Why do AI agents keep failing at complex tasks? The answer lies in a hidden flaw in how they learn: a disconnect between thinking and doing. Researchers have identified this "Thinking-Acting Gap" as the primary reason multimodal models struggle with real-world problems requiring external tools.

Current models treat reasoning and tool use as separate behaviors. When trained with standard methods, tool use is attempted in only 30% of cases. Worse, when tools are used, they fail 40% of the time, effectively silencing the learning signal exactly when it is needed most. This creates a cycle where the model never learns to use tools effectively because it rarely gets useful feedback from its mistakes.

The team introduces AXPO, a new optimization method that fixes this imbalance. Instead of treating all errors equally, AXPO targets the specific cases where tool use failed. It keeps the model’s initial reasoning intact but resamples the tool call and its outcome, paired with a smart selection process for the starting context. This approach forces the model to learn from its tool-use failures rather than ignoring them.

The results are striking. On nine multimodal benchmarks, the new method outperforms standard training by nearly 2 percentage points. Most impressively, a smaller 8-billion parameter model using AXPO surpassed the performance of a much larger 32-billion parameter model. This proves that smarter training can beat raw size, offering a more efficient path to capable AI agents.