Research map / Large language models / Language, safety and interpretability
Exploring the Translation Mechanism of Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes subspace-intervened path patching to locate translation-crucial attention heads and MLPs in LLMs, characterizes their roles, and shows fine-tuning 64 heads (<5% params) matches or beats full SFT on translation while preserving general ability.
Problem
Interpreting the internal computational mechanisms behind LLM translation, even at the word level, which remain poorly understood; prior analyses were observational (neurons, representations) rather than causal.
Approach
Refines path patching by confining the intervention to a low-dimensional translation-steering subspace (identified via SVD-based decomposition of contrastive activation differences between translation and counterfactual prompts) instead of patching full activation vectors, then uses the detected heads/MLPs for analysis and targeted fine-tuning.
Key results
- Targeted SFT of 64 attention heads (0.27B params) reaches Zh⇒En 21.3/79.1/64.3 BLEU/COMET/BLEURT versus 20.4/78.7/63.9 for Full SFT (6.7B params), while MMLU is 45.9 vs 42.6 for Full SFT.
- Crucial heads are sparse (under 5% of heads) and concentrated in layers 12–20 and the final two layers; language pairs sharing source or target language overlap over 70% in crucial heads.
- Knocking out crucial heads drops translation accuracy substantially while random-head knockout keeps accuracy within 2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune translation-crucial components instead of full parameters | LLaMA2-7B · Zh⇒En, En⇒Zh, De⇒En, En⇒De | BLEU: 15.6 -> 21.3; 17.0 -> 30.7; 24.8 -> 37.1; 13.0 -> 27.6 | Mixed |
| Fine-tune translation-crucial components instead of random components | LLaMA2-7B · Zh⇒En, En⇒Zh, De⇒En, En⇒De | BLEU: 16.9 -> 21.3; 26.4 -> 30.7; 32.5 -> 37.1; 22.7 -> 27.6 | Helped |
| Use full-parameter SFT rather than the pretrained baseline | LLaMA2-7B · MMLU | accuracy: 45.9 -> 42.6 | Hurt |
| Fine-tune translation-crucial components rather than full parameters | LLaMA2-7B · MMLU | accuracy: 45.9 -> 46.0 | No clear effect |
| Increase the number of fine-tuned attention heads from 8 to 64 (ablation) | LLaMA2-7B · Zh⇒En | BLEU: 18.7 -> 21.3 | Helped |
| Increase fine-tuned attention heads from 64 to 160 (ablation) | LLaMA2-7B · Zh⇒En | BLEU: 21.3 -> 21.3 | No clear effect |
| Increase fine-tuned attention heads from 64 to 160 (ablation) | LLaMA2-7B · MMLU | accuracy: 45.9 -> 45.3 | Hurt |
| Add MLPs to the top-64-head configuration (ablation) | LLaMA2-7B · Zh⇒En | BLEU: 21.3 -> 22.5 | Helped |
| Add MLPs to the top-64-head configuration (ablation) | LLaMA2-7B · MMLU | accuracy: 45.8 -> 42.8 | Hurt |
| Use subspace-intervened rather than standard path patching to identify top-5 heads | LLaMA2-7B · En⇒Zh | accuracy drop: -25% -> -39% | Helped |
| Use subspace-intervened rather than standard path patching to identify top-32 heads for targeted SFT | LLaMA2-7B · En⇒Zh | BLEU: 27.3 -> 28.9 | Helped |
| Knock out five heads shared by word- and sentence-level translation (ablation) | LLaMA2-7B · Sentence-level translation | PPL: +36% in PPL | Hurt |
Surprising findings
- Full-parameter SFT degrades general benchmarks (MMLU 45.9 to 42.6) while fine-tuning only attention heads preserves or slightly improves them.
- CPT, not SFT, produces significant distributional shift in translation-crucial heads (K-S p<0.00001 vs p=0.355 for SFT).
Limitations
- Analysis focuses on word-level translation and LLaMA-2-7B; generalization to other models is shown only in appendix.
- Method is claimed to be theoretically constrained for closed-source models, which cannot be analyzed.
Experimental setup
Datasets: WMT17-22 validation/parallel data Flores-200 OSCAR WMT23 WMT24 MGSM MMLU ARC SIQA
Benchmarks: WMT23 WMT24 MMLU ARC SIQA MGSM
Metrics: BLEU COMET BLEURT accuracy logit change perplexity
Baselines: standard path patching Full SFT Random SFT random-head knockout
Builds on: path patching activation patching linear representation hypothesis
Closest papers by meaning
- A Paradigm Shift in Machine Translation: Boosting Translation Performance of Large Language Models (ICLR 2024 · similarity 0.89)
- When Scaling Meets LLM Finetuning: The Effect of Data, Model and Finetuning Method (ICLR 2024 · similarity 0.87)
- How do Large Language Models Handle Multilingualism? (NeurIPS 2024 · similarity 0.87)
- Model-Aware Tokenizer Transfer (NeurIPS 2026 · similarity 0.87)
- Seeking Neural Nuggets: Knowledge Transfer in Large Language Models from a Parametric Perspective (ICLR 2024 · similarity 0.87)
- The Same but Different: Structural Similarities and Differences in Multilingual Language Modeling (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.