New ARISE-RL framework for agent self-evolution using reinforcement learning

A research team published the ARISE-RL (Agentic Rubric-Grounded Iterative Self-Evolution with Reinforcement Learning) framework on arXiv, designed to train open agents using reinforcement learning. According to the authors, it addresses the lack of verifiable correct answers and scalable evaluation criteria (rubrics) for long, open-ended agentic tasks, which leads to unstable and weak rewards and degrades fine-grained optimization signals during group-based policy learning.

The framework combines two components in a process the authors call rubric-mediated co-evolution: Generator, which creates tasks and rubrics grounded in real tool observations and is rewarded for creating valid tasks of moderate difficulty matched to the current capability of Solver, and Solver, which learns from fine-grained signals of rubric fulfillment through multi-step reasoning and tool use.

The team also introduced Reward-Gated Self-Evolution Distillation (RG-SED) — a procedure that selectively distills a memory-augmented variant of the same policy back into the policy itself, but only when memory actually improves the reward. This is intended to reduce distribution mismatch and prevent blind imitation of noisy guidance.

For evaluation, the authors introduced ECR-Bench, a benchmark suite with expert-calibrated rubrics covering single-tool deep research and multi-tool travel planning tasks. According to the authors, ARISE-RL consistently achieves the best (state-of-the-art) results across all tested benchmarks; this is a claim by the study authors, and the source does not report independent verification. The source contains no information about the institution of the authors, code, a license, or planned availability.