LLaVAProbe (Algoverse AI)

Overview

LLaVAProbe is a research effort that studies reliability signals in vision-language models, separating attention structure from answer correctness. The project produced the ICLR 2026 workshop paper "Visuals Lie, Consistency Speaks" and focused on measuring attention coherence, evolution across layers, and self-consistency during generation.

Highlights

• • Built evaluation pipelines to quantify attention clusters and spatial entropy across model layers
• • Compared reliability signals across LLaVA-1.5, PaliGemma, and Qwen2-VL
• • Showed self-consistency outperforms attention structure for predicting correctness

Metrics

• • ICLR 2026 Workshop paper accepted
• • 3 VLM families evaluated
• • ≥90% precision when self-consistency is perfect

Leadership & Impact

• • Main-authored the ICLR 2026 workshop paper and coordinated cross-family evaluation
• • Designed reliability probes and analysis that informed broader model evaluation workflows

Links