Source-linked AI summary
Large Language Models Share Representations of Latent Grammatical Concepts Across Typologically Diverse Languages
Jannik Brinkmann, Chris Wendler, Christian Bartelt, Aaron Mueller
TL;DR
The paper asks whether multilingual language models share representations of morphosyntactic concepts across languages. Using sparse autoencoders, feature overlap analysis, causal ablations, and translation interventions, it finds shared multilingual grammatical representations, including in models trained primarily on English, while steering remains imperfect.
Problem
The paper investigates whether multilingual language models encode morphosyntactic concepts such as number, gender, and tense in shared representations rather than language-specific ones.
Method
The authors train sparse autoencoders on Llama-3-8B and Aya-23-8B, identify multilingual features, test their causal importance, and intervene on them during machine translation.
Results
Up to 50% feature overlap occurs for the same grammatical concepts across typologically diverse languages, while ablating multilingual features reduces classifier performance to 61% for Llama 3.
Takeaways & Limitations
Even models trained primarily on English can develop cross-lingual grammatical abstractions, suggesting their internal lingua franca may consist of concepts rather than English words.
Takeaways & Limitations
Extreme feature clamping can produce nonsensical generations, and optimally steering model behavior remains an open problem.
Abstract
from arXiv · showhide
Human bilinguals often use similar brain regions to process multiple languages, depending on when they learned their second language and their proficiency. In large language models (LLMs), how are multiple languages learned and encoded? In this work, we explore the extent to which LLMs share representations of morphsyntactic concepts such as grammatical number, gender, and tense across languages. We train sparse autoencoders on Llama-3-8B and Aya-23-8B, and demonstrate that abstract grammatical concepts are often encoded in feature directions shared across many languages. We use causal interventions to verify the multilingual nature of these representations; specifically, we show that ablating only multilingual features decreases classifier performance to near-chance across languages. We then use these features to precisely modify model behavior in a machine translation task; this demonstrates both the generality and selectivity of these feature's roles in the network. Our findings suggest that even models trained predominantly on English data can develop robust, cross-lingual abstractions of morphosyntactic concepts.
1 Introduction
The paper asks whether multilingual language models develop shared representations of grammatical concepts across languages, especially when trained predominantly on English. Using sparse autoencoders, it investigates multilingual morphosyntactic features and their causal role in model behavior.
- Human bilinguals may process syntax in similar brain regions across languages, depending on second-language learning age and proficiency.
- Parameter-efficient multilingual models could represent grammatical concepts such as number with shared features rather than language-specific representations.
- Many high-performing multilingual language models are trained on over 90% English text, motivating questions about their strong non-English performance.
- The paper hypothesizes that generalizable abstractions enable more efficient learning of new languages, including through cross-lingual transfer.
- Sparse autoencoder features support more precise causal interventions than neuron-level analysis, helping distinguish targeted concepts from spurious related features.
- The authors train sparse autoencoders on Llama-3-8B and Aya-3-8B to locate multilingual features for morphosyntactic concepts and validate their role in generation.
2 Background
The background motivates sparse autoencoders as a way to disentangle distributed, polysemantic model representations and support causal analysis of linguistic features. It also introduces probing classifiers and attribution patching for locating and evaluating morphosyntactic representations.
- Sparse autoencoders decompose activation vectors into interpretable feature directions, addressing the many-to-many relationship between neurons and concepts.
- Single neurons are often polysemantic, activating for seemingly unrelated concepts such as citations, dialogue, HTTP requests, and Korean text.
- Gated sparse autoencoders separate feature selection from activation-magnitude estimation to reduce reconstruction biases introduced by L1 regularization.
- Rewriting model hidden states as sparse autoencoder features enables gradient-based attribution patching to identify influential components.
- Activation patching estimates causal importance by replacing a component’s activation on one input with its activation from a contrasting input.
- Attribution patching approximates causal contribution using the gradient of a target metric multiplied by the change in component activations.
- Linear probing classifiers map frozen intermediate activations to morphosyntactic labels, allowing the authors to quantify cross-lingual feature sharing.
3 Multilingual Features
The study measures whether morphosyntactic concepts are encoded by features shared across languages and tests whether these features causally support classifier predictions. Across Llama-3-8B and Aya-23-8B, multilingual features show substantial overlap, interpretable activation patterns, and concentrated causal importance.
- Measuring multilingual features: Attribution patching selects the top 32 causally influential SAE features for each language and concept, whose overlap is then measured across languages.The analysis targets residual-stream representations and computes cross-lingual overlap within each morphosyntactic concept.
- Cross-lingual overlap: Up to 50% of features identified for the same grammatical concepts overlap across typologically diverse languages, with similar patterns in Llama-3-8B and Aya-23-8B.Feature 22860 is among the most influential features across all 15 languages that inflect for grammatical gender, while overlap is greater for concepts present in more languages.
- Causal validation: Ablating multilingual features reduces probing-classifier performance across languages, indicating that classifiers rely on shared representations to detect morphosyntactic concepts.The intervention compares monolingual, multilingual, and massively multilingual feature sets.
- Causal validation: 64% average performance remains after ablating only the most multilingual features, versus 61% after ablating all multilingual features despite removing four times as many features.The most multilingual features therefore account for most of the probes’ behavior in Llama 3.
- Feature specificity: 13.9% ± 10.7% is the mean overlap between multilingual feature sets for different concept-value pairs, with 40.6% overlap for singular number and masculine gender.The higher pairwise overlap is consistent with concepts that are jointly realized morphologically.
- Interpretable representations: Selected sparse features exhibit interpretable cross-lingual activation patterns for concepts such as plural nouns and past tense.Manual inspection finds many features that are intuitively meaningful across languages.
4 Demonstrating Functional Selectivity: A Case Study in Machine Translation
The authors use sparse autoencoder features to steer grammatical concepts during translation, testing whether interventions can flip the intended concept while preserving unrelated concepts. Across controlled and Flores-101 examples, interventions are often effective and selective, but some require tuning and produce failures.
- Intervention setup: The study uses feature steering to clamp selected multilingual features to counterfactual activations during machine translation.The intervention targets tense, gender, or number features in English, French, German, and selected Turkish settings.
- Evaluation: The evaluation uses minimal-pair grammatical inputs and measures whether intervened translations express the counterfactual concept.Efficacy is computed from probe-label flips between translations generated with and without intervention.
- Efficacy: A single feature often changes the model’s prediction toward the intended grammatical value, although it is insufficient for some concepts.Figure 5 aggregates success rates over 64 examples and translation directions.
- Selectivity: Interventions are generally selective, changing the targeted grammatical concept without altering predictions for other concepts.Selectivity is measured as the proportion of translation pairs where no other concept probe changes its prediction.
- Qualitative validation: The authors also test qualitative generalization on Flores-101 because the controlled dataset is neither naturalistic nor diverse.Flores-101 supplies aligned translations across 101 languages.
- Limitations: Some interventions fail or produce degenerate generations, requiring concept- and language-specific tuning of activation values.The limitation is associated with scaling activations beyond naturally observed values and, potentially, intervening on multiple features.
5 Related Work
Related work studies multilingual representations through language-model training choices, behavioral transfer, and neuron-level analyses. This paper positions sparse autoencoder features as a finer-grained basis for examining shared morphosyntactic representations.
- Multilingual language modeling: Multilingual models differ in pretraining balance: XLM-R, M2M, mT5, and Aya use balanced multilingual corpora, whereas Llama 3 is predominantly English-trained.Llama 3 is described as trained on a distribution with over 90% English text.
- Feature-based analysis: Sparse autoencoder features provide a more fine-grained unit for analyzing and intervening on distributed model representations.The paper uses feature-based analysis to address the polysemanticity and causal imprecision associated with neuron-level interventions.
- Interpreting multilingual models: A central question is whether language models develop universal concept representations that are disentangled from particular languages.Prior work examines cross-lingual transfer and similarity across models trained on typologically related languages.
- Neuron-level analyses: Neuron-level studies find that probes can read some concepts from shared neurons across languages and that individual neurons can respond consistently to syntactic phenomena.These studies include mBERT, XLM-R, and cross-lingual probe analyses.
6 Discussion and Conclusion
The discussion argues that large-scale, even English-imbalanced pretraining can produce cross-lingual grammatical abstractions, while emphasizing a trade-off between shared representations and language-specific bias. These representations may also help explain why interventions in one language can generalize to others.
- Discussion: Large-scale pretraining with imbalanced corpora induces cross-lingually generalizable grammatical abstractions, comparable in this respect to balanced pretraining.The authors interpret this as evidence that an internal lingua franca may consist of concepts rather than English words.
- Trade-offs: Shared representations are parameter-efficient but may encode biases when languages assign different values or distributions to grammatical concepts.Examples include differing numbers of grammatical cases and differing noun genders across languages.
- Trade-offs: More language-specific features could reduce such bias but would require more parameters, leaving the optimal degree of generalization unresolved.The paper frames this as a balance between shared and language-specific representations.
- Implications: Cross-language concept sharing may help explain why preference tuning or model-editing interventions applied in one language can generalize to others.The authors connect this implication to multilingual downstream NLP tasks, including toxicity mitigation.
Limitations
The study identifies several limitations affecting interpretation and steering: SAE features may miss nonlinear concepts, human interpretations may be biased, and representation does not always predict generation.
- Steering with SAEs: Extreme SAE steering values can produce nonsensical generations, so optimal behavioral steering remains unresolved.The authors also note that SAEs can underperform alternatives for fine-grained steering.
- Understanding versus generation: High causal effects on linear probes do not necessarily translate into comparable effects on model generation.This reflects a distinction between representing a concept and using it during generation.
- Feature interpretation: Human explanations of LM features may be biased because models may use concept spaces distinct from humans.Such bias can lead to suboptimal predictions of when a feature activates.
- Non-linear features: SAEs may fail to recover concepts encoded nonlinearly because they preferentially recover linearly encoded features.They are generally better suited to binary relations than complex arrangements such as circular representations.
A Implementation Details
The implementation uses activation patching and sparse autoencoders to analyze model components and representations, with integrated-gradients attribution chosen as a more accurate approximation and gated losses used for SAE training.
- C Attribution Patching: Activation patching replaces a component’s clean-prompt activation with its value from a reference prompt where the behavior is absent.The method estimates the counterfactual impact of that replacement on model behavior.
- C Attribution Patching: Attribution patching approximates activation-patching contributions with a first-order Taylor expansion, avoiding one forward pass per component.Exact activation patching scales as O(N) forward passes for N components.
- C Attribution Patching: The experiments use an integrated-gradients attribution method that is more expensive but more accurate than the standard approximation.Its additional cost scales linearly in K, with K = 10 equally spaced interpolation values.
- D.1 Gated Sparse Autoencoders: Gated SAEs represent inputs with encoder and decoder weights, feature activations, and reconstructions, while using a sparsity-controlled training loss.The sparsity coefficient is λ > 0.
- D.1 Gated Sparse Autoencoders: The gated architecture separates selecting dictionary elements from estimating their activation magnitudes to reduce reconstruction bias from L1 regularization.The loss uses a frozen decoder reconstruction and a gated sparsity term.
- D.1 Gated Sparse Autoencoders: Table 1 reports the training parameters used for sparse autoencoders trained on Llama-3-8B and Aya-23-8B.The supplied table caption identifies the models and the parameter scope, but not individual values.
D.2 Training Parameters
The experiments assess SAE quality, multilingual feature sharing, and in-context translation across the two models, using language subsets and evaluation procedures described below.
- D.2 Training Parameters: SAE quality is evaluated by the fraction of language-model loss recovered when residual-stream activations are replaced by SAE reconstructions.This is an extrinsic metric and does not measure intrinsic feature interpretability without ground-truth features.
- E Features Across Languages: Top-32 features for each grammatical concept are assessed for strong influence across languages.The number of languages sharing influential features roughly follows a power-law distribution.
- F Translation Experiments: Translation is evaluated on Flores-101 using SacreBLEU after prompting each model with two ground-truth exemplar translations.Results are presented for Llama-3-8B and Aya-23-8B.
- F Translation Experiments: The translation setup uses a two-shot English–German prompt format that was found empirically to work best in initial two-shot experiments.The prompt contains literal double backslashes rather than newline characters.
- F Translation Experiments: Figures 15 and 16 report SacreBLEU scores for Llama-3-8B and Aya-23-8B, respectively.The supplied captions identify the model associated with each figure but provide no score values.
F.3 Directional Translation Results
Directional translation results are reported separately for each translation direction, complementing the bidirectional language-pair averages presented elsewhere.
- F.3 Directional Translation Results: Figure 17 provides results for each translation direction separately, whereas Figure 5 averages language pairs in both directions.The example comparison is English–French and French–English.
G Counterfactual Dataset
The counterfactual dataset tests gender, number, and tense across English, French, and German. It evaluates whether intervening on one multilingual feature can flip model behavior during translation between languages.
- The dataset covers three morphosyntactic concepts—gender, number, and tense—and three languages: English, French, and German.Its format is inspired by CausalGym, with templates partly taken from that dataset.
- The evaluation considers translation in both directions across language pairs involving English, French, German, and Turkish.The listed directions include Turkish–French, Turkish–German, Turkish–English, and their reverses.
- For each concept, a source-language sentence is translated into another language while a single multilingual feature is intervened on to test flipping toward an alternative concept value.The figure describes examples such as changing present-tense behavior toward past-tense behavior during translation.
- Table 3 provides examples from the counterfactual dataset.