Qwen3-VL-4B + DeltaDirect: motion direction 49.7 → 93.4. Training-only projector loss. Inference unchanged. Above even Qwen3.5 models and GPT-5. @Alibaba_Qwen @QwenDevs , your model with our loss. Details and code 👇 jong980812.github.io/which-w…

Which Way Did It Move? (NeurIPS 2026) Video-LLMs are near chance at telling if an object moves left/right/up/down. The info is in the features; the model can’t bind it to the answer. DeltaDirect fixes this: 25.9% to 85.4%, +21.9 pts on real videos. 📄 arxiv.org/abs/2605.22823

DANCE (NeurIPS 2025 Spotlight) Saliency maps mix up motion and context. Words can’t describe motion well. DANCE predicts actions through 3 disentangled concepts: motion (pose sequences), objects, and scenes. Clear explanations, competitive accuracy. 📄 arxiv.org/abs/2511.03725

ESSENTIAL (ICCV 2025 Highlight) Most video CIL methods replay dense frames, eating memory. We keep sparse features (episodic) + learnable prompts (semantic) and recover dense temporal info via cross-attn. Strong results, far less memory. 📄 arxiv.org/abs/2508.10896

PCBEAR (CVPR 2025 XAI4CV Workshop, Spotlight) Explaining action recognition with pose concepts, not static text. Static + dynamic pose concepts, discovered automatically by clustering skeleton sequences. Interpretable, with test-time intervention. 📄 arxiv.org/abs/2504.13140

CA²ST (IEEE TPAMI 2026) 🔊 The audio extension of CAST: spatial, temporal, and audio experts exchange info via Bottleneck Cross-Attention in every layer. Balanced on EK100, SSv2 & K400, and strong on audio-visual benchmarks (VGG-Sound, EPIC-SOUNDS). 📄 arxiv.org/abs/2503.23447

CAST (NeurIPS 2023) Action recognition needs what’s in the scene and how it moves. Most models lean one way. We pair a spatial expert (CLIP) and a temporal expert (VideoMAE) via bottleneck cross-attention → consistent gains on EK100, SSv2 & K400. 📄 arxiv.org/abs/2311.18825