RL-Induced Tool Use Localized to Single LLM Feature

Fine-tuning large language models (LLMs) using Reinforcement Learning (RL) is known to enhance agentic behaviors, such as the ability to use external tools. However, the precise internal mechanisms and representational changes that enable these capabilities have remained largely unexplored. This research demonstrates that RL-induced tool-calling capability can be localized to a compact and steerable set of features, termed Dedicated Feature Crosscoders (DFC). By isolating these RL-specific features in a Qwen2.5-3B model, researchers observed significant improvements in tool correctness, boosting it by over 31 percentage points. Crucially, this method also enabled a "capability spillover," passively transferring tool-calling ability to the frozen base model by nearly 7 percentage points without further retraining. This finding suggests that DFC partitioning effectively concentrates RL-introduced capabilities into a minimal set, paving the way for runtime behavioral control of agentic LLMs.