EVENT DOSSIER
A study on the BFCL v4 dataset found that large language models may experience “trajectory review” during reasoning due to issues with the design of service layer interfaces. Specifically, although the model generates grammatically correct tool invocation requests, they are intercepted due to interface restrictions during execution, resulting in zero tool invocations. Experiments showed that simply replacing the service adapter could significantly reduce the evaluation score from 0.96/0.19 to 0.00; moreover, by swapping the chat templates and parsers, it was confirmed that the problem stemmed entirely from the interaction layer interface mechanism.
Agent evaluations report a tool-call rate read off the serving stack, which can be zero while the model emits well-formed calls due to interface-induced trajectory censoring. On BFCL v4 data, changing only the serving adapter causes scores to drop from 0.96/0.19 to 0.00; a 2x2 swap on chat templates and parsers locates this effect entirely in the interactio…