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 "Visuals Lie, Consistency Speaks", accepted to the ICLR 2026 Workshop on Multimodal Intelligence, 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

• Accepted to the ICLR 2026 Workshop on Multimodal Intelligence
• 3 VLM families evaluated
• Self-consistency correlates with correctness at R = 0.429, versus ≈0.001 for spatial attention
• ≥90% precision when self-consistency is perfect

Leadership & Impact

• Second author (of nine) on the ICLR 2026 workshop paper; coordinated cross-family evaluation
• Designed reliability probes and analysis that informed broader model evaluation workflows

Links