A research team trained and evaluated 210 multimodal large language models (MLLMs) by pairing 70 vision-encoder configurations with Qwen3-1.7B, Qwen2.5-1.5B and SmolLM2-1.7B under a uniform LLaVA-style two-stage recipe, then testing them on 11 downstream benchmarks, according to a report by Machine Heart. They found that as the candidate pool of vision encoders grows more numerous and diverse, existing cross-modal evaluation metrics lose predictive power and can even fall behind purely unimodal vision-only measures. The team's proposed fix, RAVEL, is training-free: it scores a candidate encoder by how consistent image-side and text-side nearest-neighbor relations are, adding stabilized PCA whitening and patch-level similarity matching. The authors report an average Spearman correlation of 0.85 with real downstream performance across the three LLMs, 36.1% above the trained Alignment Probing method and 93.5% and 101.1% above the training-free GW and MutualNN baselines; code, the paper and the 210 MLLM checkpoints are open-sourced.