Composed Video Retrieval (CVR) aims to retrieve a target video given a reference video and a modification text. The central challenge is to identify and use video content that is relevant to the retrieval need specified by the modification text. Existing methods typically encode videos before considering this textual need, causing visual evidence to be formed in a largely text-agnostic manner. We propose StaDy-CLIP, a compact dual-pathway framework for text-conditioned evidence composition. It preserves the frozen CLIP visual encoder as a static pathway for CLIP-aligned static evidence, introduces a lightweight side pathway for dynamic evidence, and uses Text-Conditioned Composition Blocks to perform patch-wise dynamic evidence selection and textual modulation. A learnable Gallery Conditioning Prompt further enables gallery videos to be encoded within the same evidence composition framework. StaDy-CLIP achieves state-of-the-art performance on FineCVR-1M and TF-CoVR with only 7.78M trainable parameters, improving prior best results by +4.05 in Recall@1 and +1.49 in mAP@5. Ablations and encoder-transfer results validate its effectiveness and encoder-level transferability for existing CVR methods.

The growing deployment of large language models (LLMs) in socially consequential settings makes it increasingly important to understand the value systems they encode and the moral judgments they produce. However, existing evaluations are largely grounded in Western ethical traditions, leaving non-Western moral perspectives underexplored. To address this gap, we introduce Sage, a benchmark for evaluating LLMs through the lens of Classical Chinese philosophy. Sage is grounded in major Pre-Qin philosophical debates and represents normative alignment in a continuous six-dimensional value space. It includes three task types and is constructed from over 4,000 passages spanning ten philosophical schools. Experiments on eight leading LLMs reveal that models occupy highly similar regions in the proposed value space, exhibit a clear gap between self-declared philosophical identities and benchmark-measured behavior, and show systematically different reasoning patterns across task granularities. Code and data will be released after publication.