Skip to content
conftrace_
2026 EMNLP EMNLP 2026

LoRA-RG: Residual Gradient-Guided Rank Initialization and Expansion

AbstractWe study how to optimally allocate ranks in Low-Rank Adaptation (LoRA). Despite its importance, ranks are typically chosen manually and fixed across all modules. We argue that this practice is suboptimal, risking either under-parameterization, which hurts performance, or over-parameterization, which wastes memory and computation. To address this, we propose LoRA-RG, a gradient-guided framework that adjusts LoRA ranks through two stages: rank initialization and rank expansion. In the rank initialization stage, LoRA-RG estimates a minimal rank for each layer based on gradient information prior to training. During training, the rank expansion stage further increases ranks only when the current LoRA parameterization is insufficient to capture the remaining update demand. This design dynamically aligns LoRA ranks with the underlying optimization needs. We evaluate LoRA-RG on reasoning, mathematical, and code generation benchmarks, and show that it consistently achieves higher accuracy compared to fixed-rank LoRA and recent rank initialization and adaptation baselines under the same parameter budget.

More from EMNLP 2026

All EMNLP 2026 papers →