Unverified50% confidenceFactExact time
NoteBank's core retrieval results require no training, but the paper optionally trains a lightweight LoRA adapter to help the model exploit replayed evidence in downstream tasks, with the adapter enabled only during the replay and answering stages.
1
Sources
50%
Confidence
Long-term
Relevance
10/15/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedExperiments show that DoRA outperforms standard LoRA on multiple tasks while having a comparable number of trainable parameters69% similarVerifiedAfter LoRA training is complete, the low-rank matrices can be merged back into the original weights, introducing no additional latency during inference.69% similarUnverifiedThe student model in 2.0 training cannot see the hidden instructions, thereby internalizing the behavior rather than relying on prompt triggers; the backbone is the same 27B model with a second LoRA stacked on top and merged.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1040236API
curl https://kongchang.com/api/v1/knowledge/claims/1040236MCP
get_claim(id=1040236)