Unverified50% confidenceFactExact time
The paper "Margins, Not Windows: Training-Free Per-Step Lossy Speculative Decoding" proposes the AdaptiveSpec method.
1
Sources
50%
Confidence
Long-term
Relevance
9/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedCLIP采用对比学习框架，使用InfoNCE损失函数，在互联网上约4亿图文对上进行大规模预训练，无需人工标注66% similarVerifiedDuring SFT training, the model only computes loss on the assistant portion (loss masking technique), meaning it only learns how to answer, not how to ask questions.65% similarUnverifiedAdaptiveSpec is a training-free, progressively adaptive speculative decoding method.64% similarUnverifiedContrastive learning is the dominant paradigm for training text embedding models, and the InfoNCE loss function is the most commonly used training objective.64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/869677API
curl https://kongchang.com/api/v1/knowledge/claims/869677MCP
get_claim(id=869677)