DeltaMerge-LowRes composes language and task deltas for low-resource adaptation

DeltaMerge-LowRes adapts multilingual encoders to new languages and tasks by training separate language and task deltas that are recombined in weight space. The method learns a language delta from unlabeled monolingual text and a task delta from labeled English data, composing them via four rules including a novel cross-axis TIES approach.

• Cross-axis TIES improves summarization chrF by +4 to +7 on three of four African languages compared to task-only baselines.
• The method increases QA F1 by +2.32 and EM by +2.91 across evaluated cells.
• Sparsity-aware merging reduces classification ECE by 36% while maintaining parity macro-F1.

The composition rule materially changes what the merged model preserves, suppresses, and calibrates, with all JSON traces and a claim ledger released.