Nano replica of Jev featuring parallel decisions, dynamic candidate sets, and an end-to-end training and evaluation pipeline.
Evaluates multiple questions and dynamic candidate spaces concurrently in a single forward pass, logging navigation choices.
Provides open-source training infrastructure and game environment benchmarks for investigating non-generative discrete decisions.
Source code verified against pinned commit. Complete training and benchmark suite; generalization depends on dataset diversity.
Inspect a research implementation of dynamic candidate decisions without output-token decoding.
Material analysed：
The Chinese README describes a shared scalar/set-attention Choice head, a sigmoid Boolean path and a weighted expectation for Score. They should not all be interpreted as one argmax label.
Inspect the supporting source ↗evaluate_live_jev_navigation.py communicates with a worker and waits for responses. The source provides a comparison mechanism, not independently verified navigation performance.
Inspect the supporting source ↗Consider it for architecture and training experiments with dynamic options. Check the documented input roadmap before expecting full native TypeSafe structured-instruction or Noul compatibility.
Compare README input types with its roadmap, then trace the worker protocol in evaluate_live_jev_navigation.py. Separate local model runs from hosted calls in reports.