Template Gap Reconstruction for Shell Output Simulation with Small Language Models
Reusing shell-output structure cut generated tokens by 26.3–38.3%, with accuracy gains strongest on shorter gaps.
The method supplies fixed output text and asks small language models to generate the missing sections, without executing commands. The researchers built a template bank from 18.7 million outputs and evaluated reconstruction on an environment excluded from that bank. For shorter gaps, output edit similarity reached 0.964, compared with 0.812 for full-output decoding; performance was similar for longer sequences. Template selection and prediction of the remaining content still shaped response quality.
ArXiv · AI/CL/LG's note
The method supplies fixed output text and asks small language models to generate the missing sections, without executing commands. The researchers built a template bank from 18.7 million outputs and evaluated reconstruction on an environment excluded from that bank. For shorter gaps, output edit similarity reached 0.964, compared with 0.812 for full-output decoding; performance was similar for longer sequences. Template selection and prediction of the remaining content still shaped response quality.
ArXiv · AI/CL/LG's note
score 4