AI Models Learn Strategic Tool Use Beyond Simple Timing, Boosting Math Problem Accuracy

Researchers studying tool use in large reasoning models have discovered a critical gap: existing approaches focus narrowly on when to invoke external tools, but largely ignore how those tools should actually be used. This oversight leads to failures even when the underlying reasoning logic is sound.

Two Distinct Tool-Usage Patterns#

The work identifies two contrasting strategies that emerge naturally when models integrate code-based reasoning. A calculator pattern treats code as a direct execution engine—performing arithmetic and numerical operations step-by-step. By contrast, an algorithmic pattern translates the problem itself into executable code, essentially encoding the solution strategy as a program. Neither approach is universally superior; the challenge lies in recognizing which pattern suits each task.

A Two-Stage Training Method#

The research proposes a framework that tackles tool-use optimization in sequence. First, it builds the model's foundational competence with both patterns, ensuring it can employ each effectively. Second, it fine-tunes pattern selection by aligning the model's choices with preferences demonstrated by expert demonstrations or human feedback.

Significant Performance Gains#

Testing on demanding mathematics benchmarks reveals substantial improvements. On the MATH500 dataset, the pattern-aware approach raised the rate of correct code usage (Code@1) from 64.0% to 70.5%. Even more striking, performance on AIME24—a notoriously difficult competition-level exam—jumped from 26.7% to 50.0%. These gains underscore how teaching models which tool strategy to apply can dramatically enhance problem-solving accuracy and reliability.

The work appeared on arXiv as a submission to the computer science literature.