Localizing RL-Induced Tool Use to a Single Crosscoder Feature
📰 ArXiv cs.AI
arXiv:2606.26474v1 Announce Type: cross Abstract: 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{Dedica
Full Article
Title: Localizing RL-Induced Tool Use to a Single Crosscoder Feature
Abstract:
arXiv:2606.26474v1 Announce Type: cross Abstract: 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{Dedica
Abstract:
arXiv:2606.26474v1 Announce Type: cross Abstract: 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{Dedica
DeepCamp AI