RLAES trains one LLM to score essays and generate rubric feedback

A recurring problem with automated essay scoring is that it grades but does not teach. You get a number, not a paragraph explaining what the writer should fix. A new arXiv paper from Xuefeng Jin, Jiashuo Zhang, Teng Cao and Bin Yang proposes a reinforcement learning setup, RLAES, that trains a language model to do both at once, using rubric-based rewards so the feedback is graded against a checklist rather than judged only for fluency.

The measurement side is the more interesting contribution. The authors introduce Rubric-based Feedback Evaluation, or RFE, built around 166 fine-grained binary rubric items and an LLM-as-judge that scores whether each item is satisfied. Two other pieces make training tractable. Adaptive Gated Feedback Optimization activates the rubric rewards selectively to cut computational overhead, and Adjacent Contrastive Reasoning improves score calibration by contrasting adjacent scoring levels rather than treating every miss as equal.

On the standard ASAP benchmark, the authors report a quadratic weighted kappa of 0.803 for scoring, and claim feedback quality comparable to GPT-4.5 without the degradation typically seen when a score-only RL objective is bolted onto a feedback model. If that holds, it matters because most classroom and testing pipelines today either run a small scoring model with no explanation, or lean on a frontier API for the write-up. Doing both in one trained system is what makes it deployable at scale rather than only at cost.

The honest caveat is that this is a preprint, one lab, one benchmark, and the feedback quality itself is judged by another LLM, so the parity-with-GPT-4.5 claim rests on the judge being fair. The abstract does not disclose the base model size, how the 166 rubric items were built and validated, or whether the ASAP kappa was measured cross-prompt. Take the specifics as reported, not settled.

Where this gets interesting is downstream. Writing platforms and testing services have wanted rubric-grounded feedback for a decade, and a training recipe that jointly optimizes score and feedback, with a rubric-reward signal, is the sort of thing that could quietly turn up inside a scoring engine you already use.