The latest public release of nokia-applied-research/anyjev is v0.3.0. This repository is a Python library that uses a language model like a typed decision model without training. The README explains how it answers choice, yes/no, and scoring-format questions with per-option probabilities included. The provided material does not list the specific changes in v0.3.0.
The library provides a basic Decider and a self-distillation style Tacit. Tacit makes a decision in a single forward pass and, depending on the configuration, can send some decisions to additional reasoning. The core is that instead of training a new model, it reads the probability of an option from the log probability of the token placed at the correct answer position. The L0 path rotates the option order, combines the obtained values in log space, and corrects them with a prior probability estimated without labels. This is a design meant to reduce the reversal that can depend on option order when logits are read directly.
The public leaderboard figures cited in the README come from a 2026-10-12 snapshot, where the project ranks 5th of 149 based on 300 samples. The README also notes that the coverage estimate at 5% risk has a large variance. It is therefore hard to conclude from this ranking alone that performance leads in other models or situations. The repository is licensed under Apache-2.0, and the material links an arXiv technical report, 2610.00831.