New ARISE-RL framework enhances reinforcement learning for open-ended agents

Researchers have introduced ARISE-RL, a new framework designed to improve the training of open-ended agents using reinforcement learning. This framework addresses challenges like the lack of verifiable answers and scalable rubrics by coupling a task/rubric generator with a reasoning solver. ARISE-RL employs a co-evolutionary approach where the generator creates tasks based on tool observations, and the solver learns from rubric satisfaction signals. The system also incorporates Reward-Gated Self-Evolution Distillation (RG-SED) to refine policies and reduce imitation of noisy guidance. To facilitate evaluation, the researchers developed ECR-Bench, a benchmark suite for deep research and multi-tool planning tasks, demonstrating ARISE-RL's state-of-the-art performance. AI

IMPACT This framework could lead to more capable and robust open-ended AI agents, improving performance on complex reasoning and planning tasks.

RANK_REASON The cluster contains a research paper detailing a new framework and benchmark for AI agents. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• ARISE-RL
• arXiv
• ECR-Bench
• Hugging Face
• reinforcement learning

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →