EDU-Jev: Training System-One Decision Models for Education

EDU-Jev is a family of single-pass System One decision models for education tasks such as assigning a rubric score, checking whether an answer meets a marking criterion, or identifying the first wrong step in a solution. It targets fixed-option decisions where generative models require parsing, cost more at longer outputs, and express confidence in words that may not match correctness probability.

Given an assessment context and a question with predefined answer options, the model returns a probability distribution over those options rather than emitting text. A pointer head scores a marker for each option, so predictions are restricted to the allowed choices.

Across twelve education benchmarks, EDU-Jev 4B achieves an average accuracy of 67.6%. That exceeds the strongest of five commercial decision APIs by 9.4 percentage points and the best open 4B decision model by 18.7 points.

The 2B model processes about 150 decisions per second on a single H100 GPU. The paper reports this throughput as part of the single-pass design, which avoids the longer outputs that raise inference cost in generative approaches.

Training uses a single supervised stage with soft cross-entropy against the available target label distributions. Calibration is not assumed from the training objective; the paper assesses it empirically.

On four cross-domain education sets, EDU-Jev matches the strongest API. That API remains ahead on general decision-making and reasoning tasks, so EDU-Jev's advantage is specific to the education decision setting it was trained for.

The family includes 0.8B, 2B and 4B parameter variants. The smaller sizes are aimed at lower-cost, higher-throughput deployment, while the 4B model carries the reported benchmark results.