TLDR: This research investigates why large language models (LLMs) struggle with cross-lingual knowledge transfer, often hallucinating when prompted in a different language than their training data. Using small Transformer models on synthetic datasets, the study identifies a crucial learning phase where models develop either unified or separate representations of facts across languages. Unification is found to be essential for cross-lingual transfer and is influenced by parallel data, as well as the informativeness and extractability of the language feature in monolingual data. The paper introduces a ‘unification score’ that strongly predicts cross-lingual performance and validates these findings on larger LLMs like Gemma, suggesting new strategies for improving multilingual AI.
Large language models (LLMs) have revolutionized many aspects of technology, but they still face significant challenges. One puzzling issue is cross-lingual hallucination: when an LLM accurately answers a question in the language it was trained on, but provides incorrect or fabricated information when asked the same question in a different language. This problem is particularly concerning for speakers of less-resourced languages, and simply making models larger doesn’t seem to solve it.
Previous research has observed that LLMs can develop a ‘lingua franca’ (often English-based) for factual knowledge, alongside distinct language-specific ‘silos’ within their internal representations. However, the underlying reasons for these phenomena and how they develop during training have remained unclear, making it difficult to devise effective solutions.
A Controlled Environment for Study
A new research paper, titled “BEYOND THE ROSETTA STONE : U NIFICATION FORCES IN GENERALIZATION DYNAMICS”, introduces a novel approach to study this problem. Instead of analyzing large, pre-trained models as static entities, the researchers used a “Petri dish” methodology. They trained small Transformer models from scratch on specially designed synthetic multilingual datasets. This controlled environment allowed them to systematically vary data properties and observe the learning dynamics during pre-training.
The core insight from this work is the identification of a crucial learning phase where a model develops either ‘separate’ or ‘unified’ representations for the same facts across different languages. The study found that this ‘unification’ of representations is absolutely essential for successful cross-lingual knowledge transfer. If the model treats a fact expressed in English as fundamentally different from the same fact expressed in Spanish, it will struggle to transfer knowledge between the two.
What Drives Unification and Separation?
The researchers pinpointed two primary factors influencing this unification:
1. Parallel Data: As expected, expressing the same information in multiple languages (parallel data) helps models develop shared, cross-lingual representations. This is similar to how providing multiple paraphrases of a fact in a single language can improve recall.
2. Monolingual Data Properties: More surprisingly, the characteristics of the monolingual (non-parallel) portion of the dataset can actually encourage representational separation. This happens when the language of an example is both easy for the model to identify and provides a useful, albeit spurious, hint for predicting the answer. For instance, if a dataset contains more facts about Spanish cities in Spanish texts and English cities in English texts, the language itself becomes a ‘shortcut’ for the model to predict the city, rather than learning the underlying fact independently of language.
The study demonstrated this by manipulating two aspects: language informativeness (how much the language itself predicts the answer) and language extractability (how easy it is for the model to recognize the language, e.g., through distinct vocabulary or casing). When the language feature was highly informative and easy to extract, models tended to form separate representations, leading to poor cross-lingual transfer.
Predicting Performance with a Unification Score
To quantify this phenomenon, the researchers introduced a “unification score.” This metric measures how similarly the model treats semantically equivalent data points from different languages. They found a strong correlation between this unification score and cross-lingual factual recall accuracy. This means the score can be used to predict how well a model will generalize across languages, potentially even replacing the need for extensive cross-lingual test sets during model selection.
The findings were not limited to their small synthetic models. The researchers extended their analysis to larger language models, specifically the Gemma family, using the ECLeKTic dataset. They found that their unification metric correlated with cross-lingual factual accuracy in Gemma models, and that vocabulary overlap between languages (which makes the language feature harder to extract) also significantly predicted translation accuracy.
Also Read:
- Unpacking Prompt Sensitivity: A Deep Dive into LLM Robustness
- Lean Language Models Master Reasoning and Retrieval for Private AI Applications
Implications for Improving LLMs
This research offers valuable insights into the training dynamics of multilingual LLMs. It suggests that improving cross-lingual transfer isn’t just about adding more parallel data. It also involves carefully constructing monolingual datasets to discourage the formation of language-specific knowledge silos. This could mean balancing attribute frequencies across languages or making linguistic differences less obvious to the model during pre-training.
While language can be a useful signal for certain tasks, this work highlights how an overly strong or easily extractable language feature can impede true cross-lingual generalization. The study’s methodology and findings pave the way for new directions in improving how LLMs acquire and transfer knowledge across diverse languages. You can read the full paper here: Beyond the Rosetta Stone: Unification Forces in Generalization Dynamics.
