Pushing the Limits of LLM Tool Calling via Experiential Knowledge Integration and Activation

A new framework called Knowledge-Augmented Tool Execution (KATE) aims to improve how large language models handle multi-step tasks by integrating experiential knowledge, according to a study submitted in 2026 [1]. The research finds that broadening reasoning paths, rather than deepening them, more effectively activates this knowledge [2]. Large language models (LLMs) increasingly rely on external tools to function as autonomous agents, but they frequently stumble during multi-step execution. The study identifies two core problems: insufficient tool-related knowledge and ineffective activation of the knowledge the models do possess [1][2]. The researchers systematically examined three stages: knowledge acquisition, activation, and internalization [2]. In the acquisition phase, they found that simple instance-level experiential knowledge provided strong and reliable performance gains, while abstract intent-level knowledge offered limited benefits [2]. At inference time, the team discovered that prompting an LLM to expand the depth of its reasoning yielded diminishing returns. Instead, expanding the width of reasoning—using parallel sampling combined with aggregation—proved far more effective at activating latent experiential knowledge [1][2]. For knowledge internalization during post-training, the study reports that reinforcement learning outperformed supervised fine-tuning when using knowledge-augmented data [2]. These insights were combined into the KATE framework, which integrates experiential knowledge with width-expanded inference and knowledge-aware training [1]. Experiments conducted on the BFCL-V3 and AppWorld benchmarks demonstrated consistent and substantial improvements over strong baselines across various model scales [1][2]. The code for the framework has been made publicly available on GitHub [2].