Chinese language models predict many-choice decisions in one pass
Chinese Decision: A Family of Language Models for Single-Pass Probabilistic Decision-Making
Computation and Language
Summary
Many apps need to pick the most likely answer from a fixed set of options instead of writing out long answers. The authors created Chinese Decision, a set of models that look at a question and possible answers once and then pick one by giving each option a probability. Their smaller 0.8 billion parameter model performs well on general and professional topics like medicine and law. They also tested a bigger 2 billion parameter model but it did not clearly improve accuracy in specialized areas yet. The models avoid generating answers word-by-word and instead choose answers in a single step, which speeds up decision-making.
What this means in practice
- •For chinese language software developers: Build apps that need fast, probabilistic choice among fixed answers in Chinese without slow text generation.
- •For financial risk analysts: Use single-pass models to classify or score options rapidly in finance-related decision systems.
Authors
Qiang Yan
Abstract
Many language-based applications require probabilities over a finite set of decisions rather than open-ended responses. We present Chinese Decision, a family of Chinese-language models for single-pass probabilistic decision-making. Given context, an instruction, and candidate descriptions, the model extracts final-position candidate-letter logits in one forward pass and normalizes them over valid options, without autoregressive answer generation. A shared interface supports Boolean judgments, categorical choices, and ordinal scores. We report a completed 0.8B model trained with distribution cross-entropy on 13.048 million decisions and preliminary 2B checkpoints trained with an additional RLCD objective. On the full CJ-Bench of 307,900 decisions, the uncalibrated 0.8B model achieves 72.24% general-domain accuracy and 76.98% average accuracy across medicine, law, and finance using a single checkpoint, with 1.07% overall expected calibration error. The 2B model reaches 72.56% general-domain accuracy after 40.13% of one epoch but has not surpassed the completed 0.8B model on the professional-domain average. We further evaluate a validation-temperature-scaled deployment model and its Q8 quantization, finding close aggregate performance while documenting individual prediction changes. We release the training, evaluation, and deployment pipeline. The current results establish a reproducible approach to Chinese probabilistic decisions but do not isolate the effect of model scale.