Title: ToolSelf: Achieving Unified Task Execution and Self-Reconfiguration through Tool-Driven Emergent Adaptation

Abstract:

While large language model (LLM)-driven agentic systems demonstrate proficiency in managing complex, long-horizon tasks, they are often limited by static configurations established prior to execution. This inflexibility necessitates a compromise between domain-specific expertise and cross-task versatility: highly specific priors and streamlined tool sets enhance specialization but hinder transferability, whereas generic workflows and expansive action spaces increase coverage at the cost of focused guidance. Current approaches—such as pre-execution optimization, planner-worker orchestration, and configuration patching—fail to adequately address this dichotomy because they separate adaptation from execution, resulting in data loss, disjointed optimization, and unclear credit assignment.

To resolve this, we introduce ToolSelf, a paradigm for runtime self-reconfiguration driven by tools. This approach standardizes configuration updates through a unified tool interface, effectively merging execution and adaptation into the action space of a single policy. This enables the execution agent to dynamically adjust sub-goals, strategies, toolboxes, context, and context-management modes in response to task progress and feedback. Additionally, we propose Configuration-Aware Two-stage Training (CAT), a method that integrates rejection sampling fine-tuning with trajectory-level KTO reinforcement learning to embed self-reconfiguration capabilities. Evaluated across various benchmarks, zero-shot ToolSelf performs competitively with task-specialized agents. Following CAT training, ToolSelf achieves an average improvement of 28.8 points over static-configuration baselines, demonstrating a pathway toward emergent adaptivity that eliminates the need for manually injected guidance.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC