What is the HDR framework and how does it solve visual reasoning challenges in video models?

HDR (Hierarchical Denoising) integrates hierarchical latent variables into causal video generation using a tree-structured representation for coarse-to-fine reasoning, combined with Sparse Hierarchical Attention Pattern (SHAP) to reduce computational costs while maintaining logical consistency.

How much does HDR improve success rates and why does it matter?

On a six-task benchmark including maze navigation and Towers of Hanoi, HDR improved success rates from 34.22% to 60.29% (76.2% relative gain) and average progress from 76.00 to 89.56, achieving human-like logical consistency in multi-step visual reasoning for the first time.

What are HDR's data efficiency and speed metrics? What's next?

HDR retains 82.9% of full-data performance with only 2% training data and achieves 54.2× faster inference than bidirectional diffusion models. Future directions include multimodal reasoning with language models and complex dynamic environment simulation.

層級去噪機制賦能多步視覺推理：HDR框架實現低延遲流式生成與邏輯一致性

針對當前視頻模型在複雜多步視覺推理中缺乏類人邏輯一致性與低延遲流式輸出的雙重困境，本文提出了一種名為層級去噪（HDR）的統一框架。該框架將層級潛在變數整合至因果視頻生成中，通過樹狀結構組織視頻潛在表示，實現了從粗粒度到細粒度的推理過程。具體而言，粗粒度去噪層保留不確定假設以支持全局規劃，而細粒度層則逐步將其細化為具體的視覺狀態。此外，引入稀疏層級注意力模式（SHAP）以顯著降低時間注意力計算成本。在包含迷宮導航、漢諾塔等六個任務的自定義基準測試中，HDR的成功率從34.22%提升至60.29%，相對提升76.2%，平均進度從76.00提升至89.56。同時，HDR保持了0.70秒每潛在變數的低延遲流式生成速度，推理速度比雙向擴散模型快54.2倍，且在僅使用2%訓練數據的情況下仍能保持82.9%的全量數據性能。

Sources

arXiv