Unverified50% confidenceFactExact time
DASA leverages activation gradient feedback from a frozen reference model to guide the optimization of continuous synthetic input embeddings.
1
Sources
50%
Confidence
Long-term
Relevance
9/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe representative work of soft prompt techniques is Prompt Tuning proposed by Lester et al. 2021, which adds trainable continuous vectors only at the front of the input sequence while keeping model parameters frozen.69% similarUnverifiedApproaches that fall between frozen embeddings and full fine-tuning include fine-tuning only the last few layers, using parameter-efficient fine-tuning methods such as LoRA, or employing an adapter architecture.69% similarUnverifiedDASA experiments are conducted under a unified LoRA adaptation setting.66% similarUnverifiedThe human review and record-keeping mechanism provides labeled feedback data for subsequent model iteration.66% similarUnverifiedA frozen model refers to a model whose weight parameters all remain unchanged during the inference phase, with no fine-tuning, adapter training, or online learning performed.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/961002API
curl https://kongchang.com/api/v1/knowledge/claims/961002MCP
get_claim(id=961002)