Original title: Interface-Induced Trajectory Censoring
Wenbo Wang · September 2026
When researchers test artificial intelligence agents, they measure how often the model calls external tools to solve problems. However, the software interface sitting between the model and the evaluation system can hide what the model is actually doing. The model might write out a perfectly valid tool call, but a mismatched interface drops it before any tool can run. The researchers investigated this problem by keeping the model, test tasks, and random seeds fixed while only swapping the serving adapterA component that translates between the model's output format and the evaluation system.. They found that this single interface swap made a model's score leap from 0.00 to 0.96 / 0.19 on one test, and jumped parsed calls from 0 to 636 on another. Larger models generated valid calls up to 80/100 times, yet the server still counted 0/100. The tool-call rateHow often the model successfully makes calls to external tools in an evaluation. does not belong to the model alone, but to the entire interface stack measuring it.
Imagine ordering a meal at a busy restaurant where the waiter writes down your request. Even if you specify exactly what you want, the waiter might fail to pass it on to the kitchen due to miscommunication. In this paper, the AI model sends out tool calls like food orders, but the interface—the serving adapter—can silently drop them, so the evaluation system scores zero even though the model did everything right.
This affects anyone evaluating AI agents, because reported success rates can be completely wrong due to interface issues, not model performance. It means tests might unfairly reject good models or accept bad ones.
The abstract does not claim that fixing the adapter always leads to significant performance gains in all tasks. It focuses on specific benchmarks like BFCL v4 and tau-bench, and the pass rate improvement was not statistically significant.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Interface-Induced Trajectory Censoring — Wenbo Wang
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.