Open-source model beats GPT-5 in strategy game with new RL method

Researchers have developed a novel reinforcement learning technique called delayed per-step reward attribution, designed to overcome challenges in training language model agents for complex multi-agent interactions. This method allows for rewards to be computed and propagated only at the end of an episode, excluding invalid steps and ensuring stable, sample-efficient training. When applied to the MindGames Arena benchmark, an 8-billion-parameter open-source model trained with this approach outperformed significantly larger proprietary systems, including GPT-5, securing first place in both open and efficient tracks. AI

IMPACT Demonstrates a new method for training AI agents in complex environments, potentially improving performance in multi-agent strategic interactions.

RANK_REASON Academic paper detailing a new reinforcement learning method and its performance on a benchmark. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Aliaksei Korshuk
• GPT-5
• In2AI
• MindGames Arena
• NeurIPS 2025

• paper
• model release
• other

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