Unverified50% confidenceFactExact time
QiYao-I proposes a sampling-conditioned temporal manifold attention mechanism that maps real timestamps to a learnable temporal manifold feature space and injects temporal manifold biases into the attention layers.
1
Sources
50%
Confidence
Long-term
Relevance
10/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLarge model inference refers to the process by which a model, after completing training, receives user input and generates output in real time, and is extremely sensitive to latency.65% similarUnverifiedIn U-Net or DiT architectures used in diffusion models, the computation of attention layers is often the primary source of memory and time consumption64% similarUnverified上下文图谱为智能体提供了介于临时上下文和模型权重微调之间的中间层记忆机制，可在不改变底层模型的前提下持续积累经验64% similarUnverifiedThe attention sink phenomenon was systematically described by Xiao et al. from MIT in the 2023 StreamingLLM paper, referring to the tendency of Transformer models to assign a large proportion of attention weights to the first few tokens at the beginning of a sequence (typically tokens 1–4).64% similarUnverifiedDiffusion models do not logically memorize the precise appearance of the previous frame, but instead rely on attention mechanisms to maintain temporal coherence; when scenes change, the coherence constraint weakens, causing character appearance drift.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/983278API
curl https://kongchang.com/api/v1/knowledge/claims/983278MCP
get_claim(id=983278)