Research map / Vision-language and multimodal / Visual question answering
Adaptive Logit Adjustment for Debiasing Multimodal Language Models
Hoin Jung, Junyi Chai, Xiaoqian Wang
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 |
|---|---|---|---|
| replace embedding-level debiasing with ALA-N logit neutralization | LLaVA-1.5 (VQA-Task-1) · FACET VQA-Task-1 | MRC: 6.91 -> 5.96 | Improved |
| replace baseline with ALA-N | PaliGemma (VQA-Task-1) · FACET VQA-Task-1 | MRC: 5.72 -> 3.50 | Improved |
| add ALA-N to BLIP captioning | BLIP · COCO captioning | MRC: 2.52 -> 1.69 | Improved |
| add ALA-BA to CLIP-CAP captioning | CLIP-CAP · COCO captioning | MRC: 2.91 -> 2.34 | Improved |
| add ALA-N to PaliGemma stereotype task | PaliGemma · SocialCounterfactuals | Dmean: 6.08 -> 3.81 | Improved |
| add DeAR to BLIP captioning | BLIP · COCO captioning | MRC: 2.52 -> 8.51 | Hurt |
| add prompt-engineering gender instruction | LLaVA-1.5 · FACET VQA-Task-1 | MRC: 6.91 -> 12.75 | Hurt |
| add CLIP-clip to PaliGemma judge task | PaliGemma · FACET VQA-Task-3 | DWCA: -2.60 | Hurt |
| add DeAR to LLaVA judge task | LLaVA-1.5 · FACET VQA-Task-3 | DWCA: -2.23 | Hurt |
| add small logit adjustment strength lambda = 0.1 (ablation) | ALA with CLIP-CAP · COCO captioning | MRC: lambda = 0.1 | Improved |
| use lambda = 2 instead of excessively large lambda (ablation) | ALA · COCO captioning / FACET | MRC and utility: lambda = 2 | Improved |
| selective adjustment with threshold tau = 0.1 on token importance (ablation) | ALA · COCO captioning | inference time: 1.2% inference time increase vs 101.5% for VDD | Improved |
Closest papers by meaning
- Language Bias in LVLMs: From In-Depth Analysis to Simple and Effective Mitigation (ICML 2026 · similarity 0.89)
- BendVLM: Test-Time Debiasing of Vision-Language Embeddings (NeurIPS 2024 · similarity 0.89)
- A Unified Debiasing Approach for Vision-Language Models across Modalities and Tasks (NeurIPS 2024 · similarity 0.89)
- On Fairness of Unified Multimodal Large Language Model for Image Generation (NeurIPS 2025 · similarity 0.88)
- Logit-Attention Divergence: Mitigating Position Bias in Multi-Image Retrieval via Attention-Guided Calibration (ICML 2026 · similarity 0.88)
- LoMo: Local Modality Substitution for Deeper Vision-Language Fusion (NeurIPS 2026 · similarity 0.88)
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.
