Localizing RL-Induced Tool Use to a Single Crosscoder Feature
For researchers studying mechanistic interpretability of RL fine-tuning in LLMs, this work provides a method to localize and control agentic behaviors without retraining.
The paper introduces Dedicated Feature Crosscoders (DFC) to isolate RL-induced features mediating tool use in Qwen2.5-3B, achieving a +31.1 pp improvement in tool correctness via encode-decode reconstruction and a +6.8 pp capability spillover to the frozen base model.
Fine-tuning through RL 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 $\textit{Dedicated Feature Crosscoders (DFC)}$ isolate a compact set of RL-specific features that mediate tool-calling capability in $\texttt{Qwen2.5-3B}$. 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 which we call a $\textit{capability spillover}$. Our findings show that DFC partitioning concentrates RL-introduced capability into a minimal, steerable feature set that enables runtime behavioral control of agentic LLMs.