Unverified50% confidenceFactExact time
Jev is a pure classifier that has not been trained to generate text; given a set of candidate options, it outputs a choice with a confidence score.
1
Sources
50%
Confidence
Long-term
Relevance
10/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedWhat Jev does is essentially multi-label classification rather than generative reasoning, achieved by attaching multiple classification heads in parallel to the same forward pass.77% similarUnverifiedSearch reranking is one application scenario for Jev: first use inexpensive algorithms such as BM25 to retrieve candidate results, then use Jev to score the candidates for relevance.71% similarUnverifiedJev's core mechanism is to extract and number the currently interactive elements into a list, having the model only answer what to do and which number to select, rather than guessing coordinates or writing selectors.69% similarVerifiedJev offers three question types: Select, Score, and Null/Bool, with each response accompanied by a confidence value.66% similarUnverifiedIn a document value extraction test, Jev accurately selected the correct new address from a list of candidates, provided that the candidate list contains the correct answer.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/985413API
curl https://kongchang.com/api/v1/knowledge/claims/985413MCP
get_claim(id=985413)