TL;DR
This paper uses sparse autoencoders (SAEs) to probe how large language models (LLMs) process multilingual inputs, revealing that LLMs have a precise understanding of prompt structure, with headings playing a crucial role. The authors propose two methods, English Heading Prompts and Amplifying TF Features, to improve performance in low-resource languages, achieving up to 3.7% gains.
What Problem Does This Paper Address?
Large language models (LLMs) have shown impressive multilingual capabilities, but the underlying mechanisms remain largely unclear. This lack of understanding leads to performance gaps across different languages, especially for low-resource ones. While approaches like cross-lingual in-context learning (X-ICL) exist, they often focus on transfer rather than understanding intrinsic differences. This paper aims to shed light on these differences by investigating how LLMs internally represent and process multilingual prompts, using mechanistic interpretability techniques.
Methods
The core of this work is the application of Sparse Autoencoders (SAEs) to analyze the internal representations of LLMs. SAEs are unsupervised models that decompose dense embeddings into sparse, interpretable features, each potentially corresponding to a human-understandable concept.
The authors introduce two novel types of SAE features:
- Task Instruction-Focused (TF) features: These features primarily activate on the task instruction part of a prompt.
- Heading-Focused (HF) features: These features specifically activate on headings within the prompt (e.g., "Question," "Choices," "Answer").
By analyzing the patterns of these features across different languages, the researchers gain insights into how LLMs differentiate prompt segments. Based on these findings, they propose two practical methods to improve performance:
- English Heading Prompts (EHP): Replacing prompt headings with English, leveraging LLMs' English-centric nature.
- Amplifying TF Features (ATF): Steering specific TF features derived from English to enhance performance in other languages.
Datasets
The study utilizes multilingual versions of two established benchmarks:
- Multilingual-MMLU (STEM): Assesses knowledge across STEM subjects.
- Multilingual-ARC-Challenge: Focuses on reasoning based on grade-school science knowledge.
These datasets were chosen because they are widely used and contain multiple-choice questions, allowing for direct evaluation of performance. The authors selected ten languages for their experiments, grouping them into high- and low-performing categories based on their downstream performance.
Key Insights and Ideas
- LLMs have a precise understanding of prompt structure: SAEs provide concrete evidence that LLMs can distinguish between task instructions, test queries, and headings, even across different languages. This suggests a fundamental, shared understanding of prompt organization.
- Headings are surprisingly important: The existence of dedicated HF features indicates that headings play a distinct and significant role in LLM processing, beyond just being text.
- Low-performing languages have fewer TF features: A key discrepancy found is that languages with lower performance exhibit a relative deficiency in TF features compared to high-performing languages. This suggests TF features are crucial for robust language processing.
- English as a pivot language is key: The effectiveness of EHP suggests that LLMs leverage English as a primary "pivot" language, making English headings more impactful for processing.
Key Results
- Improved performance: The proposed methods, EHP and ATF, consistently improve performance, especially for low-performing languages.
- EHP alone shows significant gains.
- Combining EHP and ATF achieves average performance gains of up to 3.70% on m-MMLU and 1.94% on m-ARC for low-performing languages.
- Correlation between TF features and performance: The number of TF features strongly correlates with overall language performance, reinforcing their importance.
- Universality of features, but varying activation: While TF and HF features appear to be generally universal across languages, their activation frequency can differ, particularly being lower for low-performing languages.
Limitations and Future Work
- Label Bias: The paper notes the identification of features that may contribute to label bias (e.g., activating on headings and alphabetic choices), suggesting that steering these features could be a promising direction for future work.
- Lenient HF Features: The authors suggest that there might be more HF features than measured if "lenient" ones (those that activate on headings but less precisely) are also considered, indicating potential for further exploration.
- Further investigation of TF/HF features: The paper suggests that a deeper dive into the nuances of TF and HF features would be a valuable area for future research.
Summary updated