Unverified60% confidenceFactExact time
Adapter was proposed by Houlsby et al. in 2019; it inserts small bottleneck modules after the feed-forward network in each Transformer layer, and only the inserted modules are trained during fine-tuning.
2
Sources
60%
Confidence
Long-term
Relevance
7/13/2026
First Seen
Sources
BERT新增分类类别：增量微调的三种策略与灾难性遗忘解法
Related Claims
UnverifiedThe black-box usage of directly calling high-level APIs such as nn.Transformer encapsulated in PyTorch becomes a shortcoming when debugging models or improving architectures.64% similarUnverifiedLRCC trains a lightweight router for each Transformer block to dynamically select computation channels per token among nested low-rank paths.61% similarUnverifiedThe plugin-based split means upgrading a plugin does not touch the gateway core, and explicit plugin declarations make enabled capabilities and execution order visible to operations personnel.57% similarUnverifiedThe Transformer architecture needs to maintain Key and Value tensors for already-processed tokens during autoregressive generation.56% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/502641API
curl https://kongchang.com/api/v1/knowledge/claims/502641MCP
get_claim(id=502641)