Jev 决策模型机制科普：并行类型化决策头

atulit @atulit_gaur原推发布于 2026.09.17

在 X 上打开原推

> meet Jev > not a chatbot, not an agent, and not a model generating paragraphs > give it a state plus typed questions > Choice, Score, or Noul > it returns an answer, probability distribution, and confidence > a Choice selects one option from a predefined set > a Score maps the state onto an ordered rubric > a Noul estimates whether a statement is true, with p ∈ [0,1] > unlike an autoregressive LLM, Jev doesn't generate t₁,t₂,...,tₖ sequentially > normal generation follows p(t₁,...,tₖ)=∏ᵢp(tᵢ|t₍<ᵢ₎,prompt) > Jev encodes the state once and evaluates multiple decision heads in parallel > question₁, question₂,...questionₙ ──► shared state encoding ──► typed outputs > each head scores only the legal outputs defined by the schema > no free-form vocabulary means no malformed JSON or invented categories > hundreds of choices can use a two-stage score-then-select pipeline > the second major piece is RLCD: Reinforcement Learning for Calibrated Decisions > an 0.8 confidence prediction should be correct roughly 80% of the time > probabilities become signals that ordinary software can actually branch on > confidence low → escalate > confidence high → execute > multiple narrow judgments replace one enormous reasoning prompt > reportedly reaches roughly 70–500 ms end-to-end latency > output tokens are unmetered because there is no prose generation > the underlying network, parameter count, and exact RLCD algorithm remain undisclosed > what is public is the interface: shared encoding, parallel constrained decisions, calibrated probabilities > the idea is simple > make AI decide > let code do the rest