Research map / Vision-language and multimodal / Human and embodied modeling
Neuro-Fuzzy Concept Learning for Interpretable Large Multimodal Models
Ritik Mishra, Vanshika Gupta, M. Sajid, M. Tanveer
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use neuro-fuzzy concept learning instead of CoX-LMM Semi-NMF | Neuro-FeX on LLaVA · COCO token Bus | CLIPScore: 0.621 -> 0.667 | Improved |
| use neuro-fuzzy concept learning instead of CoX-LMM | Neuro-FeX-v0 on LLaVA · COCO token Cat | CLIPScore: 0.649 -> 0.593 | Hurt |
| use neuro-fuzzy concepts on other backbones with 30 diverse tokens | Neuro-FeX-v1 on Qwen2.5-VL · 30 diverse tokens | mean CLIPScore: 0.616 -> 0.643 | Improved |
| use K-means-initialized Gaussian membership (v1/v2) instead of SVD-initialized cosine membership (v0) (ablation) | Neuro-FeX on LLaVA · COCO token Dog overlap | overlap score: 0.067 -> 0.430 | Hurt |
| use RBF initialization instead of random initialization (ablation) | Neuro-FeX · reconstruction error | reconstruction error: 4.47 -> 0.43 | Improved |
| add fuzzy gating on top of RBF init (ablation) | Neuro-FeX · reconstruction error | reconstruction error: 0.46 -> 0.43 | No clear effect |
| decrease Gaussian scaling parameter gamma (ablation) | Neuro-FeX · Dog token, Qwen2 and LLaVA | CLIPScore | Improved |
| use Neuro-FeX instead of SAEs for monosemantic concepts | Neuro-FeX on SigLIP SoViT-400m · Monosemanticity score | mean MS: 0.610 -> 0.711 | Improved |
| use Neuro-FeX instead of Matryoshka SAE on Qwen3 | Neuro-FeX on Qwen3 · Monosemanticity score | mean MS: 0.549 -> 0.544 | No clear effect |
Closest papers by meaning
- A Concept-Based Explainability Framework for Large Multimodal Models (NeurIPS 2024 · similarity 0.91)
- Re-Imagining Multimodal Instruction Tuning: A Representation View (ICLR 2025 · similarity 0.87)
- Walking the Web of Concept-Class Relationships in Incrementally Trained Interpretable Models (AAAI 2025 · similarity 0.87)
- TokenSwap: Benchmarking and Reducing the Modality Gap in Multimodal LLMs (NeurIPS 2026 · similarity 0.86)
- Implicit Multimodal Alignment: On the Generalization of Frozen LLMs to Multimodal Inputs (NeurIPS 2024 · similarity 0.85)
- Enhancing Multimodal Large Language Models Complex Reason via Similarity Computation (AAAI 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
