Skip to content

Distribution-aware Language Neuron Identification in Multilingual Large Language Models

Conference: EMNLP 2026
Paper: ACL Anthology
Area: LLM Reasoning
Keywords: Multilingual LLM, Language-Specific Neurons, Distribution Overlap Coefficient, Activation Densities, Causal Interventions

TL;DR

This paper demonstrates that conventional methods thresholding positive activations discard essential density shapes and negative regimes, proposing Distribution-aware Language Neuron identification (DLN) via pairwise distribution overlap clustering to effectively isolate single-language neurons with up to 4.9x higher on-target damage per neuron and discover cohesive multi-language coalitions.

Background & Motivation

Multilingual large language models (mLLMs) exhibit remarkable capabilities across dozens of diverse languages within a single shared parameter space. However, how these models internally organize language-specific processing versus language-shared representations remains a central open question in mechanistic interpretability. Prior neuron-level studies locate language-specific neurons (LSNs) in feed-forward networks (MLP/FFN blocks) by measuring how selectively individual neurons fire for a given language. Dominant approaches, such as LAPE and activation-probability entropy baselines, operationalize language specificity through a binary indicator: a neuron is treated as active only when its continuous activation value is strictly positive (\(a > 0\)), and language selectivity is quantified via the entropy of normalized per-language activation rates.

This binary abstraction introduces two fundamental blind spots that clash with the distributional nature of mLLM representations. First, binarization discards all information embedded in the negative activation regime. In modern architectures employing Gated Linear Units (GLU) with activation functions like SiLU, the majority of activations across the network are negative; numerous neurons exhibit distinct distributional separation across languages exclusively below zero, yet are completely discarded as inactive by positive-rate metrics. Second, computing scalar probabilities independently across languages destroys the geometric relational structure and continuous density profile: languages with nearly identical positive activation probabilities may occupy non-overlapping continuous activation ranges, while languages showing widespread positive firing may still feature sharp magnitude shifts that encode exclusive linguistic functions.

Consequently, evaluating language neurons purely on positive firing rates leads to severe misidentifications—such as completely overlooking English-specific functional subsets and mischaracterizing cross-lingual shared circuits. Core idea: reframe language neuron identification as a continuous density clustering problem over pairwise distribution overlap coefficients across the full activation range, capturing both single-language neurons and multi-language coalitions while preserving functional negative-regime circuits.

Method

Overall Architecture

The DLN framework eliminates binary thresholding by directly estimating continuous probability density functions of neural activations for each language and grouping languages using an overlap matrix. The pipeline consists of four serial stages: (1) forward-pass estimation of continuous activation densities per neuron across languages using adaptive-range histograms; (2) construction of a pairwise distribution overlap coefficient matrix measuring shared density area; (3) single-linkage bipartition clustering (\(K=2\)) to split languages into a distinctive minority cluster and a background majority cluster; and (4) percentile-based candidate thresholding, naturally categorizing selected units into single-language neurons (SLNs) or multi-language neurons (MLNs).

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Multilingual Calibration Corpora<br/>(100k tokens per language)"] --> B["Continuous Activation Density Estimation<br/>400-bin adaptive-range histograms"]
    B --> C["Distribution Overlap Matrix Construction<br/>Pairwise overlap integrals Sj[k, k']"]
    C --> D["Single-linkage Bipartition Clustering<br/>Minimizing max between-cluster overlap (Aj, Acj)"]
    D --> E["Percentile Thresholding & Categorization<br/>Threshold τ splits SLN (|Aj|=1) / MLN (|Aj|>1)"]
    E --> F["Mean-patching Causal Intervention<br/>Replacing target activations with off-target reference"]

Key Designs

1. Continuous Density Estimation and Overlap Coefficient Matrix: capturing full-range distributional discriminability

To retain signals across both positive and negative values, DLN computes continuous probability density functions \(p_k^j(v)\) for each neuron \(j = (\ell, n)\) under language \(k \in \mathcal{L}\). Practically, density integration is realized using an adaptive histogram of 400 inner bins bounded by \([\mu_j - 3\sigma_j, \mu_j + 3\sigma_j]\) over pooled language activations. For any language pair \((k, k')\), the overlap coefficient \(S_j[k, k']\) calculates the shared distribution area: $\(S_j[k, k'] = \int \min\left(p_k^j(v), p_{k'}^j(v)\right) \mathrm{d}v\)$ This metric is mathematically equivalent to \(1 - \text{TV}(p_k^j, p_{k'}^j)\) (one minus the total variation distance), providing a bounded, scale-invariant statistic in \([0, 1]\). It captures subtle density shifts, whether situated in large positive activations or entirely within negative suppression zones, avoiding the variance-inflation pathologies of scale-bearing metrics like Wasserstein-1.

2. Single-linkage Bipartition Clustering: isolating distinctive language subsets without bimodality priors

With the pairwise matrix \(S_j\) in hand, DLN segments the language set \(\mathcal{L}\) into a bipartition \((A_j, A_j^c)\) by minimizing the maximum between-cluster overlap across all non-trivial splits: $\(\min_{A_j \subset \mathcal{L}} \max_{k \in A_j, k' \in A_j^c} S_j[k, k']\)$ This formulation corresponds exactly to single-linkage hierarchical clustering with cluster count \(K=2\). The smaller cluster \(A_j\) is designated as the candidate language subset associated with the neuron, while the larger complement \(A_j^c\) functions as the reference background. This criterion makes no rigid bimodality assumption: neurons with multi-group separations are detected via their most distinctive bipartition without requiring prior tuning of language groups.

3. Percentile Thresholding and Counterfactual Mean-patching: unified categorization and matched intervention

Neurons are ranked globally by their maximum between-cluster overlap, admitting neuron \(j\) if \(\max_{k \in A_j, k' \in A_j^c} S_j[k, k'] \le \tau\), where \(\tau\) is set to the bottom \(P\)-th percentile across all candidate neurons. The size of \(A_j\) naturally defines the neuron type: \(|A_j| = 1\) denotes a Single-Language Neuron (SLN), while \(|A_j| > 1\) denotes a Multi-Language Neuron (MLN). In causal validation, rather than using zero-ablation (which damages language-neutral negative baselines), DLN applies mean-patching: replacing the target activation \(a_{j^*}\) with the empirical mean activation over non-target languages \(\mu_{\text{off}}\). This provides a principled counterfactual that respects the neuron's continuous first-moment manifold.

Benchmark & Evaluation Protocol

The empirical framework evaluates two base autoregressive models with GLU feed-forward designs (Llama-3.1-8B and SmolLM3-3B-Base), extending to Qwen3-4B-Base, Aya-23-8B, and Llama-3.1-70B. Seven calibration languages spanning four scripts and five families are tested: English (en), Chinese (zh), French (fr), Spanish (es), Vietnamese (vi), Indonesian (id), and Japanese (ja), using 100k tokens per language from Wikipedia-20231101. Evaluation utilizes disjoint corpora: FLORES+ devtest and Wikipedia-20230601 (up to 1M tokens per language). Causal selectivity is measured via negative log-likelihood inflation \(\Delta\text{NLL}\), selectivity ratios \(\text{Sel}(\mathcal{N})\), and downstream zero-shot accuracy drops on Belebele (reading comprehension, 900 questions/lang) and MGSM (grade-school math, 250 problems/lang, 8-shot CoT).

Key Experimental Results

Main Results

At the standard operating threshold \(P = 1\%\), Table 2 and Figure 3 benchmark DLN's SLN against LAPE (Tang et al., 2024) and LSN (Zhang et al., 2026) across language modeling and downstream tasks:

Model / Method Pool Size \(\|\mathcal{N}\|\) WIKI \(\Delta\text{NLL}\) (on / off) FLORES \(\Delta\text{NLL}\) (on / off) Belebele Acc Drop (on / off) MGSM Acc Drop (on / off)
Llama-3.1-8B
LAPE 4,587 0.28 / 0.01 0.35 / 0.01 2.75 / 0.07 6.96 / 0.88
LSN 1,814 0.23 / 0.00 0.30 / 0.00 1.22 / -0.02 4.40 / 0.30
SLN (Ours) 2,972 1.20 / 0.01 1.69 / 0.01 4.94 / -0.02 9.44 / 0.18
SmolLM3-3B-Base
LAPE 3,962 0.44 / 0.02 0.49 / 0.02 2.46 / 0.14 3.60 / 0.06
LSN 1,443 0.42 / 0.00 0.48 / 0.00 1.25 / 0.01 0.72 / -0.64
SLN (Ours) 2,254 1.45 / 0.01 1.86 / 0.01 5.56 / 0.15 7.76 / -0.40

Per-neuron normalized analysis shows that SLN produces 4.9x (WIKI) and 5.4x (FLORES) higher on-target damage per neuron than LAPE on Llama-3.1-8B, and 3.1x / 3.2x over LSN. Crucially, on English, both baselines fail completely (inducing near 0 \(\Delta\text{NLL}\)), whereas SLN achieves \(\Delta\text{NLL}_{\text{en}} = 1.47\) (WIKI) and 1.87 (FLORES).

Ablation and Analysis Experiments

Decomposition of neuron sets into intersections and differences demonstrates that DLN uncovers high-impact functional neurons missed by probability-based baselines (Table 3), while site-robustness tests confirm DLN generalizes to the gate-up product \(z\)-site (Table 5):

Neuron Sub-Pool Pool Size \(\|\mathcal{N}\|\) WIKI On-Target \(\Delta\text{NLL}\) WIKI Off-Target \(\Delta\text{NLL}\) Selectivity \(\text{Sel}(\mathcal{N})\) (WIKI) Selectivity \(\text{Sel}(\mathcal{N})\) (FLORES)
Llama-3.1-8B Pool Split
LAPE \(\setminus\) SLN 3,598 0.103 0.021 4.95 6.30
LAPE \(\cap\) SLN 989 0.265 0.001 203.07 460.62
SLN \(\setminus\) LAPE 1,983 0.295 0.004 75.17 108.83
SmolLM3-3B Pool Split
LAPE \(\setminus\) SLN 3,213 0.150 0.049 3.06 4.62
LAPE \(\cap\) SLN 749 0.331 0.001 238.92 408.44
SLN \(\setminus\) LAPE 1,505 0.707 0.003 204.26 340.75

Testing identification across GLU functional sites demonstrates clear architectural robustness:

Configuration (select-site, patch-site) LAPE Selectivity LSN Selectivity SLN (Ours) Selectivity
Llama-3.1-8B
(a-site, a-site) 22.8 126.0 273.7
(z-site, a-site) 4.5 (-80%) 13.3 (-89%) 193.9 (-29%)
(z-site, z-site) 3.9 9.9 97.2
SmolLM3-3B-Base
(a-site, a-site) 28.2 89.5 296.5
(z-site, a-site) 7.4 (-74%) 8.8 (-90%) 187.3 (-37%)
(z-site, z-site) 5.3 9.2 57.4

Key Findings

  • Negative-regime neurons drive directional language switching: Between 63% and 72% of English SLNs exhibit a negative target mean (\(\mu_{\text{tgt}} < 0\)). In cross-lingual steering experiments, positive-regime English neurons govern English text generation, while negative-regime English neurons form a functional circuit that steers generation away from English into other target languages with 75% accuracy.
  • Chinese representation is predominantly organized in multi-language coalitions: Only 6.7% of Chinese-associated neurons in Llama-3.1-8B are single-language neurons; 93.3% belong to MLNs, with 57.6% forming a {zh, ja} coalition. Ablating this cluster with just 232 neurons (2.86x fewer than LRN) induces massive \(\Delta\text{NLL}\) damage of 1.24 on Chinese and 0.92 on Japanese, exposing character-level shared representations.
  • Displacement-matched control rules out magnitude artifacts: Applying an identical normalized displacement (1.6–2.2\(\sigma\)) to layer-matched random neurons yields only \(\sim\)1% of the on-target damage (0.013 vs 1.198) with zero selectivity, proving that DLN's impact stems from true functional specificity rather than perturbation size.

Highlights & Insights

  • Continuous geometric density metric: By replacing binary rate entropy with bounded overlap integrals, DLN seamlessly captures magnitude shifts and negative-domain separations invisible to indicator methods.
  • Natural SLN/MLN unification via hierarchical clustering: Single-linkage clustering on pairwise overlaps resolves both isolated single-language units and structurally coupled coalitions without arbitrary multi-language thresholds.
  • Viability at causally operative GLU sites: While the gate-up product (\(z\)-site) has sign-balanced statistics (\(P(z>0) \approx 0.5\)) that collapse entropy baselines, DLN retains robust selectivity (outperforming baselines by 15–44x).

Limitations & Future Work

  • Language diversity scope: The experimental pipeline evaluates seven representative languages; exploring wider language pools, particularly low-resource tongues, remains necessary.
  • Architecture focus: The study investigates GLU feed-forward blocks with SiLU; behavior on traditional non-gated MLPs (e.g., standard ReLU/GELU architectures) remains unexplored.
  • Magnitude compression at the \(z\)-site: Cross-language activation spread at the \(z\)-site is \(\sim\)4x narrower than at the \(a\)-site, naturally reducing the absolute magnitude of mean-patching interventions across all methods.
  • vs LAPE (Tang et al., 2024): LAPE relies on entropy over binary firing rates (\(a > 0\)), ignoring negative-range distributions and continuous density shapes. DLN recovers 4.9x higher damage per neuron and captures negative-regime English circuits.
  • vs LSN & LRN (Zhang et al., 2026): Zhang et al. augment entropy with maximum activation rates but remain restricted to positive-activation indicators. At the \(z\)-site, their performance drops by 90%, whereas DLN maintains high selectivity and isolates MLN coalitions using 2.86x fewer neurons.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Formulates language neuron identification via continuous overlap density clustering, fundamentally overcoming the binary indicator bottleneck.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorously tested across five models (up to 70B), two tasks, random displacement controls, steering interventions, and alternative GLU sites.
  • Writing Quality: ⭐⭐⭐⭐⭐ Clear exposition, elegant mathematical formulation, and deep qualitative insights into negative activation dynamics and multi-language coalitions.
  • Value: ⭐⭐⭐⭐⭐ Establishes a distribution-aware benchmark for mechanistic interpretability in mLLMs, offering valuable tools for multilingual steering and editing.