Unverified50% confidenceFactExact time
Deletion Tuning is a post-training method that zeroes out a very small number of weights in the attention gate projection by learning a binary mask, without adding new parameters.
1
Sources
50%
Confidence
Long-term
Relevance
11/10/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedZero gating is a residual injection technique: the new branch's output is multiplied by a learnable gate coefficient initialized to 0 and then added to the main branch.68% similarUnverifiedThe Skill caching mechanism reduces token consumption by reusing learned skills to skip reasoning from scratch, thereby reducing the number of reasoning rounds and context length.68% similarUnverifiedMemory Steering is a lightweight inference-time intervention method that can restore the refusal rate to near-baseline levels without retraining, and operates entirely in the compressed representation space.67% similarUnverifiedThe Abliteration technique identifies and suppresses the direction vector responsible for refusal behavior in the model's residual stream, using orthogonal projection to remove that direction from the weight matrices, thereby eliminating refusal behavior without retraining.67% similarUnverifiedThe missing-input handling mechanism refers to a design whereby a multimodal model can still perform forward propagation and produce gradients when the input of a certain modality is unavailable, with common approaches including replacing that modality's embedding with a learnable missing token, zero-filling followed by masked attention to mask invalid positions, or introducing a modality-presence embedding.66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1153747API
curl https://kongchang.com/api/v1/knowledge/claims/1153747MCP
get_claim(id=1153747)