Further reading

The parameter-efficient lineage begins with neural adapters [ Houlsby et al. , 2019 ] and LoRA [ Hu et al. , 2022 ] . AdapterFusion [ Pfeiffer et al. , 2021 ] , LoRAHub [ Huang et al. , 2024 ] , and input-aware retrieval [ Zhao et al. , 2024 ] study composition and routing rather than a single isolated adapter.

Model-merging work exposes the interference problem from several angles: task arithmetic [ Ilharco et al. , 2023 ] , TIES merging [ Yadav et al. , 2023 ] , SVD-based LoRA merging [ Tang et al. , 2025 ] , LoRA soups [ Prabhakar et al. , 2025 ] , and orthogonal-subspace control [ Zhang and Zhou , 2025 ] . Recent work on cross-task LoRA interference [ Zhang et al. , 2025a ] is especially close in motivation. ALLORA contributes a different diagnostic: the Lie bracket measures order-sensitive interaction between adapter fields, while its admission contract retains capability and safety checks that purely algebraic merging rules may not supply.