A framework that exposes when MLLMs are cheating by memorizing training data
MM-Detect, proposed in this paper, reveals hidden data contamination in multimodal LLMs through systematic testing
https://arxiv.org/abs/2411.03823
Original Problem 🤔:
MLLMs (Multimodal LLMs) show impressive performance on various benchmarks, but data contamination during training creates unfair performance comparisons. Existing contamination detection methods for LLMs don't work well for MLLMs due to their multiple modalities and training phases.
-----
Solution in this Paper 🛠️:
→ Introduces MM-Detect framework that detects contamination through two methods: Option Order Sensitivity Test and Slot Guessing for Perturbation Caption
→ Option Order Test checks if model performance changes when multiple choice options are reordered, indicating memorization
→ Slot Guessing method tests if model can predict masked words in original vs back-translated captions
→ Framework evaluates contamination at both dataset and instance levels using metrics like Correct Rate (CR) and Instance Leakage (IL)
-----
Key Insights 🔍:
→ Both open-source and proprietary MLLMs show contamination across multiple datasets
→ Training set leakage gives unfair advantages, with ~4.3% accuracy boost
→ Contamination can originate from both pre-training LLM phase and multimodal training phase
→ Higher contamination found in older benchmarks like COCO-Caption and NoCaps
-----
Results 📊:
→ MM-Detect successfully identified contamination with varying sensitivity levels (10% to 100%)
→ For intentionally contaminated models, average CR increased by 8.2% and PCR by 3.7%
→ Framework detected significant contamination in Claude-3.5-Sonnet (∆ of -5.3) and VILA1.5-3B
Share this post