PaperlayerView exact revision on arXiv

Source-linked AI summary

Understanding Language Prior of LVLMs by Contrasting Chain-of-Embedding

Lin Long, Changdae Oh, Seongheon Park, Sharon Li

arXiv:2509.23050v3cs.LGcs.AI

TL;DR

Prior input-output analyses of language prior do not reveal how visual and textual signals are integrated inside LVLMs. By contrasting layer-wise representations under visual and blind contexts, this paper identifies a Visual Integration Point and proposes Total Visual Integration; experiments find VIP consistently and TVI a reliable indicator of language prior.

  • Problem

    Prior analyses mainly probe model outputs and do not reveal how language prior manifests or is quantified within LVLMs.

    Source: §2 · p. 3

  • Method

    The framework contrasts layer-wise hidden representations under visual and blind contexts, using VIP to locate meaningful visual influence and TVI to quantify its cumulative strength.

    Source: §1 · p. 2, §3.1 · p. 3, §3.3 · p. 5

  • Results

    Across models and datasets, VIP emerges consistently, and TVI reliably indicates language-prior strength.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The framework provides an interpretable approach to understanding and quantifying language prior in LVLMs.

    Source: §6 · p. 10

    The method requires white-box access to hidden states and attention patterns, restricting applicability to open-weight models rather than commercial APIs or closed-source systems.

    Source: §6 · p. 10

Abstract

from arXiv · show

Large vision-language models (LVLMs) achieve strong performance on multimodal tasks, yet they often default to their language prior (LP) -- memorized textual patterns from pre-training while under-utilizing visual evidence. Prior analyses of LP mostly rely on input-output probing, which fails to reveal the internal mechanisms governing when and how vision influences model behavior. To address this gap, we present the first systematic analysis of language prior through the lens of chain-of-embedding, which examines the layer-wise representation dynamics within LVLMs. Our analysis reveals a universal phenomenon: each model exhibits a Visual Integration Point (VIP), a critical layer at which visual information begins to meaningfully reshape hidden representations and influence decoding for multimodal reasoning. Building on this observation, we introduce the Total Visual Integration (TVI) estimator, which aggregates representational discrepancy beyond the VIP to quantify how strongly visual query influences response generation. Across 60 model-dataset combinations spanning 10 contemporary LVLMs and 6 benchmarks, we demonstrate that VIP consistently emerges, and that TVI reliably predicts the strength of language prior. This offers a principled toolkit for diagnosing and understanding language prior in LVLMs.

1 INTRODUCTION

LVLMs can rely on memorized language patterns while underusing visual evidence, and output probing cannot reveal where that imbalance develops internally. The paper contrasts layer-wise representations with and without visual input to identify a Visual Integration Point (VIP), then uses post-VIP integration to estimate language prior (LP) through Total Visual Integration (TVI).

1 INTRODUCTION

  • Earlier LP analyses rely on input-output probing, which cannot expose latent signal integration or locate where LP begins to interfere with visual integration.Source: §1 · p. 1
  • The framework contrasts layer-wise embeddings for visual and vision-removed queries, using expected representation distance across vision-dependent and vision-independent data to locate VIP.Source: §1 · pp. 1–2, Figure 1 · p. 2
  • At VIP, visual input begins meaningfully influencing decoding: representation distance rises substantially for vision-dependent tasks but shifts less for vision-independent tasks.Source: §1 · p. 2
  • TVI aggregates post-VIP embedding distances to measure effective visual integration, with lower TVI corresponding to stronger language-prior reliance.Source: §1 · p. 2
  • Across 10 contemporary LVLMs and 6 datasets, VIP consistently emerges and TVI reliably indicates language-prior strength.The paper reports 60 model-dataset settings.Source: §1 · p. 2

2 PROBLEM STATEMENT

The problem statement asks whether LVLM language prior can be formally understood and quantified through internal states rather than output-only evaluation. It defines the data and model setting, then characterizes LP as reliance on learned textual patterns that the proposed approach aims to diagnose directly.

2 PROBLEM STATEMENT

  • The dataset consists of image-text query pairs sampled from a population distribution, with visual input and a natural-language prompt in each tuple.Source: §2 · p. 2
  • Vision-dependent examples require the associated image to answer the query, whereas vision-independent examples can be answered from text alone.Source: §2 · p. 3
  • An LVLM maps a multimodal query through decoder layers to a vocabulary probability distribution and selects the most likely answer.Source: §2 · p. 3
  • Language prior is an LVLM’s reliance on memorized textual patterns; unlike curated-dataset proxies, the proposed approach seeks direct, annotation-free, sample-wise diagnosis.Source: §2 · p. 3
  • The research question is whether internal representations can provide a formal, generalizable account of LP that output-focused evaluations do not supply.Source: §2 · p. 3

3 METHODOLOGY

The method contrasts layer-wise representations from visual-plus-text and text-only inputs to locate when vision begins influencing inference. It uses that transition, the Visual Integration Point (VIP), to define Total Visual Integration (TVI), a post-VIP measure associated with language-prior strength and visual-task performance.

3.1 CHAIN-OF-EMBEDDING AND REPRESENTATION DISTANCE

  • The chain-of-embedding is the sequence of last-token hidden-state representations across layers, contrasted between visual-plus-text and text-only inputs using a distance metric.The paper uses cosine distance by default, while noting that other distance functions can also be valid.Source: §3.1 · pp. 3–4

3.2 VISUAL INTEGRATION POINT HYPOTHESIS

  • Across Qwen2.5-VL-7B and Gemma-3-4B, a VIP marks where vision-dependent examples’ representation distances rise more sharply than vision-independent examples’.The transition appears around layers 18–20 in Qwen2.5-VL-7B and 20–22 in Gemma-3-4B, consistently across the examined datasets.Source: Figure 2 · p. 4, §3.2 · p. 5
  • After the VIP, visual-information use becomes more diverse across samples, while distance trajectories differ between models.Qwen2.5-VL-7B’s distance grows smoothly before peaking and declining; Gemma-3-4B’s remains flat early, then rises steeply after VIP.Source: §3.2 · p. 5

3.3 QUANTIFYING LANGUAGE PRIOR OF LVLMS THROUGH TOTAL VISUAL INTEGRATION

  • TVI averages representation distances across layers at and beyond the VIP, with higher scores indicating greater visual integration and lower scores corresponding to stronger language prior.The estimator contrasts visual-plus-text and text-only inputs.Source: §3.3 · p. 5
  • Post-VIP representation distances correlate strongly with correctness on VLind-Bench, whereas pre-VIP aggregation yields weak, less-significant correlations.The paper uses post-VIP aggregation for TVI.Source: §3.3 · p. 5, Table 1 · p. 5

4 EXTENDED EXPERIMENTS

Extended experiments test VIP and TVI across models, datasets, interventions, and analysis settings. They report consistent VIP emergence, TVI differences aligned with language-prior strength, and stronger TVI-correctness correlations than comparison measures.

VIP across models and datasets

  • A clear VIP appears across 60 settings spanning 10 LVLMs and 6 datasets, with larger distance increases for vision-dependent than vision-independent examples.The experiments include general VQA benchmarks and benchmarks designed for language-prior evaluation.Source: §4 · pp. 6–7

TVI and language-prior strength

  • ViLP, designed to induce stronger language prior, yields lower TVI than MMBench, which represents a weaker-language-prior setting.The paper interprets the lower and higher TVI values, respectively, as weaker and stronger visual integration.Source: §4 · p. 7, Figure 4 · p. 7

Interventional validation

  • Applying the PAI attention-correction intervention to Qwen2.5-VL-7B improves task performance and substantially increases TVI.Table 2 reports downstream performance and TVI before and after intervention.Source: §4 · p. 7, Table 2 · p. 7

Comparison with other measures

  • TVI shows stronger correlations with answer correctness than visual-attention and output-divergence approaches across the evaluated datasets and models.The authors note that attention can focus on irrelevant image regions and output-level discrepancy misses internal representation dynamics.Source: Table 3 · p. 7, §4 · p. 8

Distance-metric ablations

  • TVI remains a strong correctness indicator with L2 distance, but its effectiveness drops significantly when hidden states are projected with the logit lens.The authors suggest that the output-space projection can obscure intermediate cross-modal integration patterns.Source: §4 · p. 8, Table 4 · p. 8

Varying model scales

  • The Gemma-3 scale analysis reports VIP and dimension-normalized TVI for three model variants, and the authors say the observations support applicability across model scales.The supplied passages do not state specific scale-comparison values.Source: Figure 5 · p. 8, §4 · p. 9

Practical utility

  • During LLaVA instruction tuning, TVI was added as a regularization term alongside cross-entropy next-token loss; on MME with LLaVA-v1.5-7B, performance improved, indicating potential as a training regularizer.Source: §4 · p. 9, Table 5 · p. 9

5 THEORETICAL ANALYSIS

The theoretical analysis interprets representation divergence as a distributional discrepancy and derives bounds connecting representation distances to visual integration. It also describes when these quantities can be estimated from finite samples.

5 THEORETICAL ANALYSIS

  • Representation divergence can be interpreted as a relative distributional discrepancy between visual-dependent and text-only representations, expressed through KL divergences to a text-only density estimate.The text-only estimate is modeled as a Gaussian centered on blind representations.Source: §5 · p. 9
  • Theorem 5.2 bounds expected representation distances using empirical samples, with a hypothesis set of pseudo-dimension c and probability at least 1 − δ.The paper presents these bounds as useful when adapting a fixed model to improve effective visual integration.Source: §5 · p. 10
  • For unknown vision-dependent data, the bound's lower bound on visual integration can be increased by decreasing text-only representation distance and the H-divergence between groups.For an unknown mixture distribution, increasing the H-divergence broadens the effective range of the representation-distance bound.Source: §5 · p. 10

6 CONCLUSION

The conclusion presents chain-of-embedding contrasts and TVI as a framework for analyzing language prior, while identifying distribution shifts and white-box access as limitations. The related work covers LVLM perception and prior efforts to measure or mitigate language prior.

6 CONCLUSION

  • The framework contrasts visual and blind chain-of-embedding representations, identifies VIP, and uses TVI to estimate effective visual integration and language prior across models and tasks.The conclusion reports experiments across 9 LVLMs and 6 datasets, and says the framework provides consistent, interpretable signals.Source: §6 · p. 10
  • The method's reliability for estimating language prior still needs validation under realistic distribution shifts, which may alter representations and degrade downstream performance.It requires access to internal hidden states and attention patterns, restricting applicability to open-weight models rather than commercial APIs or closed-source systems.Source: §6 · p. 10

APPENDIX

  • The appendix figure captions divide complete experimental results among three model groups, each evaluated on six datasets.The groups are LLaVA-v1.5-7B, LLaVA-NeXT-Vicuna-7B, and LLaVA-OV-Qwen2-7B; Gemma-3-4B, SmolVLM, and Llama-3.2-11B-Vision; and Qwen2.5-VL-7B, Eagle2.5-8B, and InternVL3-8B.Source: Figure 6 · p. 17, Figure 7 · p. 18, Figure 8 · p. 19

A RELATED WORK

  • Related work describes late-fusion LVLMs, concerns about weak vision-centric performance, and studies of visual integration, language-prior evaluation, and mitigation.Prior mitigation approaches include contrastive decoding, inference-time attention reallocation, and training-time interventions.Source: A RELATED WORK · p. 20

B.1 NOTE ON VIP DETERMINATION

VIP is identified from the layer where empirical representation-distance divergence becomes persistently positive, though finite-sample estimates may fit poorly. An automatic estimator is offered for cases where manual curve inspection is impractical.

B.1 NOTE ON VIP DETERMINATION

  • Finite-sample representation-distance estimates can fit poorly, but a layer where empirical divergence exceeds a positive constant thereafter still emerges, and TVI based on it remains a strong language-prior indicator.The paper notes Figure 10 as an example of poor estimator fit.Source: §B.1 · p. 20

B.2 ALGORITHMIC ESTIMATION ON VISUAL INTEGRATION POINT

  • The automatic VIP rule compares deviation-normalized representation distance at a layer with a β-weighted historical average, using β = 2.0.It is intended for large-scale comparisons or automated pipelines where manual inspection is impractical.Source: §B.2 · p. 21
  • Automatically estimated VIPs are close to manually selected VIPs, mark where divergence curves begin separating, and yield TVI values whose correlations with downstream correctness are robust across datasets.Table 6 compares VIP detection methods using Spearman correlations between TVI and downstream correctness.Source: §B.2 · p. 21, Table 6 · p. 21

C IMPLEMENTATION DETAILS

The implementation evaluates LVLMs and benchmark datasets under specified inference and preprocessing settings, then reports a controlled comparison between irrelevant-image and vision-dependent inputs. TVI uses empirically selected model-specific VIPs and last-token hidden states.

C IMPLEMENTATION DETAILS

  • Evaluation uses official instruction-tuned checkpoints for 10 publicly available LVLMs, with generation temperature set to 0.The benchmarks comprise six datasets with binary or multiple-choice questions; MMMU samples are filtered for one image and closed-form output, and explanations are omitted.Source: C IMPLEMENTATION DETAILS · pp. 21–22
  • For VLind-Bench, annotated counterfactual statements are converted into binary yes/no questions using GPT-4o, and ViLP pairs consistently use image3 and answer3.Models answer zero-shot without reasoning steps; Table 7 provides dataset statistics.Source: C IMPLEMENTATION DETAILS · p. 22, Table 7 · p. 23
  • Average TVI is significantly lower on CommonsenseQA questions paired with irrelevant images than on standard VQA benchmarks, while Qwen2.5-VL-7B shows a reverse pre-VIP trend.The irrelevant-image baseline is vision-independent; MME, MMBench, and VLind-Bench form the vision-dependent comparison, and the authors retain agreement-based separation in the main text for better control of data attributes.Source: C IMPLEMENTATION DETAILS · pp. 22–23, Figure 10 · p. 23
  • Reported TVI values use empirically determined, model-specific VIPs, and representation distances are computed from the final generated token's hidden states.The cited VIPs range from layer 9 for LLaVA-v1.5-7B to layer 35 for Gemma-3-27B; manually selected VIPs are not necessarily optimal but are reported as effective for analysis.Source: C IMPLEMENTATION DETAILS · p. 23

D FULL RESULTS OF THE MAIN EXPERIMENT

The main experiments show a consistent VIP across evaluated models and datasets. Additional tests support the findings across different fusion architectures and larger model scales.

D FULL RESULTS OF THE MAIN EXPERIMENT

  • VIP consistently appears across the reported 9 LVLMs and 6 datasets, although its separation is less pronounced in SmolVLM.The authors suggest SmolVLM’s limited capacity may explain its less promising visual integration.Source: D FULL RESULTS OF THE MAIN EXPERIMENT · p. 24
  • Results corroborate the conclusions for cross-attention and Q-Former fusion architectures and for larger-scale models.The additional models include Llama-3.2-11B-Vision, InstructionBLIP-Vicuna-7B, Qwen2.5-VL-32B, and Qwen2.5-VL-72B.Source: D FULL RESULTS OF THE MAIN EXPERIMENT · p. 24

E FURTHER ANALYSIS

Further analyses examine what TVI and representation distances capture, how they vary with prompts and training, and where TVI–correctness alignment has exceptions. The section also contrasts latent- and output-space signals and tests an alternative TVI aggregation.

E FURTHER ANALYSIS

  • Under strong language priors, attention can concentrate pathologically on irrelevant or static image regions, so high attention weight need not indicate meaningful visual processing.The authors call this pattern an attention sink.Source: E FURTHER ANALYSIS · p. 24
  • Latent-space representation distances preserve distinctions that projection into output-token space can obscure, such as grounding versus reliance on language priors.The authors argue that surface-level outputs alone may not faithfully reflect internal decision-making.Source: E FURTHER ANALYSIS · p. 25, Figure 13 · p. 25
  • A standard-deviation-weighted TVI slightly improves effectiveness over averaging in specialized scenarios.The weighting emphasizes layers whose representation-distance distributions have higher variability across samples.Source: E FURTHER ANALYSIS · pp. 25–26, Table 9 · p. 26
  • TVI is significantly higher under generic vision-centric instructions than under misleading instructions that elicit strong language priors.The comparison includes prompts such as “Describe the image in detail.”Source: E FURTHER ANALYSIS · p. 26, Figure 14 · p. 26
  • Across visual instruction-tuning stages, VIP remains around the 12th layer while TVI rises as training progresses.The authors interpret the stable VIP position as largely established during pretraining and the TVI increase as gradual improvement in visual incorporation.Source: E FURTHER ANALYSIS · p. 27, Figure 15 · p. 27
  • TVI–correctness alignment has two stated exceptions: low TVI can accompany correct answers to low-visual-demand questions, while high TVI can accompany failure on difficult visual questions.The authors describe these cases as false negatives and false positives, respectively, and note that they introduce noise into evaluations based on TVI versus correctness.Source: E FURTHER ANALYSIS · p. 27
  • Some correctly answered ViLP samples have unexpectedly low TVI because they can be solved through knowledge leakage or lucky guessing rather than visual input.The authors characterize these cases as intrinsically less visually demanding, not as failures of TVI.Source: E FURTHER ANALYSIS · p. 27, Figure 17 · p. 28

F DETAILS ON THEORETICAL ANALYSIS AND PROOFS

The theoretical analysis characterizes representation-distance estimates as a two-sample statistic and recasts the distance as a density-estimation quantity. It also gives an information-theoretic interpretation of the expected-distance difference.

F DETAILS ON THEORETICAL ANALYSIS AND PROOFS

  • The empirical difference in expected representation distances is a two-sample test statistic with asymptotic normality under finite-second-moment conditions.The analysis also states that asymptotic normality holds for dependent samples when covariance terms are considered.Source: §F.1 · p. 28
  • With a squared l2 distance, the distance between the multimodal and text-only representations is the negative log-likelihood estimate under a unit-variance Gaussian centered on the text-only representation, up to a constant.The paper frames the distance measurement as a density-estimation problem.Source: §F.1 · p. 28
  • The expected-distance difference is interpreted as excess surprisal under the blind-representation-centered normal estimate, compensated by estimation quality.The paper presents this as an information-theoretic interpretation involving entropy, cross-entropy, and KL divergence.Source: §F.1 · p. 29

F.2 NOTATIONS AND PROBLEM SETUP FOR THEOREM F.6

This setup recasts representation-distance comparison as a binary classification problem over multimodal and text-only input distributions. It defines a distance-based hypothesis, its membership labels, and expected error, then introduces H-divergence.

F.2 NOTATIONS AND PROBLEM SETUP FOR THEOREM F.6

  • The distance between multimodal and text-only representations is used as a real-valued hypothesis for whether an input comes from the multimodal distribution rather than the text-only distribution.The setup recasts representation-distance measurement as binary classification.Source: §F.2 · p. 29
  • The labeling function assigns 1 to multimodal-distribution inputs and 0 to text-only-distribution inputs; risk is expected absolute error against those labels.Source: §F.2 · pp. 29–30
  • The setup introduces H-divergence as a measure of discrepancy between two input distributions relative to a hypothesis class.The paper notes its use in domain adaptation and VLM fine-tuning regimes.Source: §F.2 · p. 30

F.3 PROOF FOR THEOREM F.6

The proof develops high-probability empirical bounds for a representation-based discrepancy, while the accompanying statements frame language-prior measurement as a diagnostic aim and describe the study’s practices and scope.

F.3 PROOF FOR THEOREM F.6

  • The proof states bounds with probability at least 1 − δ for empirical distributions based on N samples each, defining loss as an expected hypothesis discrepancy.The hypothesis compares layer representations for a multimodal query and its text-only input.Source: §F.3 · p. 30
  • A cited lemma relates differences in expected errors across distributions to distributional discrepancy for bounded hypotheses with finite pseudo-dimension.The proof invokes this relation in deriving its proposition.Source: §F.3 · p. 30

G IMPACT STATEMENT

  • Language prior is described as over-reliance on linguistic knowledge without proper visual grounding, but lacks a clear definition or quantitative measure.The statement connects this measurement gap with heuristic or manually annotated balancing of visual and textual information.Source: G IMPACT STATEMENT · p. 31
  • The authors propose a formal framework to characterize and quantify language prior through model behavior, with possible use of its estimates in training objectives or inference strategies.The statement presents this as a principled direction for improving multimodal systems.Source: G IMPACT STATEMENT · p. 31
  • The contrastive chain-of-embeddings method is presented as inspiration for layer-wise analysis, adaptive LVLM training, and inclusive multimodal alignment research (Skean et al., 2025; Bachu et al., 2025; Oh et al., 2025b; Kim et al., 2025; Jung et al., 2024).The statement links these directions to building trustworthy multimodal AI.Source: G IMPACT STATEMENT · p. 31

H DISCLOSURE OF LLM USAGE

  • LLM tools assisted with prose editing and some utility functions, while the authors state that they independently developed the technical content, results, and analytical conclusions.The coding agent’s utility functions were double-checked.Source: H DISCLOSURE OF LLM USAGE · p. 31

I ETHICS STATEMENT

  • The study used publicly available models and benchmarks, involved no human subjects or sensitive data, and encourages validation under more diverse real-world conditions.The metrics and analyses are intended for research and diagnostic use.Source: I ETHICS STATEMENT · p. 32

Related papers

Loading 2509.23050v3…