Research map / Vision-language and multimodal / CLIP and zero-shot

VERO: Verification and Zero-Shot Feedback Acquisition for Few-Shot Multimodal Aspect-Level Sentiment Classification

AAAI 2025 · Vision-language and multimodal › CLIP and zero-shot · approach: Adapting VLMs for zero-shot classification · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VERO, an acquisition function that uses an LVLM's zero-shot prediction and self-verification divergence to select challenging unlabeled samples for few-shot fine-tuning of LLaVA on multimodal aspect-level sentiment classification.

Problem

Few-shot multimodal aspect-level sentiment classification (MALSC) requires selecting informative training samples, but existing sampling methods ignore the LVLM's own capabilities when fine-tuning large vision-language models.

Approach

Select unlabeled samples where the LVLM's zero-shot prediction and subsequent self-verification disagree, rank them by self-verification uncertainty, and mix in a small fraction of easy confirmed samples to balance fine-tuning.

Key results

• VERO with LLaVA-13b reaches 69.17% accuracy and 69.42% weighted-F1 on Twitter-2015 at 1% few-shot, surpassing MultiPoint by 3.71% accuracy and 4.44% weighted-F1.
• VERO with LLaVA-7b reaches 64.56% accuracy on Twitter-2017 at 1%, outperforming MultiPoint's 61.88%.
• At 7% few-shot, VERO attains 70.28% accuracy on Twitter-2015 and 69.38% on Twitter-2017, exceeding GMP by 3.22 and 3.18 points.

Measured effects reported in the paper

Surprising findings

• Vanilla LLaVA-7b zero-shot accuracy on Twitter-2015 (50.24%) is below that of several fine-tuned small PLM baselines, while fine-tuned LVLMs gain substantially.
• Removing easy samples (lambda=1.0, all challenging samples) hurts performance on positive and negative classes, with best results at lambda=0.9.

Limitations

• Underutilizes unlabeled data; approach is not semi-supervised.
• Evaluated only on two Twitter datasets and LLaVA backbones due to computational constraints.
• Negative-class samples remain underrepresented in selected data.

Experimental setup

Datasets: Twitter-2015 Twitter-2017

Benchmarks: Twitter-2015 Twitter-2017

Metrics: accuracy weighted-F1

Baselines: MultiPoint PVLM UP-MPF GMP LLaVA-7b LLaVA-13b LM-BFF LM-SC GFSC TomBERT

Builds on: LLaVA MultiPoint consistently distributed sampling

Closest papers by meaning

• Large Language Models are Good Prompt Learners for Low-Shot Image Classification (CVPR 2024 · similarity 0.87)
• An Empirical Study on Configuring In-Context Learning Demonstrations for Unleashing MLLMs' Sentimental Perception Capability (ICML 2025 · similarity 0.86)
• Exploring Model Editing for LLM-based Aspect-Based Sentiment Classification (AAAI 2025 · similarity 0.86)
• Making Large Vision Language Models to Be Good Few-Shot Learners (AAAI 2025 · similarity 0.85)
• LR0.FM: Low-Res Benchmark and Improving robustness for Zero-Shot Classification in Foundation Models (ICLR 2025 · similarity 0.85)
• Descriptor and Word Soups: Overcoming the Parameter Efficiency Accuracy Tradeoff for Out-of-Distribution Few-shot Learning (CVPR 2024 · similarity 0.85)