1688 multimodal & industrial AI

IndustryBench. A multi-lingual benchmark for evaluating industrial domain knowledge of LLMs.

IndustryBench-MIPU. A multi-image benchmark for evaluating industrial product understanding and attribute extraction of MLLMs.

ARBOR. ARBOR（Adaptive Rubric Buffer for Online Reward）是一个面向 search agent RL 训练的可复用 process reward 框架。核心思路是：在 RL 训练过程中按 category 维护一份 rubric memory，从 query-group 内的对比轨迹中诱导 query-local draft rubric，经 admission → consolidation → retirement 生命周期沉淀为 category-level 的 common rubric，在 outcome-only reward 信号坍缩时仍能提供组内过程区分。