ResearchBenchmarksMolecular-dynamics input-script authoring on LAMMPS 9-task transfer benchmarkFollow7.78LLM-judge Mean ScoreSIGA4.43125.30066.177.0394Jun 8, 2026Evaluation ResultsMethodMethodLinksLLM-judge Mean ScoreGainSIGABackbone=deepseek-v4-f...Backbone=deepseek-v4-flash, Best SIGA cell=M+R+S+X2026.067.783.22SIGABackbone=Claude Sonnet...Backbone=Claude Sonnet 4.6, Best SIGA cell=M+R2026.066.890.56VanillaBackbone=Claude Sonnet...Backbone=Claude Sonnet 4.62026.066.33—VanillaBackbone=deepseek-v4-f...Backbone=deepseek-v4-flash2026.064.56—