Xiaohongshu and Fudan University Build DR-Rubric for RL on Open-Ended Tasks

Xiaohongshu and Fudan University have presented DR-Rubric, a NeurIPS 2026 paper on building evaluation rubrics for reinforcement learning on open-ended tasks. Jiqizhixin covered the publication. The available excerpt describes the problem setup and contains no benchmark figures. So this piece looks at what problem the authors are tackling and what is worth checking in the paper.

Where the verifiable reward breaks down

Math is simple. There is a standard answer, the model either got it or did not, and the reward is computed automatically. A large share of recent RL successes in reasoning models rests on exactly this.

Open-ended tasks work differently. "Describe research progress in a field," "compare several technical routes," "give professional advice with justification": there is no single correct version. Errors are also hard to spot. A missing key fact, one concept swapped for another, or weak argumentation can easily hide in smooth, confident text. If a judge model hands out the reward based on overall impression, the policy quickly learns to write beautifully rather than accurately.

What a rubric is and why it is needed

The industry's common answer to this problem is to replace a single score with a set of specific criteria. A rubric breaks a "good answer" into checkable items. Are the key facts mentioned? Are terms used correctly? Is the conclusion backed by arguments? The grader answers narrow questions, and the final reward is assembled from those answers.

The weak point of this approach is the rubrics themselves. They have to come from somewhere, and each query needs its own. Written by hand, they are expensive and do not scale. Generated by a model, they may turn out shallow or repeat the model's own blind spots. Judging by the title and introduction, DR-Rubric works on exactly this link in the chain. The excerpt does not explain what the abbreviation stands for or how the rubrics are constructed.

What is known from the publication

• Authors: Xiaohongshu (RedNote) and Fudan University.
• Venue: NeurIPS 2026.
• Subject: building evaluation standards for RL on open-ended tasks.
• Task types: field overviews, comparisons of technical approaches, expert recommendations.
• Main difficulty, according to the text: deciding what counts as a good answer when there is no reference.
• Results, model scales and comparisons with baseline methods are not given in the excerpt.

What to check in the full paper

For teams that train models on open-ended tasks, a few questions determine how much the work matters.

• Source of the rubrics. Who creates them: humans, a model, or a combination, and what it costs per query.
• Quality of the criteria. How closely the rubrics match expert assessments and whether they distinguish plausible text from accurate text.
• Resistance to reward hacking. Whether the policy learns to tick items off formally, for example by mentioning the right words without substance.
• Transfer. Whether the rubrics work outside the domains they were built on.
• Task-level outcome. Whether they deliver gains on independent benchmarks, and not only on the authors' own checks.

The practical takeaway is simple. If you are running RL on tasks without an answer, do not rely on a single score from an LLM judge. Break the evaluation into atomic criteria, add penalty items for typical errors, and periodically check the grader against human annotations. Read the full DR-Rubric text with this checklist in hand: it quickly shows whether the method solves the problems of cost and rubric hacking or merely shifts them to another stage.