Balanced Twin Weight Factorization for Efficient MLLMs
BTWF reports 2–6% gains over state-of-the-art methods without expensive finetuning, at comparable computational costs.
The experiments cover six models and nine benchmarks; with expensive finetuning, the reported gains are 1–3.5%. The method reduces inference work by computing a shared transformation once for layers that receive the same input. Two balancing techniques keep larger layers from dominating the shared factors and account for their downstream effects. The authors also report that adding computation during training can improve task performance without requiring extra resources at test time.
ArXiv · AI/CL/LG's note
The experiments cover six models and nine benchmarks; with expensive finetuning, the reported gains are 1–3.5%. The method reduces inference work by computing a shared transformation once for layers that receive the same input. Two balancing techniques keep larger layers from dominating the shared factors and account for their downstream effects. The authors also report that adding computation during training can improve task performance without requiring extra resources at test time.
ArXiv · AI/CL/LG's note
score 4