"सपनों से भरे नैना, ना नींद है ना चैना"
For the eyes that are filled with dreams, there is neither sleep nor peace
— Sapnon Se Bhare Naina | Lyrics: Javed Akhtar | Singer: Shankar Mahadevan
— Cats
— Indie Games
— Sunsets
— Chocolates
— Pineapple Pizza
— Rock / 90s Music
— The Color Pink
— Bike Rides
— Painting
— Jigsaw Puzzles
— Aglio e Olio
— Gokarting
Dislikes— Crowded Places
— Messy Tables
— Loud Noises
researchMarch 21, 2026
Before applying PID to real VLMs, the paper constructs inputs where the true modality contributions are analytically derivable and checks whether the metric recovers them. This post is about why that step is necessary and what it actually shows.
Read more →
researchMarch 21, 2026
Across six models and six benchmarks, text contribution exceeds image contribution in every case but one. The numbers vary in structured ways that reveal what actually drives the imbalance -- and it is not scale, not training data, and not the benchmark.
Read more →
researchMarch 20, 2026
When you hold encoders, training, and task constant and vary only the fusion mechanism, text contribution shifts by up to 17 percentage points. Scale produces nothing comparable. The wiring between modalities shapes information flow before training starts.
Read more →
researchMarch 20, 2026
High benchmark accuracy tells you almost nothing about whether a VLM is actually using its image input. Measuring modality attribution turns out to be harder than it looks, and every existing approach shares the same structural blind spot.
Read more →
