DRACO: Closed-Form Rubric Redistribution for Outcome-Blind Agent Credit Assignment

DRACO is introduced as a fine-grained credit-assignment method for long-horizon agent training in the outcome-blind setting, where no programmatic verifier or ground-truth success signal is available. It trains with GRPO by generating multi-criteria rubrics dynamically, scoring each completed trajectory once, and redistributing that judgment over responsible steps in closed form. On AppWorld, the preprint reports a 15.9-point gain over the base model and a 5.3-point gain over GRPO trained with a sparse ground-truth reward, despite not using a verifier itself. It also reports an out-of-domain Tau-Bench gain of 5.3 points over the base model even without a frontier judge, beating both ground-truth-reward training and other rubric-based training settings.¹

Context

The method is motivated by the observation that reinforcement learning from verifiable rewards works when a task has a programmatic checker, but most long-horizon agent domains do not. In the absence of success labels, multi-criteria rubrics are a common reward source, yet scoring a rubric only once per trajectory gives a weak scalar signal across many steps. DRACO generates rubrics adaptively so that the reward surface tracks the policy's evolving capability, while its closed-form redistribution allows differentiated per-step advantages without introducing a trained credit-attribution module.¹

Assessment

The key news is that credit assignment need not inherit the usual constraint of a verifiable reward: a trajectory-level judgment captured as a rubric can be converted into step-wise training advantages through a deterministic closed-form rule. Making rubric quality dynamic during training is also substantively new because it lets the reward adapt as the policy changes, which a fixed rubric cannot do. The AppWorld and Tau-Bench results suggest that this substitution can be competitive with sparse ground-truth reward, though the size and reliability of that advantage cannot be assessed from the abstract alone.¹

Caveats

The present source is an abstract rather than a full paper, so no variance, hyperparameter sensitivity, or baseline detail is available. I would revise the central claim downward if static rubrics matched the reported results, if the closed-form redistribution step added no benefit over a single scalar trajectory score, or if the Tau-Bench result depends on a specific unseen judge. Conversely, code release and ablations that separate dynamic rubric generation from credit redistribution would substantially increase confidence in the mechanism.¹