivating Type-1 neurons (0.2% of all neurons) significantly degraded performance, whereas baseline-neuron deactivation had almost no effect; this tendency held across models.Source: §6.2 · pp. 8–9
  • On MMLU-ProX, deactivating Type-1 neurons produces results highly consistent with the MKQA findings.MMLU-ProX is described as a multilingual benchmark designed to evaluate comprehensive reasoning ability.Source: §6.2 · p. 9

7 Discussion, Future Studies and Conclusion

The discussion links transfer direction to neuron language specificity and proposes that partly shared transfer trajectories explain language-family overlap. The paper concludes that transfer neurons are essential for reasoning and cross-lingual representation movement.

7 Discussion, Future Studies and Conclusion

  • Because Type-1 neurons target a shared English latent space, their overlap across languages increases as representations approach that space in middle layers.The authors contrast this language-agnostic pattern with greater language specificity among Japanese Type-1 neurons and Type-2 neurons.Source: §7 · p. 9, Figure 8 · p. 9
  • Transfer neurons include both language-specific and less language-specific types, extending prior language-specific-neuron findings to Type-1 and non-language-specific Type-2 neurons.The authors suggest partly shared transfer trajectories, especially among linguistically and spatially proximate languages, may explain language-family specificity.Source: §7 · p. 9
  • The conclusion states that transfer neurons are essential for reasoning and mediate movement between language-specific and shared semantic latent spaces.The authors propose selectively fine-tuning these neurons after multilingual pretraining as a possible way to align low-resource-language representations with the English-centric space.Source: §7 · p. 9

Limitations

The study analyzes representational transfer as a linear transformation within the MLP and leaves other modules unexamined. Its language and dataset evaluation also leaves other settings for future investigation.

Limitations

  • The representational-transfer analysis is limited to linear transformations in the MLP and does not consider other modules that may induce alternative vector movements.The authors derive the linear-shift framing from the MLP’s weighted sum of activations and value vectors.Source: Limitations · p. 10
  • Evaluation covers three LLMs and four languages per model, primarily using Tatoeba and MKQA; other languages and datasets, including longer inputs, warrant investigation.The tested languages include Dutch and Italian, which are linguistically close to English, and Japanese and Korean, which are not.Source: Limitations · p. 10

A Experimental Settings

The experiments use three multilingual LLMs and multilingual sentence and QA data, with sampled representations and train-test separation for identifying and testing transfer neurons. Across the models, the English latent-space centroid moves non-linearly toward the middle layers, motivating layer-wise centroids for neuron identification.

A Experimental Settings

  • Experiments use LLaMA3-8B, Mistral-7B, and Aya expanse-8B, with Tatoeba for similarity and language-specificity analyses and MKQA for QA and dimensionality reduction.They sample 1k sentences per language for most analyses and 1k hidden states per layer and language for the SVD experiments.Source: A Experimental Settings · p. 12
  • Transfer-neuron identification and deactivation use a 50:50 train-test split, and similarity remeasurement uses sentences separate from those used to identify neurons.The authors describe this as a test of generalization performance.Source: A Experimental Settings · p. 12

B Trajectory of the English Latent Space Moving towards Middle Layers

  • Trajectory similarity compares the actual layer-wise centroid path with a linear path from decoder layer 1 to the middle layer.A lower similarity score indicates non-linear movement; the authors set the middle layer to 10 based on observed alignment patterns.Source: B Trajectory of the English Latent… · pp. 12–13
  • The English latent-space trajectory moves non-linearly toward the middle layers across all three adopted models.The authors therefore use layer-wise centroids to identify neurons contributing to shifts toward the moving English latent space.Source: B Trajectory of the English Latent… · p. 13

C.1 Similarity of Hidden States and Activation Patterns for Parallel Sentences across Languages

Across languages, parallel sentence representations and their MLP activation patterns are more similar than non-parallel ones in the middle layers, supporting a shared semantic latent space. Additional analyses report separability, latent-space dimensionality and distance patterns, and define the cross-language mutual k-NN alignment measure.

C.1 Similarity of Hidden States and Activation Patterns for Parallel Sentences across Languages

  • Parallel sentences show a larger similarity advantage over non-parallel sentences in middle layers for both hidden states and MLP activation patterns, consistently across language pairs and models.The authors interpret this pattern as evidence for a shared semantic latent space in the middle layers.Source: §C.1 · pp. 13–14, Figure 10 · p. 14, Figure 11 · p. 15

C.2 Investigating Linear Separability between Parallel and Non-Parallel Inner Representations with Logistic Regression Model

  • Most layers after the initial few achieve approximately over 70% test accuracy in classifying parallel versus non-parallel sentence pairs from concatenated hidden states.The authors report higher classification accuracy in layers with greater similarity between language latent spaces.Source: §C.2 · p. 14, Figure 12 · p. 16

D.1 Dimensionality of Latent Spaces Across Layers

  • Across languages and layers, Fig. 13 shows each language latent space has lower dimensionality than hidden states, indicating that these latent spaces lie in the hidden-state space.Source: §D.1 · p. 14

D.2 Distance among Language Latent Spaces

  • Language latent spaces remain relatively distant in the initial and final layers but become closer in the middle layers, where language-agnostic semantic processing and reasoning mainly occur.The reported centroid-distance results cover Figs. 14–16.Source: §D.2 · p. 15

E.1 Formalization of the Computation for Mutual k-NN Alignment Metric across Layers and Input Language Pairs

  • The mutual k-NN alignment metric is adapted to compare layer-wise English–L2 latent spaces within one model using parallel sentence representations.The setup pairs English and L2 sentences expressing the same meaning and measures the average intersection of their k-nearest-neighbor sets.Source: §E.1 · pp. 15–16

E.2 Result of Mistral-7B

The appendices report that neuron deactivation affects latent-space transitions, with Type-1 and Type-2 neurons associated with different directions of movement. They also describe how neurons are operationally defined and detected, and report language-family and overlap patterns for the identified neurons.

E.2 Result of Mistral-7B

  • Mistral-7B shows results similar to LLaMA3-8B and Aya expanse-8B on mutual k-NN alignment, while Dutch and Italian tend to have more similar latent spaces to English than more distant language-family inputs.The passage does not provide numerical comparisons.Source: §E.2 · p. 16

G.1 The Definition of Neurons in This Study

  • In this study, neurons are defined as the element-wise product activation values, which the authors consider more indicative of neuron activity in gated MLPs than the nonlinear activation output alone.The up-projection output can reduce the actual activation value even when the nonlinear activation output is large, attenuating its impact on subsequent neurons.Source: §G.1 · p. 16

G.2.1 Kernel-Based Similarity between Latent Spaces

  • Deactivating the top-1k Type-1 neurons substantially reduces English–L2 kernel-based similarity, especially in middle layers, while relatively high similarity remains for Dutch and Italian.The authors connect the reduction to Type-1 neurons' role in shifting representations toward the shared semantic latent space; they attribute remaining similarity for Dutch and Italian to their spaces being close even in early layers.Source: §G.2.1 · p. 17

G.2.3 Quantitative Distance among Language Latent Spaces

  • Deactivating Type-1 neurons hinders movement toward the English latent space more than deactivating randomly sampled neurons, consistently across languages and models.The reported distance results appear in Figs. 28–30.Source: §G.2.3 · p. 17

G.3.1 Deactivating Type-2 Transfer Neurons Significantly Inhibit Spatial Transition from the Shared Semantic Latent Space to the Language-Specific Latent Spaces

  • Deactivating 0.2% of model neurons as top-1k Type-2 neurons delays movement toward language-specific latent spaces, particularly in final layers, relative to a same-layer random baseline.The distance analysis also reports inhibited movement toward language-specific spaces in final layers under Type-2 deactivation.Source: §G.3.1 · pp. 17–18, §G.3.2 · p. 18

H.2 Overlap Ratio of Transfer Neurons

  • Type-1 neuron overlap across language pairs increases as representations move closer toward the shared latent space, whereas Type-2 overlap decreases as representations move toward language-specific output spaces.Source: §H.2 · p. 18

H.3 Detecting Language-Specific Neurons

  • Language-specific neurons are detected using correlation ratio rather than average precision, allowing one score to capture positive and negative activation-label correlations; these neurons tend to occur in initial and final layers and have a distribution similar to transfer neurons.The correlation-ratio analysis uses language-labeled sentences; the passage reports the distributional pattern, not values from the referenced tables.Source: §H.3 · p. 18

H.5.2 Correlation Ratio for Language-Family Specificity

  • Language-family specificity is particularly evident in Type-2 neurons when English, Dutch, and Italian are grouped together and Japanese and Korean as another family.The correlation ratio is used to summarize association between the language-family labels and neuron activations.Source: §H.5.2 · p. 19, §H.6 · p. 19

H.7 Hypothesis Testing for Reported Correlation Values

The appendix tests whether transfer-neuron activation values vary with language or language-family labels using ANOVA, with a Mann–Whitney U test as a non-parametric check. It also reports cross-lingual deactivation effects and reasoning-task outcomes for MKQA and MMLU-ProX.

H.7 Hypothesis Testing for Reported Correlation Values

  • Approximately 80–95% of top-1k transfer neurons are statistically significant, and Mann–Whitney U results are highly consistent with ANOVA.The reported significance proportions use p < 0.05 and correlation-ratio thresholds of 0.1 and 0.25; the analyses address language and language-family associations.Source: §H.7 · pp. 19–20

I The Effects of Transfer Neurons on Internal Representations under Cross-Lingual Deactivation

  • Cross-lingual Type-1 deactivation has a noticeable effect on L2 hidden-state dynamics, whereas Type-2 deactivation has little to no influence except when the target language spaces are close.The comparison is against the L2 deactivation baseline and is consistent with Type-1 neurons being less language-specific and Type-2 neurons more language-specific.Source: I The Effects of Transfer Neurons on… · p. 20

J.1 Multilingual Knowledge QA (MKQA)

  • The MKQA deactivation results are reported as highly consistent with the paper's earlier §6.2 results.Source: §J.1 · p. 20

J.2 MMLU-ProX

  • On MMLU-ProX, deactivating top-1k Type-1 neurons significantly degrades performance, while deactivating a same-layer random set causes negligible change; the authors describe this as highlighting Type-1 neurons' critical role in reasoning.The reported languages are Japanese, Korean, and French because the evaluation tool had limited language coverage.Source: §J.2 · pp. 20–21

J.3 Open-Ended Text Generations while Deactivating Transfer Neurons

This section examines open-ended generations after deactivating top-1k Type-1 or Type-2 neurons, using manual comparisons across three models and illustrative examples. It describes degraded generation with Type-1 deactivation and language changes, including Korean/Japanese code-switching-like output, with Type-2 deactivation.

J.3 Open-Ended Text Generations while Deactivating Transfer Neurons

  • The authors evaluated generation quality on PolyWrite, comparing outputs across languages under no intervention, top-1k Type-1 neuron intervention, and top-1k Type-2 neuron intervention using greedy decoding.PolyWrite is a multilingual open-ended generation task with no single correct answer (Ji et al., 2025). The authors used machine translation to assess Dutch, Korean, and Italian outputs because they were not proficient in those languages.Source: §J.3 · p. 21
  • Type-1 deactivation made LLaMA3-8B generate gibberish across languages, while Mistral-7B and Aya expanse-8B showed model- and language-specific declines in fluency, coherence, or sentence variety.Mistral-7B showed repetition and fragmentation, particularly in Italian and Dutch; Aya expanse-8B remained generally coherent, with mild regressions that varied by language.Source: §J.3 · p. 21
  • With Type-2 neurons deactivated, all three models sometimes produced coherent answers in another language, often linguistically close to the input, despite fluency or repetition problems.LLaMA3-8B also produced gibberish in many samples; Mistral-7B showed looping and fragmented or prematurely cut-off outputs, while Aya expanse-8B often repeated itself and sometimes code-switched.Source: §J.3 · pp. 21–22
  • After deactivating the top-1k Type-2 neurons, example generations changed output language from input in Mistral-7B and Aya expanse-8B.Source: Table 39 · p. 55, Table 40 · p. 56
  • After deactivation, Aya expanse-8B generated an example in a code-switching–like form in Korean and Japanese.Source: Table 41 · p. 57

Related papers

Loading 2509.17030v1…