- Published on
Turning GLM-5.3-Flash into a Jev-like decision model
- Authors
- Name
- DeepHorus
GLM-5.3-Flash has been turned into a Jev-like “System One” model that selects predefined options and returns their probabilities in one forward pass. Johannes Hötter and Marko Rosenmüller, PhD, report performance on par with TypeSafe’s Jev on a public-data benchmark. The setup also enables typed image decisions, which the authors say Jev cannot handle.
Software often asks LLMs categorical questions, such as which team should handle a ticket or whether a contract clause belongs under liability. Standard prompting can produce JSON, but generating the full object—and hundreds of reasoning tokens first—can add cost and latency. It also leaves confidence unavailable unless explicitly requested.
Jev and Laya are specialized models that accept a state and named options, then return a choice and probability for each. The team tested whether an off-the-shelf LLM could reproduce those properties without fine-tuning. It could.
The prompt numbers each option and supplies the state, question and options as JSON. It ends with “choice_index:”, making the next token an option index rather than the start of a complete object. The system reads probabilities for every valid index at that one position, normalizes them and selects the highest-scoring answer.
Researchers ran GLM-5.3-Flash on vLLM in Privatemode. They used /chat/completions with continue_final_message and add_generation_prompt set to false, continuing a prefilled assistant turn. Images can accompany the text.
On the benchmark, the system matched Jev’s decision accuracy, correctness and speed. This could benefit high-volume software that needs structured choices and confidence values without lengthy generated responses. It shows a shipped LLM can perform a dedicated decision role without task-specific training.
The approach requires evaluation across other models and workloads; no numerical accuracy or latency results are provided. Its significance is faster, probability-aware decisions with a familiar LLM. Multimodal support could expand the architecture beyond text.
