A research team has found a way to meaningfully fix one of vision-language models' most stubborn weaknesses without touching the underlying weights.
Spatial reasoning — understanding where objects are relative to each other in an image — has long been a weak spot for vision-language models. The standard fix is fine-tuning, which typically means millions of additional parameters via methods like LoRA. ScAle takes a different path: it learns a small set of scalar coefficients that rescale activations in selected transformer layers while leaving the pretrained backbone completely frozen. The whole adaptation runs on roughly 1,000 trainable parameters.
On the SpatialEval benchmark, ScAle achieved up to 134.1% relative accuracy gains — a striking number given the near-zero parameter cost. The method also held its own on real-world visual question-answering datasets (COCOQA and VGQA) without degrading performance on non-spatial tasks, which is the usual tradeoff when you push a model hard in one direction. The implication for deployment is practical: teams could slot ScAle into existing pipelines without retraining or storing large adapter weights.
Most parameter-efficient fine-tuning research chases performance parity with full fine-tuning; ScAle is explicitly not trying to win that race. It recovers a substantial fraction of standard PEFT performance while staying architecture-agnostic — meaning it ports across model families rather than being tied to one vendor's design. Whether those gains hold outside controlled benchmarks, and in more complex real-world scenes, remains the open question.