●builderIf you're fine-tuning multilingual models for reasoning tasks, SOLAR-style soft token alignment is a candidate auxiliary objective to reduce language-dependent answer inconsistency.
●researcherSOLAR targets the generation-stage divergence rather than representation-stage inconsistency, offering a new training signal for cross-lingual reasoning alignment.