Readout mistakes can mislead LLM tool selection

People who build “tool-using” AI systems often focus on whether the model can spot the right option. But a recent study suggests the bigger failure point can show up later—after the model has already looked at the tools.

The paper examines a benchmark called BFCL. In the results, the correct tool ended up with the highest attention about 80% of the time. Yet the tool that was ultimately selected had only a 21% chance. That mismatch matters because it points to a system behavior that isn’t just “the model can’t see the right tool.” Instead, something changes between noticing a tool and committing to it.

The authors also report that the “gold” (correct) tool was under-attended in only about 10% of failures. Put plainly, many errors aren’t explained by the model ignoring the right choice in its attention signals. That pushes the likely culprit toward the final readout step—where the model translates what it has processed into an actual tool decision.

They compared different fixes. Prompt repairs helped only up to 23%. By contrast, interventions aimed at the readout improved results much more, with gains reported in a range from 59% to 91%. Another approach—a better selector—also moved the needle, raising BFCL by 11.9 points and Seal-Tools by 14.9.

For builders, the practical takeaway is simple: improving the model’s “tool awareness” may not be enough if the last step is where the choices go wrong. If the system’s final output can override or distort what the model already attended to, then debugging has to reach beyond the tool-selection stage.

Source: arXiv