Skip to content

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

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

ICLR 2026; chain-of-embedding contrast exposes language prior dominance in LVLMs

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

LVLMs systematically over-rely on language priors (memorized textual patterns) while under-utilizing visual evidence, yet prior analyses only probe input–output behavior. This paper introduces a mechanistic framework based on chain-of-embedding analysis that identifies a Visual Integration Point (VIP)—a specific layer where visual information first meaningfully reshapes hidden representations—and a Total Visual Integration (TVI) estimator that quantifies language prior strength per sample without any annotation.

Problem

Existing language-prior analyses in LVLMs rely on black-box input–output probing (curated counterfactual datasets, modality-conflicting queries), which cannot reveal where inside the model visual evidence is integrated or ignored, cannot disentangle LP mechanisms from surface statistics, and cannot provide per-sample, annotation-free diagnostics.

Method

The framework operates on the chain-of-embedding \((Z^1, \ldots, Z^L)\)—the sequence of last-token hidden states across all \(L\) decoder layers. Two forward passes are run per input: \(Z^l_\text{vis} := f_l(x_v, x_t)\) (vision+text) and \(Z^l_\text{blind} := f_l(\emptyset, x_t)\) (text only). The expected per-layer representation distance is:

\[D^l(\mathcal{P}_*, F_\theta) := \mathbb{E}_{(X_v, X_t) \sim \mathcal{P}_*}\left[d(Z^l_\text{vis}, Z^l_\text{blind})\right]\]

using cosine distance by default. The VIP \(l^*\) is the layer where \(D^l(P_{VT}, F_\theta) - D^l(P_T, F_\theta)\) transitions from \(\approx 0\) to \(> \tau\), distinguishing vision-dependent (\(P_{VT}\)) from vision-independent (\(P_T\)) examples. Samples are partitioned into these groups by checking whether removing visual input changes the model's prediction. The TVI estimator aggregates post-VIP representational discrepancy for a single input \(x\):

\[\text{TVI}(l^*; x, F_\theta) = \frac{1}{L - l^* + 1} \sum_{l=l^*}^{L} d(z^l_\text{vis}, z^l_\text{blind})\]

Low TVI → strong language prior; high TVI → effective visual grounding.

Key Contributions

  • First mechanistic, layer-wise framework for analyzing language prior in LVLMs via contrastive chain-of-embedding.
  • Discovery and formalization of the Visual Integration Point (VIP): a universal, model-intrinsic layer transition observed across 10 LVLMs and 6 benchmarks (60 settings).
  • The TVI estimator: an annotation-free, per-sample proxy for language prior strength grounded in internal representation dynamics.
  • Theoretical analysis and analytic bounds on TVI; interventional validation showing TVI tracks deliberate increases in visual attention.

Results

  • VIP universality: VIP consistently emerges across all 60 model–dataset combinations; VIP location is dataset-agnostic within each model (e.g., layers 18–20 for Qwen2.5-VL-7B, layers 20–22 for Gemma-3-4B).
  • TVI vs. correctness correlation (Spearman's \(\rho\), post-\(l^*\)):
  • Qwen2.5-VL-7B on VLind-Bench: \(\rho = 0.7241\) (\(p < 0.001\)); pre-\(l^*\): \(\rho = 0.1489\)
  • Gemma3-4B on VLind-Bench: \(\rho = 0.7174\) (\(p < 0.001\)); pre-\(l^*\): \(\rho = 0.4659\)
  • TVI vs. competing proxies (Spearman's \(\rho\) with correctness on VLind-Bench/ViLP):
  • TVI: 0.7155 / 0.6335 (Qwen2.5-VL-7B); Visual Attention: 0.0871 / −0.0364; Output Divergence: 0.2978 / 0.5084
  • TVI consistently outperforms visual-attention and output-divergence baselines across all models and datasets
  • Interventional validation: Applying PAI (attention-correction) to Qwen2.5-VL-7B raises accuracy from 50.00% → 52.33% and TVI from 0.038 → 0.144, confirming TVI tracks actual visual integration changes.
  • Strong vs. weak LP datasets: ViLP (strong LP) yields lower TVI than MMBench (weak LP) across models, confirming construct validity.

Limitations

  • VIP threshold \(\tau\) and layer selection are manually chosen per model; an automatic selection method is deferred to the appendix and not fully validated.
  • The vision-independent/dependent partition (\(D_{VT}\) vs. \(D_T\)) is derived from prediction agreement—a noisy proxy that may misclassify samples where the model happens to get the right answer by language prior.
  • TVI requires two full forward passes per sample, doubling inference cost for diagnostic use.
  • Evaluated only on decoder-only LVLM architectures; applicability to encoder-decoder models (e.g., BLIP-2's cross-attention design) is unverified.
  • Theoretical bounds are derived but not empirically tightened across the full 60-setting sweep.

Relevance to Vision-Language Models

This work provides the first principled, mechanistic decomposition of where and how strongly LVLMs integrate visual evidence, directly addressing one of the field's core failure modes—hallucination and language-prior dominance—at the representation level rather than the output level. The VIP concept has immediate diagnostic utility: it localizes the depth at which multimodal fusion becomes task-relevant, informing architecture choices (e.g., where to intervene with attention corrections or adapter layers). TVI's superior correlation with downstream accuracy over attention-based and output-divergence proxies establishes it as a practical, annotation-free tool for model selection, benchmark design, and mitigation evaluation. For researchers tracking VLMs, this offers a reusable interpretability toolkit applicable to any transformer-based LVLM without requiring curated adversarial datasets.