Yaping Li

Weizhe Chen, Miao Zhang, Junpeng Jiang et al.

Hybrid attention architectures are becoming an increasingly important paradigm for improving LLM inference efficiency while preserving model quality, making hybrid architecture design a central problem. Existing designs often rely on manual empirical rules or proxy-based selector signals for layer-wise operator allocation. Recent NAS-style systems such as Jet-Nemotron demonstrate the promise of automated hybrid architecture search. However, Jet-Nemotron's PostNAS search stages alone use 200B tokens, making such search pipelines difficult to use as routine methods for hybrid architecture design. We introduce DASH, a fast differentiable search framework for hybrid attention architecture design, which relaxes discrete layer-wise attention operator placement into continuous architecture logits, prepares reusable teacher-aligned linear candidates, and performs architecture-only search with model and operator weights frozen to significantly enhance search efficiency. On Qwen2.5-3B-Instruct, DASH consistently outperforms a comprehensive suite of existing selector-style hybrid attention design baselines, showing that direct differentiable search can discover stronger hybrid architectures. Moreover, DASH achieves stronger RULER performance than released Jet-Nemotron models while remaining competitive on overlapping short-context and general benchmarks. Notably, each DASH search run uses only 12.3M tokens and takes about 20 minutes on a single RTX Pro 6000 GPU, corresponding to merely 0.006% of the PostNAS search tokens reported by Jet-Nemotron. These results suggest that high-quality hybrid attention architectures can be obtained through minutes-level differentiable search, providing a promising direction for hybrid architecture design.

Haidong Wang, Zhiyong Li, Yaping Li et al.

Online multi-object tracking (MOT) is a longstanding task for computer vision and intelligent vehicle platform. At present, the main paradigm is tracking-by-detection, and the main difficulty of this paradigm is how to associate current candidate detections with historical tracklets. However, in the MOT scenarios, each historical tracklet is composed of an object sequence, while each candidate detection is just a flat image, which lacks temporal features of the object sequence. The feature difference between current candidate detections and historical tracklets makes the object association much harder. Therefore, we propose a Spatial-Temporal Mutual Representation Learning (STURE) approach which learns spatial-temporal representations between current candidate detections and historical sequences in a mutual representation space. For historical trackelets, the detection learning network is forced to match the representations of sequence learning network in a mutual representation space. The proposed approach is capable of extracting more distinguishing detection and sequence representations by using various designed losses in object association. As a result, spatial-temporal feature is learned mutually to reinforce the current detection features, and the feature difference can be relieved. To prove the robustness of the STURE, it is applied to the public MOT challenge benchmarks and performs well compared with various state-of-the-art online MOT trackers based on identity-preserving metrics.