ToolAnchor: Anchoring Counterfactual Context to Boost Agentic Tool-use Capability

Short summary

ToolAnchor addresses behavioral inertia in LLM agents— their tendency to favor familiar tools over newly introduced ones. The framework uses teacher models to generate counterfactual anchor contexts at critical decision points, verifies them via student rollouts, and internalizes successful interventions through agentic post-training. Evaluations across GAIA, BrowseComp, and VDR-Bench show competitive performance under expanded toolsets.

• •ToolAnchor tackles behavioral inertia where LLM agents prefer familiar tools over new ones
• •Teacher models hypothesize counterfactual contexts, verified via student rollouts and internalized through post-training
• •Competitive results across GAIA, BrowseComp, and VDR-Bench benchmarks under expanded toolsets

Generated with AI, which can make mistakes.