Community · E2 · artifact verified
Use a decision model as an agent world model
A bilingual study asking whether JEV can serve as a world model for LLM agents - predicting what happens next from typed state questions - benchmarked against generative LLMs on identical predictions, with run manifests and a phase-0 API check archived.
01 · Role in the system
What Jev does here
Agent stacks usually ask a generative LLM to imagine the consequences of an action. This study tests a different shape: the current state becomes typed questions and JEV answers with calibrated probabilities - what will the environment return, will this action succeed - no generation at all. The same prediction tasks go to generative LLM baselines for comparison, and the repo archives the methodology: a run manifest per experiment, a phase-0 script that checks the live API before burning tokens, and a dedicated Jev client. README is bilingual (English and Korean), reflecting where the agent-research question came from.
02 · Control boundary
Where Jev sits
Jev as the world-model layer of an agent stack: state descriptions become typed prediction questions answered with calibrated probabilities, benchmarked head-to-head against generative-LLM imagination on identical tasks.
Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.
03 · Known limits
What this evidence does not prove
- Research scope; the world-model role is tested on bounded prediction tasks, not open-ended environments.
- 0 stars; single-author study with Korean and English docs.
- Baselines are the author's generative-LLM runs, not third-party replications.
04 · Attribution
Public sources
This is a Community record: the project was published by a third-party community author.
- thddydgnl ↗Community · github · public · checked 2026-10-08