Skip to content

Vision Function Layer in Multimodal LLMs

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper discovers that visual-functional decoding in MLLMs is localized to narrow 2–3-layer blocks called Vision Function Layers (VFLs), with a consistent hierarchical ordering (recognition → counting → grounding → OCR) across model families and scales. The authors introduce Visual Token Swapping as a causal probing framework to identify these layers, then exploit this structure for parameter-efficient fine-tuning (VFL-LoRA) and automated data selection (VFL-Select), achieving strong practical gains.

Problem

Despite MLLMs' impressive visual capabilities, it is unknown how distinct visual functions (OCR, counting, recognition, grounding) are represented and distributed across the LLM backbone's decoder layers. Existing interpretability work focuses on token importance or cross-modal interaction but does not isolate individual functional roles at the layer level, leaving the internal mechanism a black box.

Method

Visual Token Swapping (VTS): At decoding layer \(k\), the KV-cache vision tokens \(U^{(k)}\) are replaced with tokens from a source image \(\tilde{U}^{(k)}\) constructed to differ in exactly one visual attribute (e.g., a different word for OCR, a different object count). The "change rate" in model output quantifies how causally influential layer \(k\) is for that function. Paired datasets are curated per function: OCR (arXiv word renders on blank canvas), Counting (CLEVR-adapted), Recognition (COCO object-present vs. blank), Grounding (same object at varied positions).

Visual Token Dropping: For general benchmarks where controlled pairs are infeasible, all vision tokens from layer \(k\) onward are dropped; performance degradation identifies function-critical layer depths.

VFL-LoRA: LoRA adapters are applied only to layers identified as VFLs for the target function (e.g., layers 10–17, 20–23 for count-function in Qwen2.5-VL-7B), reducing tunable parameters while avoiding forgetting in non-targeted VFLs.

VFL-Select: A data-utility score \(R_k(x,y) = P(y|U^{(\leq k)},W) / P(y|U^{(\leq k-1)},W)\) measures per-sample reliance on layer \(k\). Data is partitioned by the layer of highest influence and then uniformly sampled per partition to build a balanced, function-diverse training subset. A small proxy model (TinyLLaVA-0.5B) is used for the classification to enable scalability to 20M-sample pools.

Key Contributions

  • Visual Token Swapping framework: a causal, layer-targeted probing method that replaces KV-cache vision tokens at specific layers to isolate per-function layer contributions.
  • VFL discovery: visual functions localize to 2–3 layers; for Qwen2.5-VL-7B: recognition layers 0–10, counting layers 14–16, grounding layer ~18, OCR layers 22–24.
  • Cross-model consistency: the recognition→counting→grounding→OCR ordering holds across LLaVA-v1.5 (7B, 13B) and Qwen2.5-VL (3B, 7B), from 3B to 70B scale.
  • VFL-LoRA: function-targeted PEFT with ~50% fewer tunable parameters than standard LoRA, better out-of-domain generalization.
  • VFL-Select: automated data selection surpassing human-expert curation; achieves 98% of full-data performance with 20% of LLaVA-665k.

Results

  • VFL localization (Qwen2.5-VL-7B): Count peaks at layer 12 (87.4% change rate), Grounding at layer 18 (100%), OCR at layer 22 (92.8%); other layers contribute negligibly.
  • VFL-LoRA vs. standard LoRA (Qwen2.5-VL-7B): 0.9% of parameters (vs. 1.9%), average CV-Bench in-domain 85.0 vs. 84.4; out-of-domain average 75.0 vs. 74.3; CV-Count 72.6 vs. 70.9. Parameters: ~155M vs. ~309M.
  • VFL-Select vs. Random/Expert (665k pool, 665k subset): SQA-I 86.0 (VFL) vs. 60.4 (Random) vs. 72.1 (Expert); MME-P 1526.3 (VFL) vs. 1410.8 (Random) vs. 1421.0 (Expert).
  • VFL-Select on LLaVA-665k 20% subset: 99.5% of full-data performance on shallow-layer benchmarks, 97.4% on deep-layer benchmarks; outperforms D2-Pruning, EL2N, and COINCIDE.
  • Token dropping: On MMMU, dropping some vision function layers improves performance by up to 1.8%, indicating functional redundancy.

Limitations

  • VFL localization is demonstrated primarily on Qwen2.5-VL and LLaVA families; generalization to architectures with substantially different visual connectors (e.g., Q-Former-based) is not fully validated.
  • Recognition function violates the clean localization pattern—it peaks early but maintains distributed influence across almost all layers, making it harder to target precisely.
  • VFL-LoRA shows no consistent gain on language-prior-dominated subtasks (e.g., CV-Distance), where the targeted visual layers are not the bottleneck.
  • VFL-Select classification relies on a proxy model's functional hierarchy, assuming transfer of layer ordering to the target model; this assumption is validated empirically but not theoretically proven.
  • The paired dataset construction for probing is hand-designed per function and may not cover all atomic visual functions in real-world MLLM capabilities.

Relevance to Vision-Language Models

This paper directly advances mechanistic interpretability of MLLMs by revealing that, contrary to the prevailing "black box" view, visual processing is organized into narrow, functionally specialized decoder layers with a consistent cross-model hierarchy. For researchers tracking VLMs, this has immediate implications for PEFT strategies (layer-targeted LoRA reduces forgetting), data curation (function-aware sampling beats expert selection), and model compression (redundant VFLs can be pruned without hurting non-reliant tasks). The cross-architecture consistency finding—spanning LLaVA to Qwen2.5-VL across model scales—suggests VFLs may be a fundamental emergent property of multimodal instruction tuning rather than an artifact of any specific design.