ResearchBenchmarksContinual Instruction Tuning on MMLU, ScienceQA, GSM8K, and HumanEvalContinual Instruction Tuning on MMLU, ScienceQA, GSM8K, and HumanEvalFollow74.6Average AccuracyHiP-LoRA71.68872.44473.273.956Apr 20, 2026Evaluation resultsMethodMethodLinksAverage AccuracyForgetting RateHiP-LoRAconfig=fullconfig=full2026.0474.66.9PiSSArank (r)=16rank (r)=162026.0473.110.2LoRArank (r)=16rank (r)=162026.0471.813.6