Zijie Xin 辛梓杰

Omni-modal large language models (om-LLMs) achieve unified audio-visual understanding by encoding video and audio into temporally aligned token sequences interleaved at the window level. However, processing these dense non-textual tokens throughout the LLM incurs substantial computational overhead. Although training-free token selection can reduce this cost, existing methods either focus on visual-only inputs or prune om-LLM tokens only before the LLM with fixed per-modality ratios, failing to capture how cross-modal token importance evolves across layers. To address this limitation, we first analyze the layer-wise token dependency of om-LLMs. We find that visual and audio dependencies follow a block-wise pattern and gradually weaken with depth, indicating that many late-layer non-textual tokens become redundant after cross-modal fusion. Motivated by this observation, we propose SEATS, a training-free, stage-adaptive token selection method for efficient om-LLM inference. Before the LLM, SEATS removes spatiotemporal redundancy via attention-weighted diversity selection. Inside the LLM, it progressively prunes tokens across blocks and dynamically allocates the retention budget from temporal windows to modalities using query relevance scores. In late layers, it removes all remaining non-textual tokens once cross-modal fusion is complete. Experiments on Qwen2.5-Omni and Qwen3-Omni demonstrate that SEATS effectively improves inference efficiency. Retaining only 10% of visual and audio tokens, it achieves a 9.3\times FLOPs reduction and a 4.8\times prefill speedup while preserving 96.3% of the original performance.

Omni-proactive streaming video understanding, i.e., autonomously deciding when to speak and what to say from continuous audio-visual streams, is an emerging capability of omni-modal large language models. Existing benchmarks fall short in three key aspects: they rely primarily on visual signals, adopt polling or fixed-timestamp protocols instead of true proactive evaluation, and cover only a limited range of tasks, preventing reliable assessment and differentiation of omni-proactive streaming models. We present OmniPro, the first benchmark to jointly evaluate omni-modal perception, proactive responding, and diverse video understanding tasks. It comprises 2,700 human-verified samples spanning 9 sub-tasks and 3 cognitive levels, covering 6 basic video understanding capabilities. Notably, 84% of samples require audio signals (speech or non-speech), and each sample is annotated with modality-isolation labels to enable fine-grained multimodal analysis. We further introduce a dual-mode evaluation protocol: \textitProbe mode assesses content understanding by querying the model before and after each ground-truth trigger, while \textitOnline mode evaluates full proactive ability by requiring models to autonomously decide when to respond in streaming input. Evaluating 11 representative models reveals three key findings: (1) audio provides consistent gains but with highly variable utilization across models, (2) performance degrades significantly over time, indicating limited long-horizon robustness, and (3) non-speech audio perception remains the weakest dimension.

Adding proper background music helps complete a short video to be shared. Previous work tackles the task by video-to-music retrieval (V2MR), aiming to find the most suitable music track from a collection to match the content of a given query video. In practice, however, music tracks are typically much longer than the query video, necessitating (manual) trimming of the retrieved music to a shorter segment that matches the video duration. In order to bridge the gap between the practical need for music moment localization and V2MR, we propose a new task termed Music Grounding by Short Video (MGSV). To tackle the new task, we introduce a new benchmark, MGSV-EC, which comprises a diverse set of 53k short videos associated with 35k different music moments from 4k unique music tracks. Furthermore, we develop a new baseline method, MaDe, which performs both video-to-music matching and music moment detection within a unified end-to-end deep network. Extensive experiments on MGSV-EC not only highlight the challenging nature of MGSV but also set MaDe as a strong baseline.

Ad-hoc Video Search (AVS) involves using a textual query to search for multiple relevant videos in a large collection of unlabeled short videos. The main challenge of AVS is the visual diversity of relevant videos. A simple query such as "Find shots of a man and a woman dancing together indoors" can span a multitude of environments, from brightly lit halls and shadowy bars to dance scenes in black-and-white animations. It is therefore essential to retrieve relevant videos as comprehensively as possible. Current solutions for the AVS task primarily fuse multiple features into one or more common spaces, yet overlook the need for diverse spaces. To fully exploit the expressive capability of individual features, we propose LPD, short for Learning Partially Decorrelated common spaces. LPD incorporates two key innovations: feature-specific common space construction and the de-correlation loss. Specifically, LPD learns a separate common space for each video and text feature, and employs de-correlation loss to diversify the ordering of negative samples across different spaces. To enhance the consistency of multi-space convergence, we designed an entropy-based fair multi-space triplet ranking loss. Extensive experiments on the TRECVID AVS benchmarks (2016-2023) justify the effectiveness of LPD. Moreover, diversity visualizations of LPD’s spaces highlight its ability to enhance result diversity.

For text-to-video retrieval (T2VR), which aims to retrieve unlabeled videos by ad-hoc textual queries, CLIP-based methods currently lead the way. Compared to CLIP4Clip which is efficient and compact, state-of-the-art models tend to compute video-text similarity through fine-grained cross-modal feature interaction and matching, putting their scalability for large-scale T2VR applications into doubt. We propose TeachCLIP, enabling a CLIP4Clip based student network to learn from more advanced yet computationally intensive models. In order to create a learning channel to convey fine-grained cross-modal knowledge from a heavy model to the student, we add to CLIP4Clip a simple Attentional frame-Feature Aggregation (AFA) block, which by design adds no extra storage / computation overhead at the retrieval stage. Frame-text relevance scores calculated by the teacher network are used as soft labels to supervise the attentive weights produced by AFA. Extensive experiments on multiple public datasets justify the viability of the proposed method. TeachCLIP has the same efficiency and compactness as CLIP4Clip, yet has near-SOTA effectiveness.