Multi-Target Few-Shot Prompting with Diffusion Language Models
Examples improved accuracy when diffusion language models filled multiple masked regions in the same text.
The method uses demonstrated edits to guide changes across several locations, such as formatting updates or API migrations. It addresses edits that depend on context on both sides of each gap, which the authors describe as awkward for causal autoregressive models. The authors also derive a likelihood-ratio reward through selective masking to favor fills supported by the examples. Across few-shot benchmarks and software refactoring case studies, demonstrations improved accuracy, with the additional guidance delivering further gains in several settings.
ArXiv · AI/CL/LG's note
The method uses demonstrated edits to guide changes across several locations, such as formatting updates or API migrations. It addresses edits that depend on context on both sides of each gap, which the authors describe as awkward for causal autoregressive models. The authors also derive a likelihood-ratio reward through selective masking to favor fills supported by the examples. Across few-shot benchmarks and software refactoring case studies, demonstrations improved accuracy, with the additional guidance delivering further gains in several settings.
ArXiv · AI/CL/LG's note
score 5