RAVEL predicts vision encoder performance without training: 0.85 Spearman across 210 MLLMs
机器之心 · wechat · 2026-10-17
研究团队训练并评测了 210 个 MLLM（70 种视觉编码器 × 3 个 LLM），发现现有跨模态评测指标在候选池扩大后预测能力显著下降甚至不如纯视觉指标。他们提出免训练方法 RANCEL：基于图文表征近邻一致性，配合带稳定项的 PCA 白化与 patch 级双向匹配，平均 Spearman 相关系数达 0.85，大幅超越需训练的 Alignment Probing（+36.1%）及其他免训练方法（+93.5%/+101.1%）。代码、评测框架与全部 210 个模型权重均已开源。
More from Multimodal
- Dev builds Mumbai-inspired 3D manga background generator from OpenStreetMap data — CurieuxExplorer · 2026-10-18
- Omni 1'1 flash video demo generates atmospheric foggy Victorian haunted house — Dense-Instruction210 · 2026-10-18
- Flux 3 klein distilled tested: int8 workflow delivers surprisingly good 2MP images — Ok_Tale7582 · 2026-10-18
- AI-generated clip with cinematic truck-driver shot fuels 'feature film soon' talk — 1ili0 · 2026-10-18
- AI Videos Are Getting So Good That Real Content Now Gets Doubted — gouterz · 2026-10-18
- Qwen-Image-2.1 LoRA Unlocks Nearly Any Pose, Works for Both T2I and Editing — metalfans · 2026-10-18