Unverified50% confidenceFactExact time
Through self-supervised contrastive learning, RaRe aligns the CoT rationale embeddings generated by the MLLM with textualized relation descriptions, bringing the rationale representations close to the corresponding relation semantics in the semantic space.
1
Sources
50%
Confidence
Long-term
Relevance
11/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedMainstream text embedding models are typically trained based on contrastive learning, mapping semantically similar text pairs to nearby positions in vector space and pushing unrelated text pairs further apart.67% similarUnverifiedMasked language modeling (MLM) for text and masked autoencoders (MAE) for images share the "mask-predict" self-supervised learning paradigm.65% similarUnverifiedContextual biasing is a technique that introduces an external vocabulary or phrase list at the decoding stage to increase the recognition probability of specific word sequences by adjusting language model scores or using methods such as shallow fusion.64% similarUnverifiedScenario-based intent understanding relies on the semantic parsing capability of large language models to break down everyday-life descriptions into style tags, category combinations, and spatial relationship constraints, and usually requires intent slot filling working together with constraint reasoning to accomplish.64% similarVerifiedLLMs learn the statistical patterns of word co-occurrence, grammatical structures, and rhetorical patterns in human language by training on massive text corpora, but the generation process does not involve understanding of content or intent.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1158234API
curl https://kongchang.com/api/v1/knowledge/claims/1158234MCP
get_claim(id=1158234)