Unverified50% confidenceFactExact time
Unlike BERT, which usually makes only one prediction, a dLLM starts from full masking or a high masking rate and iterates repeatedly, gradually reducing the mask until the entire sentence is filled in.
1
Sources
50%
Confidence
Long-term
Relevance
10/27/2026
First Seen
Sources
Related Entities
Related Claims
Unverified只有在LLM推理比embedding加向量检索慢数十倍以上时，语义缓存才能同时实现降本与提速71% similarUnverified预测编码与数据压缩中的差分编码思想一致，只传输变化量而非完整信号从而降低冗余67% similarUnverifiedCompared to having an LLM perform the same classification task, a dedicated decision model typically has lower latency, more controllable costs, and more well-formed output formats, but at the cost of open-ended expressive capability67% similarUnverifiedConstrained decoding applies a mask to the vocabulary at each decoding step, setting the probability of tokens that violate constraints to zero; the mask computation itself has overhead, and complex grammars slow down decoding.67% similarUnverifiedIn LLM inference, the decode phase is mainly limited by memory bandwidth, so a drop in SM frequency does not cause a proportional drop in speed; the prefill phase is compute-limited and is more severely affected by throttling.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1102713API
curl https://kongchang.com/api/v1/knowledge/claims/1102713MCP
get_claim(id=1102713)