Exploring the Translation Mechanism of Large Language Models

Conference: NeurIPS 2025 arXiv: 2502.11806 Code: Available (link provided in paper) Area: Multilingual Translation Keywords: translation mechanism, mechanistic interpretability, attention head, path patching, subspace intervention

TL;DR

This paper proposes a subspace-intervened path patching method for fine-grained causal analysis of the translation mechanism in LLMs. The study finds that translation is driven by a sparse set of attention heads comprising fewer than 5% of all heads, categorized into three functional roles: source heads, indicator heads, and positional heads. MLP layers integrate these features into an English-centric intermediate representation, and fine-tuning only 64 critical heads achieves performance comparable to full-parameter fine-tuning.

Background & Motivation

Background: LLMs exhibit strong multilingual translation capabilities, yet the internal core translation mechanism—even for basic word-level translation—remains poorly understood. Prior analyses have largely remained at the level of surface observations (neuron activation patterns, intermediate representation visualization) rather than revealing causal computational mechanisms.

Limitations of Prior Work: (1) Conventional path patching intervenes on entire activation vectors, resulting in coarse granularity and noisy estimates; (2) systematic study of the translation mechanism in decoder-only LLMs is lacking (prior work focused on encoder-decoder architectures); (3) it is unclear which attention heads serve which functions or how MLP layers participate in translation.

Key Challenge: There is a need to precisely localize translation-relevant causal effects within the high-dimensional activation space of LLMs while filtering out activation dimensions unrelated to translation.

Goal: Systematically answer three questions: which components are critical for translation? What behavioral patterns do these components exhibit? Can fine-tuning these components improve translation performance?

Key Insight: The linear representation hypothesis—linear subspaces of activation vectors constitute the most interpretable model components. A "translation-oriented subspace" is extracted by contrasting positive/negative data pairs (with/without translation logic), and interventions are applied exclusively within this subspace.

Core Idea: Perform path patching within the translation-oriented subspace to precisely localize critical translation components, uncovering a sparse translation circuit consisting of three functionally differentiated attention head types and English-centric MLP processing.

Method

Overall Architecture

A three-stage systematic framework: (1) identify critical translation components via subspace-intervened path patching; (2) analyze the functional roles of critical attention heads and the representational properties of MLP layers; (3) design and validate a targeted fine-tuning strategy based on the findings.

Key Designs

1. Subspace-Intervened Path Patching:

   • Function: Performs causal intervention within the "translation-oriented subspace" of component activations rather than on the entire activation vector.
   • Mechanism: Contrastive data pairs (with translation logic \(X_+\) vs. without translation logic \(X_-\)) are used to compute an activation difference matrix \(\mathbf{M}_c\). Orthogonal decomposition (optimization objective Eq. 1) partitions it into a general translation-oriented subspace \(\mathbf{S}_c\) and a dataset-specific subspace \(\mathbf{E}_c\). Intervention replaces only the component along the \(\mathbf{S}_c\) direction: \(\tilde{\mathbf{a}}_c = W_cW_c^T\mathbf{a}_c(X_-) + (I-W_cW_c^T)\mathbf{a}_c(X_+)\)
   • Design Motivation: Standard path patching, which replaces the entire activation vector, introduces interference from dimensions unrelated to translation; subspace projection precisely isolates the translation signal.

2. Three Functionally Differentiated Attention Head Types:

   • Function: Critical heads are categorized into source heads (attending to source-language tokens), indicator heads (attending to translation instruction tokens such as "Chinese:"), and positional heads (maintaining sequential position information) based on attention weight analysis.
   • Mechanism: The attention distribution of each critical head is analyzed for its alignment with source-language words, instruction tokens, and positional indices.
   • Design Motivation: To understand why these heads are important—each extracts a different type of information required for the translation task.

3. English-Centric MLP Processing:

   • Function: Demonstrates that MLP layers integrate multilingual features extracted by attention heads into an English-centric intermediate representation.
   • Mechanism: The correlation between hidden representations at each MLP layer and the token embeddings of English, source-language, and target-language tokens is measured; intermediate MLP representations are found to be highly correlated with English embeddings.
   • Design Motivation: Provides causal-level confirmation of the previously proposed hypothesis that LLMs use English as an implicit computational pivot.

Validation: Targeted Fine-Tuning

Fine-tuning only the 64 identified critical attention heads (<5% of parameters) matches or exceeds full-parameter fine-tuning performance on both word-level and sentence-level translation.

Key Experimental Results

Critical Component Statistics (LLaMA2-7B)

Metric	Value
Proportion of translation-critical attention heads	<5%
Head overlap rate across language pairs (same source/target)	>70%
Head overlap rate across bidirectional translation pairs	>60%
Layer range of critical head concentration	Layers 12–20 + last 2 layers

Targeted Fine-Tuning vs. Full-Parameter Fine-Tuning

Configuration	Parameters	Translation Performance
Full-parameter fine-tuning	100%	Baseline
Fine-tune 64 critical heads only	<5%	Comparable or better
Fine-tune top-5 shared heads	—	Word-level −39% logits

Knockout Validation

Operation	Change in Translation Accuracy
Incrementally knock out critical heads	Significant drop (90% → <30%)
Incrementally knock out random heads	Fluctuation <2%
Knock out critical MLPs	Similarly significant drop

Key Findings

• Translation-critical heads exhibit strong cross-lingual transferability—different translation directions share a large proportion of critical heads.
• The translation criticality of MLP layers is concentrated in layers 15 and beyond; the final MLP layer accounts for up to 50% of the change in target token logits.
• Low-resource languages (Swahili, Bengali, Arabic) exhibit the same sparsity and transferability patterns.

Highlights & Insights

• Subspace-projected path patching: By intervening within a translation-specific subspace, this approach substantially improves the precision and interpretability of causal analysis, and is generalizable to mechanistic analysis of other tasks.
• Functional differentiation among three head types: The translation process is clearly decomposed into three sub-functions: extracting source-language content, recognizing translation task signals, and maintaining positional information.
• <5% parameter fine-tuning experiment: Mechanistic understanding is directly translated into a practical parameter-efficient fine-tuning strategy.

Limitations & Future Work

• Analysis is primarily conducted at the word level: Although sentence-level transfer is validated, the core analysis remains in a simplified word-level setting.
• Only LLaMA2-7B is analyzed: The translation mechanism in larger models (70B+) may differ.
• Counterfactual template design: The negative sample construction (replacing translation instruction tokens) may not fully isolate the translation signal.

Related Work & Insights

• vs. Voita et al. (2019) encoder-decoder head pruning: This paper performs analogous analysis on decoder-only LLMs for the first time, finding that the conclusions regarding sparsity and functional differentiation are consistent across architectures.
• vs. Wendler et al. (2024) English pivot hypothesis: The English-centric processing hypothesis is confirmed at the causal level (rather than merely at the correlational level).

Rating

• Novelty: ⭐⭐⭐⭐ Subspace path patching and the three-category head taxonomy are original contributions.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Six translation directions + low-resource languages + sentence-level validation + mathematical reasoning transfer + knockout experiments + fine-tuning.
• Writing Quality: ⭐⭐⭐⭐ Analysis proceeds in a well-structured, layered manner with clear figures and tables.
• Value: ⭐⭐⭐⭐⭐ Represents a significant advance in understanding the LLM translation mechanism; the targeted fine-tuning strategy has direct practical utility.

Related Papers

• [NeurIPS 2025] XIFBench: Evaluating Large Language Models on Multilingual Instruction Following
• [ACL 2026] Exploring Two-Phase Continual Instruction Fine-tuning for Multilingual Adaptation in Large Language Models
• [NeurIPS 2025] ParallelPrompt: Extracting Parallelism from Large Language Model Queries
• [AAAI 2026] Consensus-Aligned Neuron Efficient Fine-Tuning Large Language Models for Multi-Domain Machine Translation
• [AAAI 2026] Focusing on Language: Revealing and Exploiting Language Attention Heads in Multilingual Large Language Models