IBM's DRACO: 15.9-Point AppWorld Gain via Dynamic Rubrics
TL;DR
- DRACO generates rubrics during training, scores them per trajectory, then redistributes that judgment across steps as differentiated GRPO advantages.
- On AppWorld it gains 15.9 points over the base model and 5.3 points over GRPO trained with a sparse ground-truth reward.
- On out-of-domain Tau-Bench it gains 5.3 points over the base model without any frontier judge.
IBM researchers say a rubric-based RL recipe for long-horizon agents can beat a verifier-trained baseline in the very setting where verifiers usually win. The paper, DRACO on arXiv, reports a 15.9-point AppWorld gain over the base model and a 5.3-point gain over GRPO trained with a sparse ground-truth reward, despite DRACO itself using no verifier.
The method, submitted September 3 by Shubham Gandhi, Saurabh Goyal, Kiran Kate and Yara Rizk, generates evaluation rubrics dynamically as the policy evolves, scores them once per completed trajectory, then splits that scalar judgment across the individual steps a rubric flags to produce per-step advantages in GRPO. The redistribution "is closed-form and does not introduce any trained attribution module," the authors write. On out-of-domain Tau-Bench, they report a further "5.3 points over the base model even without a frontier judge," beating both ground-truth-reward training and other rubric-based setups.
The paper lands in an unusually busy stretch for agent-RL credit-assignment work on our Agents tracker, which has logged 410 agent stories in the last 90 days. Neither the abstract nor the arXiv listing names the judge model used to grade rubrics, and results are reported on only two benchmarks. Code lives at github.com/IBM/draco.
Originally reported by arxiv.org
Read the original article →Original headline: IBM's DRACO Paper Redistributes Rubric Rewards Across Steps, Lifts AppWorld Agent Score 15.9 Points