NeurIPS 2026

Multimodal Concepts for Interpreting Large Vision–Language Models

Md Abdul Kadir1,2  ·  Omair Shahzad Bhatti1  ·  Daniel Sonntag1,2
1German Research Center for Artificial Intelligence (DFKI), Saarbrücken   2University of Oldenburg
A 2x2 grid of a hotdog, bear, rabbit and tiger; for each LVLM prediction TGCL shows the top attributed concepts with their image patches and decoded tokens.
Given a grid image, the LVLM predicts the main objects and TGCL retrieves the top attributed concepts from its learned dictionary. Retrieved patches and decoded concept tokens show each prediction is grounded in related visual evidence.

Abstract

Large vision–language models (LVLMs) are inherently opaque, making it difficult to determine whether their outputs are grounded in visual evidence or driven by language-model priors. Existing concept-based interpretability methods are limited to classification settings, rely on proxy models, or require predefined tokens, and thus do not extend to LVLMs.

We propose Text-Guided Concept Learning (TGCL), a weakly supervised framework for extracting multimodal concept vectors in LVLMs without token supervision. TGCL builds concept-to-image mappings from data and extracts patch-level activations via concept-guided probing. It then formulates concept learning as a contrastive disentanglement problem, isolating concept-specific patches from background patches to produce sparse, stable, and semantically aligned concept vectors.

We conduct experiments on four datasets (ImageNet, MSCOCO, CIFAR-100, DTD) with three recent LVLMs. Across ImageNet and MSCOCO, TGCL consistently outperforms state-of-the-art baselines, with gains of up to:

+20pts concept sparsity
+8pts stability
−30pts concept overlap
+20pts attribution alignment
+4pts faithfulness

Method

TGCL moves concept extraction from token-level representations to patch-level activations guided by textual concepts. Concept learning is a linear contrastive disentanglement problem: keeping the transformation linear keeps concept vectors in the model's own activation space, rather than explaining one black box with another.

Seven-step TGCL pipeline: concept discovery, concept-example bags, binary prompting, hidden activation extraction, binary dictionary learning, positive vector identification, and visual/textual grounding.
Overview of TGCL. A concept vocabulary and weak concept-example bags are built from unlabeled images; each crop is probed with a binary concept prompt, and a two-basis dictionary separates concept-present from concept-absent directions.
“Does the image contain ck? If yes, output ck; otherwise, output No ck.”
  1. 1Concept discovery & bags

    The LVLM lists the concepts in each image. For each concept, crops (random or SAM proposals) from supporting images form a bag that mixes concept-present and concept-absent patches — no labels needed.

  2. 2Concept-guided probing

    Each crop gets the binary prompt above. Residuals are averaged over the shared concept-token span only (dropping “No”), so the contrast comes from the image, not the token identity.

  3. 3Binary dictionary learning

    Per concept, the activations S are factorized as S ≈ VU⊤ with just two bases (concept vs. no-concept). K small independent dictionaries avoid atom interference and scale to large vocabularies.

  4. 4Grounding & attribution

    The positive basis is picked via maximum-activated crops. Each vector is stored with visual (top patches) and textual (output-head tokens) groundings; predictions are attributed by cosine similarity to residual activations.

Results

Concept vector quality

Gemma-3n-E4B on ImageNet and MSCOCO. Higher sparsity, lower instability and overlap are better.

MethodDecomp.ImageNetMSCOCO
Sparsity ↑Instability ↓Overlap ↓Sparsity ↑Instability ↓Overlap ↓
CRAFTSNMF0.420.020.690.420.000.81
SAE0.390.550.020.370.810.04
SIMPLE0.420.000.520.420.330.59
MCDSNMF0.420.020.700.420.000.78
SAE0.390.670.020.370.780.05
SIMPLE0.420.000.670.430.230.45
CoX-LMMSNMF0.430.060.270.430.000.32
SAE0.420.800.020.380.860.03
SIMPLE0.440.000.340.430.090.36
TGCL (ours)SNMF0.610.000.010.630.010.02
SAE0.360.960.780.421.000.04
SIMPLE0.450.000.530.440.010.68

Best values in bold, as in the paper. TGCL with SNMF gives the highest sparsity and the lowest instability and overlap.

Attribution quality

Gemma-3n-E4B. In the combined setting the model labels each cell of a 2×2 grid of unrelated images, and attribution is scored against those predictions.

MethodDatasetMetricRandomText-onlyImage-onlyCombined
TGCL (ours)ImageNetCLIPScore0.52–0.620.67
BERTScore0.710.780.840.86
MSCOCOCLIPScore0.48–0.600.64
BERTScore0.710.890.910.93
CoX-LMMImageNetCLIPScore0.49–0.570.58
BERTScore0.710.740.750.82
MSCOCOCLIPScore0.51–0.570.55
BERTScore0.710.830.790.73

Means shown; standard deviations, DTD, CIFAR-100, Qwen2-VL-7B and Qwen2.5-VL-7B are in the paper.

Faithfulness

CoX-LMM

CoX-LMM concept deletion curves CoX-LMM concept insertion curves

TGCL (ours)

TGCL concept deletion curves TGCL concept insertion curves
Concept deletion (left of each pair) and insertion (right) on MSCOCO with Gemma-3n-E4B. CoX-LMM's curves are flat with weak separation across ranks; TGCL keeps a clear order (Top-1 > Top-2 > Top-3), with sharper drops under deletion and stronger recovery under insertion.
MethodRanked orderRandom orderRank spread ↑
Ins ↑Del ↓Gap ↑InsDelGap
CRAFT0.7810.2170.5640.5100.4910.0190.016
MCD0.7930.1850.6080.5120.4890.0230.016
CoX-LMM0.7710.1940.5770.5120.4880.0240.016
TGCL (ours)random crops0.8150.1870.6280.5170.4890.0280.086
TGCL (ours)SAM crops0.8260.1760.6500.5160.4790.0370.108

Concept insertion/deletion AUC on MSCOCO-10 with Gemma-3n-E4B, mean over 500 images; Gap = Ins − Del. TGCL (SAM) is better than every baseline with p < 0.001 (Welch's t-test, Holm-corrected). Standard deviations and effect sizes are in the paper.

Monosemantic concepts

ImageNet concepts from CoX-LMM (left), which mix unrelated tokens, next to TGCL concepts (right) such as fur, dog and stripes.
ImageNet concepts. CoX-LMM (left) produces polysemantic vectors — “canine” tied to “hot dog”, lion represented as tiger. TGCL (right) finds fine-grained, monosemantic concepts such as fur, dog and stripe.

More attribution examples

Each prediction on a 2×2 grid, with its two most attributed concepts: attribution score, top activating patches, and decoded tokens.

BibTeX

@inproceedings{kadir2026tgcl,
  title     = {Multimodal Concepts for Interpreting Large
               Vision--Language Models},
  author    = {Kadir, Md Abdul and Bhatti, Omair Shahzad
               and Sonntag, Daniel},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}