Multiplicative LoRA Weights

• Multiplicative LoRA weights are dynamic scaling factors for low-rank adaptation that modulate base model weights or adapter updates for enhanced transfer learning.
• They employ instance-, module-, and rank-level policies, including per-token fusion gates, to achieve stable gradients and improved performance.
• Dynamic scaling enables precise, context-aware integration of multiple LoRA modules, leading to higher accuracy in both classification and generative tasks.

Multiplicative LoRA weights extend the low-rank adaptation (LoRA) framework for parameter-efficient fine-tuning of large-scale deep learning models by introducing dynamic, explicit multiplicative scaling factors. These factors modulate either the base model weights, the adapter updates, or the fusion of multiple LoRA modules. Unlike the standard additive approach, multiplicative LoRA weights enable finer control over the contribution of pre-trained model components and their adapters during transfer learning, and they address theoretical and empirical weaknesses of fixed or improperly-scaled adaptation. Multiplicative schemes encompass instance-level, module-level, and rank-based scaling policies, as well as dynamic per-token fusion gates for multi-LoRA combination.

1. Multiplicative LoRA Weight Formulations

Multiplicative LoRA weights are applied in distinct settings, with three principal variants established in recent work:

1. Base Weight Scaling (α-LoRA): Each row (or scalar, or per-layer block) of the pre-trained base matrix $W$ is scaled by a trainable parameter $\alpha$; the LoRA update becomes $W' = \alpha \circ W + AB$, where $AB$ is the standard low-rank LoRA adapter ($A \in \mathbb{R}^{d_{\text{out}} \times r}$, $B \in \mathbb{R}^{r \times d_{\text{in}}}$). For LLMs, the row-wise form $W'_{i,:} = \alpha_i W_{i,:} + (AB)_{i,:}$ is common. This reparameterization introduces negligible parameter and compute overhead since $$\#(\alpha) = d_\text{out} \ll d_\text{out}d_\text{in}$$ (Firdoussi et al., 24 Oct 2025).
2. Adapter Rank Scaling (rsLoRA): The LoRA additive update is modified by a deterministic rank-dependent factor $\gamma_r$. The original LoRA sets $\gamma_r = \alpha / r$, but theoretical analysis shows optimal stability and learning at $\alpha$0. Thus, the effective weight is $\alpha$1, termed rank-stabilized LoRA (rsLoRA) (Kalajdzievski, 2023).
3. Dynamic Fusion Scaling (LoRA-Flow): For combining multiple pre-trained LoRA adapters, a dynamic gate outputs per-token, per-layer multiplicative weights $\alpha$2 for each LoRA module $\alpha$3 at decoding step $\alpha$4 and transformer layer $\alpha$5. The final output at step $\alpha$6 is $\alpha$7, where the fusion weights are obtained via a softmax gate conditioned on the current hidden state (Wang et al., 2024).

2. Theoretical Rationale for Multiplicative Scaling

Base Weight Rescaling: α-LoRA and RMT Analysis

The α-LoRA formulation addresses the mismatch between pre-trained and target tasks in low-resource or partially aligned transfer. Random Matrix Theory (RMT) provides a formal analysis in the high-dimensional binary classification setting: For a source classifier $\alpha$8 and target data $\alpha$9, the fine-tuned classifier is $W' = \alpha \circ W + AB$0, with $W' = \alpha \circ W + AB$1 the regularized target adapter. The asymptotic decision statistic $W' = \alpha \circ W + AB$2 is Gaussian, with explicit expressions for mean $W' = \alpha \circ W + AB$3 and variance $W' = \alpha \circ W + AB$4, and the test accuracy depends strongly on $W' = \alpha \circ W + AB$5. The optimal scaling $W' = \alpha \circ W + AB$6 unless tasks are perfectly aligned ($W' = \alpha \circ W + AB$7), and the improvement is pronounced for $W' = \alpha \circ W + AB$8 (parameter-inefficient regimes) (Firdoussi et al., 24 Oct 2025). This analysis demonstrates that additive adapters under- or over-utilize the pre-trained weights and a learned $W' = \alpha \circ W + AB$9 corrects the weighting.

Adapter Rank-Dependence: Stability in LoRA

Standard LoRA's choice of scaling factor $AB$0 for rank-$AB$1 adapters causes both forward activations and backward gradients to collapse as $AB$2 increases, rendering large-rank adaptation ineffective. The proper criterion is that the magnitude of output activations and the norm of gradients should remain $AB$3 as $AB$4. Theoretical analysis (see Theorem 3.1) proves that only $AB$5 ensures stability, hence the rank-stabilized LoRA (rsLoRA) update $AB$6 (Kalajdzievski, 2023).

Dynamic Per-Token Fusion: Contextualized Contribution

In settings with multiple LoRA adapters, static task- or module-level fusion weights fail to capture token-level task heterogeneity. LoRA-Flow employs a gating mechanism whereby fusion weights $AB$7 are generated by a small softmax gate conditioned on the current hidden state at each token and layer, allowing precise contextual control (Wang et al., 2024). Experiments show significant improvements in generative tasks demanding adaptive skill composition.

3. Training, Initialization, and Implementation

α-LoRA Parameterization

The scaling vector $AB$8 is initialized to $AB$9, matching standard LoRA at the start of fine-tuning. For LLMs, $A \in \mathbb{R}^{d_{\text{out}} \times r}$0 can be per-output-row. It is trained with a dedicated optimizer (Adam or AdamW) and higher learning rate ($A \in \mathbb{R}^{d_{\text{out}} \times r}$1 or $A \in \mathbb{R}^{d_{\text{out}} \times r}$2, versus LoRA adapter's $A \in \mathbb{R}^{d_{\text{out}} \times r}$3), and is updated every $A \in \mathbb{R}^{d_{\text{out}} \times r}$4 steps using fresh batches to minimize overfitting (Firdoussi et al., 24 Oct 2025). Typical values of $A \in \mathbb{R}^{d_{\text{out}} \times r}$5 remain in $A \in \mathbb{R}^{d_{\text{out}} \times r}$6 during standard LLM tuning, with no special norm constraint beyond standard AdamW weight decay.

rsLoRA Scaling

For rsLoRA, implementation involves replacing the scaling factor from $A \in \mathbb{R}^{d_{\text{out}} \times r}$7 to $A \in \mathbb{R}^{d_{\text{out}} \times r}$8 within the LoRA module. No other modifications are required. The value of $A \in \mathbb{R}^{d_{\text{out}} \times r}$9 can be kept as for small-rank LoRA, but may be tuned for stability (Kalajdzievski, 2023).

LoRA-Flow Fusion Gates

The fusion gates in LoRA-Flow are parameterized by per-layer matrices $B \in \mathbb{R}^{r \times d_{\text{in}}}$0 and biases $B \in \mathbb{R}^{r \times d_{\text{in}}}$1. Gates are trained with only the fusion parameters updated; all base and LoRA adapters are frozen. Training proceeds using cross-entropy loss on few-shot data, needing few parameters ($B \in \mathbb{R}^{r \times d_{\text{in}}}$20.2% of LoRA adapter size) and robust to overfitting in low-resource settings (Wang et al., 2024).

4. Empirical Results and Comparative Performance

α-LoRA Empirics

In high-dimensional linear transfer (Amazon Reviews, 400D features, $B \in \mathbb{R}^{r \times d_{\text{in}}}$3), learned $B \in \mathbb{R}^{r \times d_{\text{in}}}$4 (e.g., $B \in \mathbb{R}^{r \times d_{\text{in}}}$5 in Books→DVD) yields $B \in \mathbb{R}^{r \times d_{\text{in}}}$6–$B \in \mathbb{R}^{r \times d_{\text{in}}}$7 percentage points over vanilla LoRA ($B \in \mathbb{R}^{r \times d_{\text{in}}}$8)—Books→DVD: 64.12% (from scratch) → 75.67% (vanilla LoRA) → 77.35% (α-LoRA). On LLMs (roberta-base, LoRA rank 8, GLUE), α-LoRA consistently outperforms vanilla LoRA, with accuracy increases ranging from $B \in \mathbb{R}^{r \times d_{\text{in}}}$9 to $W'_{i,:} = \alpha_i W_{i,:} + (AB)_{i,:}$0 points depending on the task (Firdoussi et al., 24 Oct 2025).

rsLoRA Scaling

When training Llama 2-7B on OpenOrca with increasing adapter ranks, the perplexity curves under standard LoRA ($W'_{i,:} = \alpha_i W_{i,:} + (AB)_{i,:}$1) are nearly identical and insensitive to rank; for rsLoRA ($W'_{i,:} = \alpha_i W_{i,:} + (AB)_{i,:}$2), performance improves monotonically with $W'_{i,:} = \alpha_i W_{i,:} + (AB)_{i,:}$3. The average parameter gradient norm under standard LoRA vanishes for high $W'_{i,:} = \alpha_i W_{i,:} + (AB)_{i,:}$4, in contrast to rsLoRA where it remains stable across ranks. This confirms rsLoRA's stable learning and utility of high-rank adaptation (Kalajdzievski, 2023).

LoRA-Flow Fusion

Combining multiple LoRA adapters on Llama-2 models, LoRA-Flow achieves Math (MGSM) accuracy of 37.6% versus 28.7% with task-level fusion and 13.9% with static fusion; similar improvements are seen in code generation (HumanEval) and ablated gate granularity (layer-level outperforms module/step-level). In multilingual tasks (Llama-2-13B), LoRA-Flow reaches 41.2%/35.4% (math/code) versus 40.0%/34.2% for the best static fusion. In few-shot settings, LoRA-Flow consistently exceeds training new or task-specific LoRA modules (Wang et al., 2024).

5. Practical Recommendations and Application Scenarios

When to use α-LoRA:

Ideal for low-resource tuning ($W'_{i,:} = \alpha_i W_{i,:} + (AB)_{i,:}$5 small), or when target tasks are only partially aligned with pre-training. Also suited for scenarios where small relative shifts in pretrained weights matter (e.g., cross-domain transfer). The compute and memory overhead is negligible—on roberta-base, the additional $W'_{i,:} = \alpha_i W_{i,:} + (AB)_{i,:}$6 parameters constitute $W'_{i,:} = \alpha_i W_{i,:} + (AB)_{i,:}$7 extra.

When to use rsLoRA:

Advantageous when high-rank adapters are needed for more expressive adaptation, enabling a smooth compute–performance trade-off. The best practice is to use the largest rank $W'_{i,:} = \alpha_i W_{i,:} + (AB)_{i,:}$8 permissible by hardware constraints and tune $W'_{i,:} = \alpha_i W_{i,:} + (AB)_{i,:}$9 as needed.

Dynamic Fusion with LoRA-Flow:

Appropriate for generative and multitask settings demanding token-wise skill composition, e.g., multilingual LLMs tackling mixed-domain tasks. Fusion gates are compact and train efficiently with minimal examples.

Implementation caveats:

For α-LoRA, ensure $$\#(\alpha) = d_\text{out} \ll d_\text{out}d_\text{in}$$0 is optimized with distinct batches and learning rates to prevent overfitting, and for rsLoRA, simply update the scaling law to $$\#(\alpha) = d_\text{out} \ll d_\text{out}d_\text{in}$$1.

6. Impact, Limitations, and Future Directions

Multiplicative LoRA weights offer an additional degree of freedom over additive-only frameworks, theoretically guaranteeing improved or equivalent asymptotic generalization in alignment-mismatched and low-data regimes, and empirically enhancing transfer accuracy in both linear and LLM tasks (Firdoussi et al., 24 Oct 2025). rsLoRA reactivates the use of large-rank adapters previously ineffective under standard scaling, allowing performance scaling commensurate with training resources (Kalajdzievski, 2023). In multi-LoRA fusion tasks, dynamic multiplicative gating significantly outperforms static weights and enables granular, contextual adaptation (Wang et al., 2024).

The incremental overheads in parameters and computation are minimal, making multiplicative LoRA extensions broadly applicable within the current PEFT and transfer learning ecosystem. In regimes where downstream data is abundant or tasks are strongly aligned, the improvement from multiplicative scaling may diminish, and standard additive adapters suffice. A plausible implication is that future work may explore more fine-grained or adaptive multiplicative schemes, such as hybrid gating or hierarchical scaling, especially as multi-LoRA and meta-learning approaches proliferate.