简体中文 · English

Chinese-Jev: A Chinese System One decision model

L1 objective and model head. Candidate marker representations, candidate-probability supervision and RLCD-style calibration are implemented. The paper's model trained on 10 million examples was not trained locally.

Paper information

Field	Details
Paper	Chinese-Jev
Institution	Fudan University (first author affiliation)
First public release	2026-09-29 (arXiv v1)
Original source code	No standalone repository released; the project page says models, data and code are planned.
Local adapter/method	chinese-jev
Local reproduction code	src/auto_research/foundation_latest_20260930.py

Original paper summary

Background and main changes

Chinese-Jev targets closed-output tasks such as classification, routing and candidate selection. A lightweight encoder reads Chinese context and candidates and scores hidden states at candidate markers directly. Heterogeneous labels become candidate-probability distributions. Pretraining uses 10 million general examples before separate medical, legal and finance fine-tuning.

flowchart LR
  X[Chinese task and candidates] --> E[Encoder]
  E --> M[Candidate marker hidden states]
  M --> P[Softmax decision distribution]
  Y[Probability targets] --> CE[Cross-entropy]
  P --> CE
  P --> R[RLCD calibration]

Key figure from the paper

Figure 1 shows the composition of general Chinese decision data. From the paper; copyright belongs to the authors.

Core equations

Candidate \(i\) receives \(s_i=w^\top h_{m_i}\) and \(p_i=\mathrm{softmax}(s)_i\). The local objective combines soft-label cross-entropy \(-\sum_i q_i\log p_i\) with \(-\log p_i\) weighted by reward-centered advantage. It preserves candidate decisions rather than autoregressive generation.

Offline and online results

The paper reports 1.24% higher general-task accuracy than closed Jev, 20.3× speed, a 4.0% medical gain after domain fine-tuning and roughly 15ms mean latency. These are paper results. See local three-seed mechanism metrics.

Local reproduction and limits

ChineseJevHead runs forward and backward with probability normalization and finite-gradient checks. No paper data, full backbone, mobile INT8 deployment or paper-scale performance comparison.