On the Similarity of Circuits across Languages: a Case Study on the Subject-verb Agreement Task (2410.06496v1)

Abstract: Several algorithms implemented by LLMs have recently been successfully reversed-engineered. However, these findings have been concentrated on specific tasks and models, leaving it unclear how universal circuits are across different settings. In this paper, we study the circuits implemented by Gemma 2B for solving the subject-verb agreement task across two different languages, English and Spanish. We discover that both circuits are highly consistent, being mainly driven by a particular attention head writing a `subject number' signal to the last residual stream, which is read by a small set of neurons in the final MLPs. Notably, this subject number signal is represented as a direction in the residual stream space, and is language-independent. We demonstrate that this direction has a causal effect on the model predictions, effectively flipping the Spanish predicted verb number by intervening with the direction found in English. Finally, we present evidence of similar behavior in other models within the Gemma 1 and Gemma 2 families.

• The paper demonstrates that language model circuits for subject-verb agreement show consistent behavior across English and Spanish.
• It employs methods like activation patching and attention pattern analysis to pinpoint the critical L13H7 attention head encoding subject number.
• The findings support the design of robust multilingual models by exploiting shared circuitry to minimize language-specific tuning.

Essay on the Similarity of Circuits across Languages in LLMs

The paper "On the Similarity of Circuits across Languages: A Case Study on the Subject-verb Agreement Task" by Javier Ferrando and Marta R. Costa-jussà tackles a nuanced aspect of mechanistic interpretability within LLMs. Specifically, it explores the internal mechanisms employed by the Gemma 2B LLM when solving the subject-verb agreement (SVA) task across English and Spanish. This study is particularly pertinent in assessing the robustness and universality of model behavior across linguistic boundaries.

Methodology and Findings

The research draws on a robust suite of techniques prevalent in circuit analysis, including direct logit attribution, activation patching, and attention pattern analysis. These methods collectively aim to identify the components and pathways—referred to as circuits—responsible for specific linguistic behaviors within LLMs.

One of the primary revelations of this study is the consistent circuitry observed in the Gemma 2B model across both English and Spanish. A particular attention head (L13H7) emerges as pivotal in encoding a 'subject number' signal in a specific subspace of the residual stream. Remarkably, this signal's direction is language-independent, and its manipulation can causally influence model predictions, allowing for the intentional flipping of verb numbers across languages. Notably, the research uncovers that this signal is primarily read by a small complement of neurons in the model's later layers, highlighting a directional influence on predictions.

Implications

The insights gained from this research extend beyond a mere dissection of model internals; they provide evidence supporting the transferability of identified circuits across languages. This understanding is invaluable as it suggests a more generalized architectural behavior inherent in such models. From a theoretical standpoint, these findings could inform future model architectures to enhance cross-lingual capabilities. Practically, this work might assist in the development of LLMs that are more robust and reliable in multilingual settings, potentially reducing the need for language-specific fine-tuning in some applications.

Numerical Results and Validation

An intriguing aspect of the study is the rigor applied in validation across different models within the Gemma family. The circuits identified in Gemma 2B were replicated in both Gemma 7B and Gemma 2 2B, reinforcing the generality and robustness of the findings. Activation patching demonstrated similar behaviors, especially the presence of a singular influential attention head responsible for encoding subject number, lending credibility to the study's claims.

Future Directions

The paper opens several avenues for future research:

1. Broader Language Scope: Expanding this analysis to other languages, particularly those structurally distinct from English and Spanish, would further validate the universality of the identified circuits.
2. Model Variability: Investigating whether these circuits manifest similarly in other LLM architectures could provide insights into model-independent features of language processing.
3. Enhanced Multilingual Models: Leveraging the understanding of such shared circuits could guide the design of more efficient multilingual models, optimizing resource usage by utilizing fewer language-specific parameters.

Conclusion

This paper provides a meticulous examination of how LLMs internally handle syntactic agreement tasks, revealing shared circuits across languages. The systematic approach and thorough validation exemplify a high level of research rigor, contributing meaningfully to the field of AI interpretability. As AI systems continue to integrate into global applications, understanding the universality and specificity of their internal workings remains a critical pursuit.