Improve final-answer accuracy of PTC-Decoder agents

Improve the final-answer accuracy of PTC-Decoder agents beyond the demonstrated step-level tool-adherence improvements, for which final-answer accuracy remains unresolved.

Background

PTC-Decoder is a training-free decoding framework for small LLMs deployed on offline, resource-constrained edge devices. It combines Plan-to-Act, which forces the agent to generate a plan before executing atomic tools, with TC-Decoder, which applies token-level hard constraints to tool names while leaving parameter generation unconstrained.

The reported experiments show statistically significant gains in overall performance and tool-execution reliability across seven small LLMs. However, the framework primarily constrains tool invocation and does not ensure that the tools are used with correct parameters or that the resulting information is synthesized into a correct final response. The paper therefore identifies final-answer accuracy as an unresolved challenge requiring further improvement.