Unverified50% confidenceFactExact time
RCD is a training-free decoding-stage method that amplifies tokens whose probability is boosted by the rationale by contrasting the next-token distributions with and without the rationale.
1
Sources
50%
Confidence
Long-term
Relevance
10/22/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedRCSD (Reinforcement and Continual Learning for Domain-Specific Speculative Decoding) is a training framework that directly aligns draft model optimization with a system-level objective (maximizing acceptance length).72% similarUnverifiedThe RCD method has boundaries: what it amplifies is the reasoning signal already encoded in the model itself; if the model has no usable reasoning knowledge for a certain type of problem, contrastive decoding can hardly conjure a correct path out of thin air.69% similarUnverifiedImplementations of semantic ID technology typically first use a pretrained model (such as BERT or CLIP) to extract item semantic vectors, then discretize the continuous vectors into multi-level code sequences using techniques such as Residual Quantization VAE (RQ-VAE).66% similarUnverifiedGradient Confusion Defense (GCD) is a wrapper layer that requires no retraining and takes effect at test time, directly rewriting the probability vector the model returns to the querying party.63% similarUnverifiedRLCD constructs positive and negative sample pairs for contrastive learning, combines reinforcement learning signals to distill model output distributions, and has advantages in scenarios that do not require manually annotated rewards.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1084517API
curl https://kongchang.com/api/v1/knowledge/claims/1084517MCP
get_claim(id=1084517)