Original title: DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent Training
Shubham Gandhi, Saurabh Goyal, Kiran Kate, Yara Rizk · September 2026
Training an AI agent across many steps is hard when there is no automatic checker to grade the final result. People often use a rubric to give one overall score, but that single number fails to show which exact steps were good or bad. To solve this, researchers built DRACO. It makes new rubricsLists of specific criteria used to evaluate a task, often applied at the end. on the fly as the agent improves during training. After scoring a finished run, DRACO uses a set math formula to assign credit directly back to the specific steps responsible for each rubric item, without needing an extra trained model. In tests on AppWorld, DRACO beat the base model by 15.9 points. It also beat standard training that actually had access to true reward signals.
Imagine you're coaching a soccer team through a long match, but you only know the final score. Deciding which players deserve praise or correction based on a single number is tough. Instead, you watch the game and use a checklist of key plays—like passes, shots, and defense—to evaluate each moment. In this paper, DRACO does something similar for AI training: it uses a checklist (rubrics) to judge the whole task at the end, then traces back which steps were responsible for each point on the checklist, allowing the AI to learn from specific actions.
Training AI agents for long, complex tasks usually requires a perfect reward signal, which is rarely available in real-world scenarios. DRACO makes it possible to train effective agents by judging outcomes with flexible rubrics, helping developers build systems for multi-step problems where success is hard to measure.
The abstract doesn't specify limitations, but it focuses on settings without ground-truth success signals and relies on dynamically generated rubrics which may require careful design.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
DRACO: Fine-Grained Credit Assignment with Dynamic Rubrics for Long-Horizon Agent Training — Shubham Gandhi, Saurabh Goyal, Kiran Kate et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.