Previous work argued that when an agent only needs to choose from a bounded set of options, generating an answer token by token is unnecessary. This article tests seven open decision models, called Jev-style models, against a generative LLM baseline on banking requests. These models score the available options and return a typed decision, such as a support route or tool selection, instead of writing out an answer.
The benchmark covered both fixed route lists and changing candidate menus, which matter because agent tools and actions often change with the workflow. Kev reached 98% accuracy on both fixed routes and changing candidates. Verdict reached 89% on changing candidates, compared with 66% for the Qwen3-0.6B baseline, which scored 83% on fixed routes. JSON generation still beat the three Qwen readout configurations.
The decision models also generated zero answer tokens, while Qwen JSON generation produced about 10 per request. Median request latency was 125–130 ms for Kev and 17–18 ms for Verdict, compared with 331–357 ms for the LLM baseline across the two panels.
The authors frame candidate scoring as a more flexible alternative to classifiers with a permanent output slot for each route, which need retraining when new routes are added. A starter kit and setup guides for each project are available for readers who want to test the approaches themselves.
Source: Towards Data Science · Summarized by HeadlinesBriefing