Scientific manuscript generation using large language models

A method, comprising:

• receiving, by one or more processors coupled to memory, an input including at least one machine-readable data object including resources divided by one or more input node identifiers, wherein each of the one or more input node identifiers corresponds with a location of the resources within the at least one machine-readable data object;
• generating, by the one or more processors, a multi-level resource mapping encoding associations between the one or more input node identifiers and one or more output node identifiers;
• constructing, by the one or more processors, a set of model input parameters for each of the one or more output node identifiers, each set of model input parameters based on a selection of the resources from the at least one machine-readable data object, the selection based on an association with one or more corresponding input node identifiers encoded by the multi-level resource mapping;
• providing, by the one or more processors, each set of model input parameters to a first generative model to generate an output data object for each of the one or more output node identifiers;
• generating, by the one or more processors, at least one metric for each output data object, the at least one metric based on a comparison of the output data object to a corresponding structural or coverage parameter;
• in response to the at least one metric failing to satisfy the corresponding structural or coverage parameter, automatically generating, by the one or more processors, a modified set of model input parameters to regenerate the output data object;
• and
• generating, by the one or more processors, based on a schema of a word processing application, a function that causes the word processing application to create an annotation for each output data object by (i) identifying, using a second generative model, a textual or graphical representation presented in a first region of a graphical user interface of the word processing application for each output data object and (ii) populating a body of the annotation with an indicator linking each textual or graphical representation to at least one of the one or more input node identifiers, the annotation configured to be presented in a second region of the graphical user interface of the word processing application, wherein the indicator presents the location of the resources on the word processing application via the annotation.

A non-transitory computer-readable storage medium having one or more instructions stored thereon, the one or more instructions executable by one or more processors to:

• receive an input including at least one machine-readable data object including resources divided by one or more input node identifiers, wherein each of the one or more input node identifiers corresponds with a location of the resources within the at least one machine-readable data object;
• generate a multi-level resource mapping encoding associations between the one or more input node identifiers and one or more output node identifiers;
• construct, for each of the one or more output node identifiers, a set of model input parameters, each set of model input parameters based on a selection of at least a portion of the resources from the at least one machine-readable data object, the selection based on an association with one or more corresponding input node identifiers encoded by the multi-level resource mapping;
• provide each set of model input parameters to a first generative model to generate an output data object for each of the one or more output node identifiers;
• generate at least one metric for each output data object, the at least one metric based on a comparison of the output data object to a corresponding structural or coverage parameter;
• in response to the at least one metric failing to satisfy the corresponding structural or coverage parameter, automatically generate a modified set of model input parameters to regenerate the output data object;
• and
• generate, based on a schema of a word processing application, a function that causes the word processing application to create an annotation for each output data object by (i) identifying, using a second generative model, a textual or graphical representation presented in a first region of a graphical user interface of the word processing application for each output data object and (ii) populating a body of the annotation with an indicator linking each textual or graphical representation to at least one of the one or more input node identifiers, the annotation configured to be presented in a second region of the graphical user interface of the word processing application, wherein the indicator presents the location of the resources on the word processing application via the annotation.