The Parts Are Greater Than the Sum: Automated Task Sequencing for Efficient Training of Multi-Policy LLMs
A multi-policy PEFT method groups and sequences compatible tasks into separate QLoRA paths, reducing interference without adding trainable capacity; it reached 44.78 on TRACE.
The proposed framework automatically groups and sequences heterogeneous tasks, then assigns compatible paths to independent **QLoRA adapters**. On TRACE, automatic multi-policy PEFT reached **44.78** under the **same trainable capacity** as the compared approaches.
For builders fine-tuning one model across varied workloads, adapter capacity may be less important than deciding which tasks share an optimization path. Separating incompatible tasks can limit interference while retaining transfer among compatible ones.
The proposed framework automatically groups and sequences heterogeneous tasks, then assigns compatible paths to independent **QLoRA adapters**. On TRACE, automatic multi-policy PEFT reached **44.78** under the **same trainable capacity** as the compared approaches. For builders fine-tuning one model across varied workloads, adapter capacity may be less important than deciding which tasks share an optimization path. Separating incompatible tasks can limit interference while retaining transfer among compatible ones. The supplied results cover the TRACE benchmark and report the best aggregate score, but not deployment cost or behavior outside that evaluation. It remains unclear how reliably the automatic grouping transfers to other task mixes.
This makes task compatibility and training order a first-class model-design choice: under fixed trainable capacity, separate adapter paths can preserve transfer among compatible tasks while limiting interference among incompatible ones. It reinforces specialization and upstream data composition as efficiency levers, but the single-benchmark aggregate result does not establish that its automatic grouping will transfer to other workloads or deployments.