PaperlayerView exact revision on arXiv

Source-linked AI summary

The Semantic Hub Hypothesis: Language Models Share Semantic Representations Across Languages and Modalities

Zhaofeng Wu, Xinyan Velocity Yu, Dani Yogatama, Jiasen Lu, Yoon Kim

arXiv:2411.04986v3cs.CL

TL;DR

The paper asks how language models process diverse languages and modalities, and proposes that they use a shared semantic representation space. It tests representations and intervenes in that space, finding cross-type similarities and predictable effects on outputs.

  • Problem

    Language models process diverse languages and modalities, raising the question of how one set of parameters handles these distinct data types when they can share semantic concepts.

    Source: §1 · p. 2

  • Method

    The paper tests whether semantically similar inputs across data types share intermediate representations, probes those representations through the dominant data type, and intervenes in the shared space.

    Source: §1 · p. 2

  • Results

    Across languages, arithmetic, code, formal semantic structures, visual inputs, and audio, the authors find evidence of shared representations, and interventions through the dominant language predictably change outputs for other data types.

    Source: §1 · p. 2, §3 · p. 4, §7 · p. 14

  • Takeaways & Limitations

    The findings support the semantic hub hypothesis and suggest that models represent semantically similar inputs from different modalities near one another.

    Source: §5 · p. 13

    Formal semantics receives no intervention experiment because a natural-language prefix usually licenses only one possible thematic role to follow.

    Source: §4 · p. 12

Abstract

from arXiv · show

Modern language models 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.

1 INTRODUCTION

The paper asks how language models process heterogeneous inputs and proposes that they use a shared semantic representation space. It presents evidence that this space spans data types, is scaffolded by the dominant language, and affects model behavior.

1 INTRODUCTION

  • The semantic hub hypothesis proposes that models map surface-distinct inputs with shared concepts into a common space for computation.The paper draws an analogy to the human brain’s transmodal semantic hub, which integrates semantic information from modality-specific spokes.Source: §1 · p. 2
  • Representations for semantically similar inputs across languages, arithmetic, code, formal structures, and multimodal inputs are close in intermediate layers.The paper also reports that these representations can be interpreted to an extent through the model’s dominant data type, such as English for an English-dominant model processing Chinese.Source: §1 · p. 2
  • Intervening in the shared space through the dominant language predictably changes outputs for other data types, indicating that the space affects model behavior.The authors contrast this with a merely vestigial byproduct of training on broad data.Source: §1 · p. 2
  • Unlike work aligning representations from separately trained models, this paper studies a single model trained on multiple data types and finds shared processing without explicit alignment transformations.The authors suggest these findings may inform interpretation and future efforts to control models.Source: §1 · p. 2

2 THE SEMANTIC HUB HYPOTHESIS

The paper hypothesizes that a model maps semantically similar inputs from different data types into a shared intermediate representation space. It tests relative similarity between representations and whether the shared space is anchored in the model’s dominant data type.

2 THE SEMANTIC HUB HYPOTHESIS

  • The hypothesis is that semantically similar inputs from different data types map to similar representations in a shared, modality-agnostic space.The model is described as encoding inputs in early layers, computing in middle layers, and verbalizing outputs in end layers.Source: §2 · p. 3
  • The test compares whether a representation is more similar to a semantically related input than to an unrelated sequence.The paper uses relative rather than absolute similarity because absolute scores in high-dimensional spaces are difficult to interpret.Source: §2 · p. 3
  • For models with a dominant training data type, the hypothesis predicts that shared representations are closer to a corresponding continuation in that type than to one in a non-dominant type.The example is an English-dominant model encoding a Chinese prefix near the English continuation “write” rather than its Chinese translation.Source: §2 · p. 3, Figure 2 · p. 3

2.1 METHOD: TESTING THE SEMANTIC HUB HYPOTHESIS

  • The method tests related-input similarity with intermediate hidden states and cosine similarity, and probes dominant-type anchoring with the logit lens.The logit lens applies the output operation to intermediate layers; tests generally use short verbalizations identifiable by a single token, while some multimodal cases test only related-input similarity.Source: §2.1 · pp. 3–4

3 EVIDENCE OF A SEMANTIC HUB

Across languages, arithmetic, code, formal semantics, and visual and audio inputs, the experiments find evidence of shared representations in intermediate model layers. In several settings, logit-lens analyses also connect these representations to English or another dominant-language concept.

3 EVIDENCE OF A SEMANTIC HUB

  • Across languages, arithmetic, code, formal structures, images, and audio, the experiments find evidence of a shared representation space.Source: §3 · p. 4

3.1 MULTILINGUAL

  • Translation pairs exceed nonmatching-similarity baselines most in middle layers; dominant-language anchoring (except in BLOOM) gives way to the input language in final layers.For English-dominant models processing Chinese, English is more salient in intermediate layers; the trend flips for a Chinese-dominant model.Source: §3.1 · pp. 4–5, Figure 4 · p. 5

3.2 ARITHMETIC

  • Arithmetic expressions with the same value but distinct surface forms have higher similarity than different-value expressions, with translation similarity above baseline in early-middle layers.Representations also pass through the space occupied by English number words before numeral tokens dominate later.Source: §3.2 · p. 6

3.3 CODE

  • Code representations align with English semantic continuations in intermediate layers, then converge on syntax-constrained Python tokens in final layers.Examples include English “except” for a loop condition and “and” for a list continuation; constructing semantically equivalent English–Python prefix pairs was difficult.Source: Figure 8 · p. 7, §3.3 · pp. 7–8, Figure 7 · p. 7

3.4 FORMAL SEMANTICS

  • COGS sentence–structure matches remain stronger than lexical and positional controls; for Llama-3, prior-adjusted thematic-role tokens align with intermediate states, unlike in Llama-2.Source: Figure 10 · p. 8, §3.4 · p. 9

3.5 VISUAL INPUT

  • Image-caption pairs are more similar than nonmatching pairs, and color-patch states sometimes align with color words in intermediate layers.The image-caption similarity emerges under autoregressive training without an explicit cross-modal similarity objective; the study does not test dominant-data-type anchoring for visual inputs.Source: Figure 11 · p. 9, §3.5 · p. 10

3.6 AUDIO

  • On 1,000 VGGSound samples, matching audio-label pairs have more similar representations than nonmatching pairs in intermediate layers.The audio experiments do not investigate dominant-data-type anchoring.Source: §3.6 · p. 10

4 INTERVENING IN THE SEMANTIC HUB

The section tests whether the semantic hub affects behavior by intervening in dominant-language representations while models process other data types. The interventions produce predictable changes across languages, arithmetic, code, images, and audio, though some settings limit what can be tested.

Intervention results

  • English-space interventions produce predictable output changes in non-dominant data types, supporting a causal role for the semantic hub.Figure 12 summarizes steering for arithmetic, code, images, and audio.Source: §4 · p. 10, Figure 12 · p. 12

Multilingual

  • English and text-language sentiment steering are similarly effective for Llama-3, usually without statistically significant decreases in fluency or relevance.The reported comparison covers Spanish and Chinese; Llama-2 shows similar trends.Source: §4 · p. 11, Table 1 · p. 11

Arithmetic and code

  • As intervention strength increased, arithmetic steering produced the expected decremented number followed by “+1” in up to >90% of instances, while code steering produced “range(0, end)” in up to 67%.For 846 addition expressions, arithmetic intervention was applied at layer 25 for Llama-3 and layer 30 for Llama-2; the range intervention added a vector at layer 17.Source: §4 · pp. 11–12

Visual input

  • Above 80% of visual color interventions succeeded when the model had to report both the changed color and the untouched color correctly.The two-color setup controls for the possibility that the injected word simply biases the answer lexically.Source: §4 · p. 12

Audio

  • For animal sounds, stronger interventions made SALMONN more likely to predict in the steered mammal or non-mammal direction.The experiment used 1,000 single-word animal sounds and measured normalized “Yes” probability.Source: §4 · p. 13

Intervention design and scope

  • Interventions vary by data type, and some tests are constrained: formal-semantics prefixes license only one thematic role, while Chameleon-7B cannot answer tested higher-reasoning color questions unaided.The authors also note that interventions are not guaranteed to produce sensible or correct outputs.Source: §4 · pp. 10, 12, 13

5 DISCUSSION

The discussion interprets the findings as evidence that models use a shared semantic space across data types, often organized around a dominant language. It also cautions that reliance on the dominant data type may have costs.

Discussion

  • Across languages and modalities, the findings support a semantic hub in which semantically similar inputs are represented near one another.The authors connect this proposal to the hypothesized semantic hub in the human brain and to findings that multilingual models often reason best in English.Source: §5 · p. 13

Risks of dominant-language reliance

  • Reliance on a dominant language could inherit its training-data biases and potentially align ideologies across languages, harming inclusivity.The authors present this as a conjectured risk, not an established effect.Source: §5 · p. 13
  • Representing arithmetic through a dominant language may be undesirable given structural differences and linguistic variation in number expressions and bases.The discussion cites Danish number composition and the Kalam base-23 numeral system as examples.Source: §5 · p. 13

6 RELATED WORK

The paper distinguishes its focus on aligned representations within one multimodal model from prior work aligning representations across separately trained models. It also situates its analyses among research on layer dynamics and methods for interpreting hidden states.

Representation alignment between separately trained models

  • Unlike work mapping representations across separately trained models, this work studies a single static model whose data-type representations align without a transformation.The related work also notes prior cross-lingual embedding alignment and mappings between visual and text representations.Source: §6 · p. 14

Representation evolution throughout layers

  • The paper’s layer-by-layer analyses complement research examining how transformer representations evolve and relate to reasoning, factuality, knowledge, pruning, and early exiting.Other work has characterized representation changes algorithmically.Source: §6 · p. 14

Inspecting model hidden states

  • The authors use the logit lens as a simple, training-free interpretability method, while noting that more expressive trained alternatives carry overfitting risks.They identify testing the hypothesis with alternative methods as future work.Source: §6 · p. 14

7 CONCLUSION

The conclusion reports evidence for shared intermediate representations across data types and predictable behavior changes when those representations are intervened on through the model’s dominant language.

Conclusion

  • Evidence across multiple models and data types supports the semantic hub hypothesis, and dominant-language interventions lead to predictable behavior changes.The dominant language is usually English.Source: §7 · p. 14

A EXPERIMENTAL DETAILS FOR §3

These appendices specify tokenization, filtering, prompting, and dataset choices for the multilingual, code, and vision-language experiments.

A EXPERIMENTAL DETAILS FOR §3

  • Token-level logit-lens probabilities use whichever prefix-space configuration maximizes the probability, which can differ by setting.For example, after “+” in “a=b+c”, “c” is more likely than “_c”.Source: A EXPERIMENTAL DETAILS FOR §3 · p. 20

A.1 MULTILINGUAL

  • For Experiment 1, parallel sentences are placed in “{English Sentence} This represents” templates to reduce mid-sentence code-switching that could confound analysis.Other tested templates gave similar results, and GALE sentence pairs were retained only when both transcripts were nonempty.Source: §A.1 · p. 20
  • For Experiment 2, only positions with a valid Jieba-segmented Chinese word and a single-token English translation survive the aggressive tokenization filter.The example leaves only 今天 because its translation meets the single-token cutoff.Source: §A.1 · p. 20

A.2 CODE

  • For MBPP function-call arguments, the analysis retains 540 cases after excluding unsuitable calls, generic roles, identical role and argument names, and unlocatable token positions.The dataset combines all splits for 974 programs.Source: §A.2 · p. 21

A.3 VISION-LANGUAGE

  • For vision-language experiments, image patches use task-specific prompts, and a patch corresponds to a segmentation label when that segment covers more than half its pixels.Caption and segmentation experiments use MSCOCO 2017; segmentation uses its panoptic labels.Source: §A.3 · p. 21

B EXPERIMENTAL DETAILS FOR §4

The intervention experiments use argmax decoding and mostly rely on the pyvene library.

B EXPERIMENTAL DETAILS FOR §4

B.1 MULTILINGUAL

The multilingual interventions steer sentiment across languages, with Llama-2 results reported alongside the Llama-3 results. Image-patch representations also better match corresponding English words than non-corresponding labels.

B.1 MULTILINGUAL

  • For each language, the study samples 1,000 training instances and uses language instructions to mitigate Llama models’ tendency to switch back to English.Sentiment, fluency, and relevance are evaluated with multilingual models; ActAdd uses layer 17, a scaling coefficient of 5, and generation settings adopted without tuning.Source: §B.1 · p. 21
  • The authors report similar trends when using Mistral-Nemo-Base for fluency evaluation.They also tried this model in addition to Llama-3.1-70B.Source: §B.1 · p. 21
  • Cross-lingual steering is consistently successful, sometimes exceeding monolingual steering, without substantial damage to fluency or relevance.Table 2 reports mean sentiment, disfluency (perplexity), and relevance, with standard deviations across 10 seeds.Source: Table 2 · p. 22
  • Image representations better match semantically corresponding English words than baseline words without correspondence to the patch.Figure 13 compares logit-lens probabilities for caption nouns and segmentation labels against non-corresponding baselines.Source: Figure 13 · p. 22
  • Llama-2 intervention results show trends similar to those reported for Llama-3.Table 2 reports Llama-2 steering results.Source: §B.1 · p. 22, Table 2 · p. 22

C TOKEN-LEVEL MULTIMODAL EXPERIMENTS USING THE LOGIT LENS

The multimodal experiments use a token-level logit-lens-style test of representation alignment.

C TOKEN-LEVEL MULTIMODAL EXPERIMENTS USING THE LOGIT LENS

  • At individual tokens such as image patches, the test measures logit-lens alignment between a token’s hidden representation and its natural-language description token.The authors note that they do not perform sequence-level logit-lens analysis for multimodal experiments.Source: C TOKEN-LEVEL MULTIMODAL EXPERIMENTS… · p. 22

C.1 VISUAL INPUT.

Matching image captions align better with image-patch representations than unmatched captions, while segmentation-label alignment differs between LLaVA and Chameleon; audio states are closer to corresponding label words, but only position-agnostic analysis is possible.

C.1 VISUAL INPUT.

  • Matching captions align better with image-patch representations than unmatched captions for both LLaVA and Chameleon.Alignment scores sum logit-lens probabilities for caption nouns across patches and layers.Source: §C.1 · p. 22
  • LLaVA patches align much better to corresponding segmentation labels than random labels, while Chameleon shows only a nonsignificant effect in one middle layer.A label corresponds to a patch when it covers more than half of its pixels.Source: §C.1 · p. 23

C.2 AUDIO

  • Audio hidden states are closer to semantically corresponding English label words than to a random label, as shown by SALMONN’s logit-lens probabilities.The captioning study uses 1,000 samples; label-word probabilities provide a lower bound because some label words may not be represented in the audio states.Source: §C.2 · p. 23, Figure 14 · p. 23
  • SALMONN evaluations are position-agnostic, preventing fine-grained patch-style analysis such as alignment with segmentation labels.Unlike image patches, audio inputs have no correspondence between individual input positions and locations in the audio.Source: §C.2 · p. 23
Loading 2411.04986v3…