Unverified50% confidenceFactExact time
NCP-ArchPreview proposes having language models model discrete concepts spanning multiple tokens in latent space through "Next Concept Prediction" (NCP).
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedNCP-ArchPreview predicts future concepts through a dedicated concept module and feeds the concept representations back to the token level to guide subsequent token generation.76% similarUnverifiedNCP-ArchPreview constructs a discrete concept vocabulary from the model's own hidden states via product quantization (PQ), requiring no external annotations.73% similarUnverifiedThe Coconut (Chain of Continuous Thought) research proposed by the MIT team allows models to perform multi-step reasoning in a continuous latent space without discretizing intermediate states into text.63% similarUnverified上下文偏置（Contextual biasing）在解码阶段介入，通过向模型提供期望词表提高特定词的输出概率，实现方式包括浅层融合、深层偏置和 TCPGen 等架构63% similarUnverifiedLatent space reasoning allows the model to complete thinking in a continuous vector space composed of hidden states, only decoding the result into text at the final stage.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/907836API
curl https://kongchang.com/api/v1/knowledge/claims/907836MCP
get_claim(id=907836)