Localizing RL-Induced Tool Use to a Single Crosscoder Feature

Reinforcement learning fine-tuning reshapes the internal representations of language models to enable agentic behaviors such as tool use, yet the mechanistic basis of these changes remains poorly understood. While RL substantially improves structured tool-call generation, it is unclear which features emerge, which are preserved, and whether identified features can be leveraged for retraining-free behavioral control. In this work, we show that Dedicated Feature Crosscoders (DFC) isolate a compact set of RL-specific features that causally mediate tool-calling capability in Qwen2.5-3B. Across a $48$-crosscoder hyperparameter sweep, encode-decode reconstruction improves the RL model's tool correctness by $+31.1 \pm {9.7}$\,pp and passively transfers tool-calling ability to the frozen base model by $+6.8 \pm 5.0$\,pp - capability spillover effect arising from tool-relevant signal leaking into shared decoder weights during joint training. Our findings show that DFC partitioning concentrates RL-introduced capability into a minimal, steerable feature set which enables precise retraining-free behavioral control of agentic LLMs.