Your Language Model Is Already a Decision Model (github.com)

A recent analysis reveals that the Qwen3.5-9B language model functions effectively as a decision model, capable of selecting actions based on its next-token probabilities without requiring decision-specific training. Tested alongside Jev 1.13.0, Qwen3.5-9B demonstrated competitive accuracy across various benchmarks, such as achieving 80.5% on JevBench and performing well in tasks related to phishing and decision-making scenarios. Notably, Qwen3.5-9B consistently outpaced its predecessor, Qwen3-8B, across multiple benchmarks, highlighting its enhanced capabilities without the need for specialized training datasets. This finding is significant for the AI/ML community as it challenges the traditional view that decision models must undergo distinctive training. By employing a language model's inherent capabilities, researchers can streamline decision-making processes, potentially reducing time and computational resource requirements. The study presents a novel approach to integrating language models into decision-making architectures, suggesting that the interplay between states and candidate actions can effectively utilize existing models, paving the way for advancements in automated decision-making and multi-task learning scenarios.