The Semantic Hub Hypothesis: Language Models Share Semantic Representations Across Languages and Modalities (2411.04986v3)

Abstract: Modern LLMs can process inputs across diverse languages and modalities. We hypothesize that models acquire this capability through learning a shared representation space across heterogeneous data types (e.g., different languages and modalities), which places semantically similar inputs near one another, even if they are from different modalities/languages. We term this the semantic hub hypothesis, following the hub-and-spoke model from neuroscience (Patterson et al., 2007) which posits that semantic knowledge in the human brain is organized through a transmodal semantic "hub" which integrates information from various modality-specific "spokes" regions. We first show that model representations for semantically equivalent inputs in different languages are similar in the intermediate layers, and that this space can be interpreted using the model's dominant pretraining language via the logit lens. This tendency extends to other data types, including arithmetic expressions, code, and visual/audio inputs. Interventions in the shared representation space in one data type also predictably affect model outputs in other data types, suggesting that this shared representations space is not simply a vestigial byproduct of large-scale training on broad data, but something that is actively utilized by the model during input processing.

• The paper provides empirical evidence that LLMs share a unified semantic hub across languages and modalities, demonstrated through cosine similarity analyses of hidden layer representations.
• Intervention experiments using Activation Addition (ActAdd) show that perturbing the shared space causally influences model outputs across diverse data types.
• The study uncovers a dominant English bias in models, highlighting implications for fine-tuning, bias mitigation, and improved cross-modal processing.

Semantic Hub Hypothesis: Shared Representation Across Languages and Modalities

The paper investigates the premise that LLMs and multimodal models acquire the capability to process diverse linguistic and non-linguistic data types through a unified representation framework, termed the "semantic hub." This model posits that, similar to the human brain's transmodal semantic hub, artificial models utilize a shared representation space where semantically equivalent inputs across different modalities, such as text in various languages, code, arithmetic expressions, and visual/audio inputs, are internally processed in a similar manner.

Key Contributions

1. Shared Representation Analysis: The authors utilize empirical tests to support the semantic hub hypothesis. They show that LLMs represent semantically similar inputs from different data types closely in their intermediate layers. This conclusion is drawn from evaluating the cosine similarity of hidden layer representations across inputs from multiple modalities and translations in different languages. The findings affirm that LLMs scaffold these shared spaces using the most prevalent pretraining data type, typically English, enabling a comprehensible interpretation of this space through a technique termed the "logit lens."
2. Numerical and Multimodal Results: The paper presents robust numerical results validating the hypothesis across various models and datasets. For instance, language dominant models, such as Llama-2 and Llama-3, show a clear middle-layer bias toward English when processing multilingual texts, even when the input is non-English. Similar trends emerged in the arithmetic, code, formal semantics, visual, and auditory data, where LMs manifested a shared representation in line with the semantic hub hypothesis.
3. Intervention Experiments: Further, the paper deploys intervention methods, including Activation Addition (ActAdd), to causally affect model outputs. By introducing semantic perturbations in unseen languages at intermediate layers, the authors reveal how interventions conducted in the shared space influence model performance and output. This empirical approach underscores the causal impacts of the shared representation space, substantiating its role as more than a mere byproduct of data amalgamation during model training.

Implications and Future Directions

The findings contribute new insights into the architectural design and functional modality of LLMs. By confirming a shared transmodal representation space, the research advances the understanding of LLM's internal mechanisms, suggesting models process and generalize semantic data analogously to hypothesized human cognitive processes. This paradigm carries practical implications for model interpretability, fine-tuning, and controlled manipulations via informed interventions.

Moreover, this work raises pertinent discussions about potential intrinsic biases in LLMs due to dominant language training data, hinting at the need for balanced multilingual corpora. By aligning arithmetic and non-linguistic expressions within a linguistic scaffold, the paper provokes considerations of computational efficiency versus accuracy trade-offs that merit further exploration.

In conclusion, the Semantic Hub Hypothesis posits a functional blueprint for LLMs, fostering a myriad of practical applications and foundational research in dissecting cross-modal and multilingual processing frameworks. It creates prospects to refine fine-tuning methodologies, enhance model efficiency, and mitigate biases, thus steering future developments in artificial intelligence towards more inclusive and nuanced understanding architectures.