A new reinforcement learning method fixes a subtle flaw in how AI agents learn from trial and error.
Most agent training methods, including the widely used GRPO, judge a multi-step task by a single final score. If an agent uses a tool, searches for information, and then reasons its way to an answer, the whole sequence gets one reward number, leaving the optimizer to guess which step actually mattered. Researchers propose Reinforcement Learning with Decomposed Subtasks (RLDS), which instead splits the reward across a fixed set of subtasks and assigns credit to the specific steps that drove each one. Tested on four agent benchmarks, the approach gained 11.5 points on ScienceWorld and 9.8 points on FrozenLake, while showing no meaningful change on HotpotQA and DeepResearch.
That split matters because it tells you where decomposed credit is worth the trouble: tasks that blend distinct, unrelated skills benefit, while tasks built from one repeated skill do not. ScienceWorld also ran 10.9 percent faster per training step under this method, since its long rollouts spread the extra reward-splitting overhead across more work.
The catch is that this is not a universal upgrade, it is a targeted one. The gains show up exactly where the researchers' own diagnostics predicted they would and vanish where they did not, which is a useful honesty check but also a reminder that decomposing rewards adds complexity that only pays off for a specific kind of messy, multi-skill task.