Cross-Version ∆-Transfer of QLoRA-Trained Adapters¶
Conference: EMNLP 2026
Paper: ACL Anthology
Area: LLM Reasoning
Keywords: QLoRA, LoRA, Cross-Version Transfer, Model Merging, Parameter-Efficient Fine-Tuning
TL;DR¶
This paper demonstrates that BF16 LoRA adapters trained on a 4-bit quantized base (NF4) transfer across model versions (e.g., Llama-3-8B-Instruct to Llama-3.1-8B-Instruct) as effectively as full-precision LoRA adapters across 20 downstream tasks (absolute score gap \(\le 0.054\)), while cutting post-load deployment VRAM by 49%–63% under NF4 runtime serving and showing that the empirical task-optimal merge scale \(\lambda^*\) systematically exceeds the weight-space Frobenius distance optimum.
Background & Motivation¶
Fine-tuning large language models on complex domain-specific tasks requires substantial computational resources. Low-Rank Adaptation (LoRA) reduces trainable parameters by factorizing weight updates into low-rank matrices, while QLoRA pushes memory savings further by freezing the base model in 4-bit NormalFloat (NF4) and dequantizing it to 16-bit brain floating point (BF16) strictly during forward and backward matrix multiplications. Concurrently, the paradigm of model merging and Task Arithmetic treats fine-tuning weight deltas (\(\Delta W\)) as modular, reusable task vectors. Recent efforts such as Param\(\Delta\) and Fine-tuning Transfer have proven that adapters trained on an earlier base model version can be directly transferred to an updated base model (such as migrating from Llama 3.0 to Llama 3.1) at essentially zero training cost, eliminating the need to retrain domain adapters whenever foundational model providers issue point releases.
However, existing investigations into cross-version and cross-model delta transfer rest on an unexamined foundational assumption: the adapter to be transferred must originate from a full-precision (BF16) fine-tuning run. In practical research and industrial deployments, compute constraints make QLoRA the de facto standard adaptation workflow. Because base weights are held in NF4 throughout training, the stochastic perturbation and quantization errors inherently affect gradient dynamics. Whether the resulting BF16 adapter retains functional fidelity and structural alignment when transplanted onto a newer base model was completely uncharted territory. Furthermore, prior transfer frameworks invariably hardcode a naive scaling factor \(\lambda = 1.0\) or assume weight-space geometric proximity governs parameter composition, leaving the scaling dynamics of cross-precision adapter injection poorly understood.
To resolve these foundational questions, this work decouples training-time base precision from inference-time evaluation to systematically investigate whether QLoRA-derived adapters exhibit cross-version transferability on par with LoRA. In addition, it rigorously interrogates the hyperparameter landscape of adapter scaling factors across tasks and architectural families. Core idea: Training an adapter on an NF4-quantized base does not compromise its cross-version transferability to newer base models (maintaining a score gap \(\le 0.054\) across 20 diverse tasks compared to full-precision LoRA), while revealing that empirical task-optimal merge scales \(\lambda^*\) systematically exceed weight-space Frobenius projections \(\lambda_{\mathrm{opt}}\) by an average of 4.35\(\times\).
Method¶
Overall Architecture¶
The paper formulates a two-phase pipeline to study and execute cross-version adapter delta transfer across differing base precisions. In Phase 1 (Low-Cost Training), a source base model \(M_{\mathrm{src}}\) (specifically Llama-3-8B-Instruct) is loaded in frozen 4-bit NF4 precision with double quantization, and low-rank adapter matrices \(A \in \mathbb{R}^{r \times d_{\mathrm{in}}}\) and \(B \in \mathbb{R}^{d_{\mathrm{out}} \times r}\) (\(r=8, \alpha=16\)) are optimized in BF16, yielding the task update \(\Delta W_{\mathrm{src}} = \frac{\alpha}{r} BA\). In Phase 2 (Cross-Version Transfer and Joint Search), this BF16 adapter is injected into a target base model \(M_{\mathrm{tgt}}\) of a newer generation and full precision (Llama-3.1-8B-Instruct in BF16) modulated by a merge scaling factor \(\lambda\): $\(W_{\mathrm{final}} = W_{\mathrm{tgt}} + \lambda \cdot \Delta W_{\mathrm{src}}\)$ To ensure an equitable baseline comparison where performance discrepancies reflect solely training-time quantization rather than inference-time execution differences, primary evaluations are conducted on a BF16 target base. For operational serving, the adapter is separately deployed atop an NF4 target base via runtime scaling. Selection of hyperparameters bypasses naive decoupled search through an internal joint grid search over training checkpoints and scaling factors \((CP^*, \lambda^*)\) on split TestA, followed by clean validation on held-out split TestB.
%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
A["Source Base Llama-3.0 (NF4 Frozen)"] --> B["Low-Cost QLoRA Training<br/>BF16 adapter update on NF4 base"]
B --> C["Cross-Version Parameter Injection<br/>W_final = W_tgt + λ·ΔW_src"]
C --> D["Joint Checkpoint & Scale Search<br/>Joint (CP*, λ*) optimization on TestA"]
D --> E["Cross-Precision & Cross-Family Verification<br/>Weight-space geometric analysis & NF4 serving"]
E --> F["Downstream Evaluation Output (TestB)"]
Key Designs¶
1. Low-Cost QLoRA Training: Quantized Base Freezing with High-Precision Adapter Gradients
To circumvent the prohibitive GPU VRAM demands of full-precision fine-tuning across dozens of candidate benchmarks, the source training stage strictly follows the QLoRA protocol. The base model weights \(W_{\mathrm{src}}\) are quantized into NF4 with double quantization enabled via bitsandbytes, while LoRA decomposition matrices \(A\) and \(B\) are instantiated in BF16. Base weights are dequantized to BF16 on-the-fly exclusively during forward and backward matrix multiplications, and gradient backpropagation updates only the BF16 adapter parameters. Consequently, the transferred delta \(\Delta W_{\mathrm{src}} = \frac{\alpha}{r} BA\) resides natively in BF16 while reducing peak training VRAM from 29.4–30.2 GB down to 21.5–24.6 GB (a 20%–30% reduction), validating that lightweight local fine-tuning can produce reusable weight vectors.
2. Cross-Version Parameter Injection: Non-Invasive Adapter Re-Scaling Across Precision Regimes
To eliminate the compute waste of training domain models from scratch whenever base models evolve, this design grafts the source delta \(\Delta W_{\mathrm{src}}\) onto target base weights \(W_{\mathrm{tgt}}\). In contrast to existing frameworks that blindly assume identity scaling \(\lambda = 1.0\), the framework introduces a parameterized task-specific scaling factor \(\lambda \ge 0\). Technically, rather than permanently baking the weights via destructive merge_and_unload, the scaling is injected at runtime by overriding the PEFT library's BaseTunerLayer.scaling attribute. To isolate variables for rigorous empirical attribution, the benchmark comparison runs both LoRA Transfer and QLoRA Transfer on a BF16 target base, ensuring the performance gap isolates training-time base precision from inference-time quantization artifacts. In deployment phases, loading the target base in NF4 cuts steady-state serving memory by 49%–63%.
3. Joint Checkpoint & Scale Search: Unbiased Dual-Parameter Grid Optimization
To address the severe structural bias uncovered during pilot investigations—where a conventional two-stage search that first selects the top checkpoint at \(\lambda = 1.0\) and subsequently sweeps \(\lambda\) yielded \(\lambda^* = 1.0\) in 4 out of 4 trials and missed latent peak accuracy by up to 6.7 percentage points—the pipeline institutes a joint \((CP, \lambda)\) grid search. Operating over task-specific internal validation partitions (TestA, formed by doc_id % 2 == 0), the search jointly identifies:
$\((CP^*, \lambda^*) = \arg\max_{CP, \lambda} \mathrm{Score}_{\mathrm{TestA}}\left(W_{\mathrm{tgt}} + \lambda \cdot \Delta W_{\mathrm{src}}(CP)\right)\)$
On 20 core tasks, Experiment 1 explores a fine-grained 21-point grid \(\lambda \in \{0.0, 0.1, \dots, 2.0\}\), whereas deployment setups employ a pruned 5-value smart grid informed by empirical distributions. High Spearman rank correlation between TestA and held-out TestB (\(\rho = 0.96\) for generation, \(\rho = 0.80\) for classification) proves that the joint search generalizes robustly without validation overfitting.
4. Cross-Precision & Cross-Family Verification: Dissecting Weight-Space Projections vs Functional Optima
To test whether expensive search procedures could be replaced by a closed-form analytic prediction, the framework introduces a geometric baseline \(\lambda_{\mathrm{opt}}\). By computing the least-squares projection of the Direct Adaptation delta \(\Delta W_{\mathrm{DA}}\) onto the 1D ray spanned by \(\Delta W_{\mathrm{src}}\) under the Frobenius norm: $\(\lambda_{\mathrm{opt}} = \max\left(0, \frac{\langle \Delta W_{\mathrm{DA}}, \Delta W_{\mathrm{src}} \rangle_F}{\|\Delta W_{\mathrm{src}}\|_F^2}\right) = \cos(\Delta W_{\mathrm{DA}}, \Delta W_{\mathrm{src}}) \cdot \frac{\|\Delta W_{\mathrm{DA}}\|_F}{\|\Delta W_{\mathrm{src}}\|_F}\)$ Inner products are accumulated across all 224 submodules strictly in float64 to prevent catastrophic floating-point cancellation. Across all 44 experimental settings (22 tasks \(\times\) 2 methods on Llama), empirical evaluation uncovers a universal \(\lambda \text{-gap}\): the task-optimal \(\lambda^*\) strictly exceeds \(\lambda_{\mathrm{opt}}\) (mean ratio 4.35\(\times\)). This reveals that functional task activation operates in a regime far beyond simple Euclidean parameter alignment, an observation subsequently replicated across Mistral and Qwen model families.
Benchmark & Evaluation Protocol¶
Source adapters are trained via paged AdamW 8-bit using a learning rate of \(1 \times 10^{-5}\), zero weight decay, 0.03 warmup ratio with linear decay, and an effective batch size of 12 across 5 epochs. Target evaluation evaluates 22 diverse NLP tasks structured into Group A (10 text generation benchmarks), Group B (10 classification and reasoning tasks), and Group C (2 practical deployment benchmarks). Text generation benchmarks are evaluated using sacrebleu corpus BLEU alongside auxiliary BERTScore-F1 under greedy decoding. Multiple-choice classification benchmarks are evaluated using prompt option log-likelihood via lm-evaluation-harness. Mathematical reasoning on GSM8K is scored via exact string match on chain-of-thought outputs under standard strict filtering.
Key Experimental Results¶
Main Results¶
The primary evaluation in Experiment 1 establishes the direct transfer performance of QLoRA-trained adapters compared against full-precision LoRA-trained adapters on the target base Llama-3.1-8B-Instruct (BF16). Below are the TestB corpus BLEU results across all 10 Group A generation tasks (numbers in parentheses indicate the selected \(\lambda^*\)):
| Task | Train Examples \(n_{\mathrm{tr}}\) | Base 3.1 | LoRA Transfer (\(\lambda^*\)) | QLoRA Transfer (\(\lambda^*\)) | Score Gap \(\Delta\) (Q - L) |
|---|---|---|---|---|---|
| MBPP | 474 | 0.062 | 0.131 (1.8) | 0.152 (1.8) | +0.021 |
| MeQSum | 800 | 0.011 | 0.235 (1.5) | 0.236 (1.4) | +0.001 |
| SciTLDR | 1,992 | 0.042 | 0.106 (0.9) | 0.104 (1.1) | -0.002 |
| ViGGO | 5,103 | 0.092 | 0.238 (1.1) | 0.259 (1.1) | +0.021 |
| NarrativeQA | 10,000 | 0.056 | 0.216 (0.8) | 0.220 (0.8) | +0.004 |
| SAMSum | 10,000 | 0.047 | 0.186 (0.9) | 0.195 (0.9) | +0.009 |
| DialogSum | 10,000 | 0.030 | 0.194 (1.1) | 0.198 (1.2) | +0.004 |
| AESLC | 14,438 | 0.014 | 0.078 (0.5) | 0.093 (0.6) | +0.015 |
| E2E NLG | 33,525 | 0.040 | 0.259 (1.1) | 0.267 (1.1) | +0.008 |
| WebNLG | 35,425 | 0.186 | 0.329 (0.7) | 0.331 (0.6) | +0.002 |
Across all 20 Group A and Group B benchmarks, the absolute evaluation-score gap between QLoRA Transfer and LoRA Transfer is at most 0.054 (observed on WinoGrande, 0.709 vs 0.763), and remains \(\le 0.030\) on 18 out of 20 tasks. Within an equivalence margin of \(|\Delta| \le 0.02\), Group A yields 8 ties and 2 QLoRA wins, while Group B yields 8 ties and 2 LoRA wins, demonstrating that training adapters under 4-bit base quantization preserves cross-version transferability.
Ablation & Practical Deployment¶
The table below presents Experiment 2 (Group C deployment evaluation), assessing transfer against direct adaptation (DA) alongside paired bootstrap significance testing (10,000 resamples) and steady-state inference VRAM:
| Task | Method | Joint Search | DA (\(\lambda=1.0\)) | \(\Delta\) (Joint - DA) | \(p\)-value | Deploy Steady-State VRAM |
|---|---|---|---|---|---|---|
| CommonGen (Gen.) | LoRA | 0.319 | 0.305 | +0.014 | – | 15.1 GB (BF16) |
| CommonGen (Gen.) | QLoRA | 0.337 | 0.319 | +0.018 | – | 5.7 GB (NF4) |
| CommonsenseQA (Clf.) | LoRA | 0.821 | 0.795 | +0.026 | 0.038 | 19.6 GB (BF16) |
| CommonsenseQA (Clf.) | QLoRA | 0.764 | 0.797 | -0.033 | 0.014 | 10.0 GB (NF4) |
Furthermore, replicating the entire experimental matrix across three independently developed open-weight model families (Llama 3.0\(\to\)3.1, Mistral v0.1\(\to\)v0.3, and Qwen2\(\to\)2.5) yields: 1. Method Equivalence (\(|\Delta| \le 0.02\)): 54/60 settings (90.0%) exhibit practical equivalence, with mean absolute difference across all settings measuring only 0.008. 2. Prevalence of the \(\lambda \text{-Gap}\) (\(\lambda^* \succ \lambda_{\mathrm{opt}}\)): In 111 out of 124 cross-family settings (89.5%), the empirical scale outperforms the analytic Frobenius minimizer, yielding an average performance advantage of +0.067. 3. NF4 Deployment Retention (\(|\Delta_{\mathrm{NF4-BF16}}| \le 0.02\)): Serving transferred adapters on quantized NF4 bases maintains performance parity on 53/60 settings (88.3%) with a mean gap of just 0.011.
Key Findings¶
- 4-Bit Quantization Preserves Weight Transfer Dynamics: Discarding base model floating-point precision during adaptation training does not degrade the functional reusability of low-rank adapters across model generations.
- Breakdown of Weight-Space Geometric Heuristics: Across all 44 Llama settings, the empirical optimum \(\lambda^*\) is strictly larger than the Frobenius distance minimizer \(\lambda_{\mathrm{opt}}\) (mean ratio \(4.35\times\)). Relying on geometric distance to direct-adaptation weights in parameter space fails to find the optimal operational point.
- Bimodal Task-Type Preference for Scaling Factor \(\lambda\): Multiple-choice classification and formal reasoning benchmarks systematically favor heightened adapter scales (\(\lambda^* \ge 1.5\) to \(2.0\)), whereas long-form open generation tasks favor moderate scales (\(\lambda^* < 1.0\)) to avoid lexical degeneration.
Highlights & Insights¶
- Democratizing LLM Version Migration: Validates an end-to-end low-compute lifecycle: researchers fine-tune models cheaply on consumer GPUs using QLoRA, transplant weights across base model upgrades without retraining, and serve them in 4-bit memory footprints (5.7–10.0 GB).
- Discovery of the \(\lambda \text{-Gap}\): Provides strong empirical evidence that in deep non-linear neural networks, functional capability activation does not correlate linearly with Euclidean parameter proximity, offering theoretical insights for model merging and task vector surgery.
- Rigorous Cross-Family Replicability: Methodological findings hold not only on Llama 3, but replicate consistently on Mistral and Qwen families, establishing generalizable principles for adapter transfer.
Limitations & Future Work¶
- Architectural Scope and Scaling Restrictions: Evaluations are restricted to 7B–8B parameter instruction-tuned models and NF4 quantization; behavior on 70B+ scales, post-training quantization schemes (e.g., AWQ, GPTQ), or quantization-aware adapters (e.g., QA-LoRA, LoftQ) remains unverified.
- Negative Outlier on Complex Commonsense QA: On CommonsenseQA, QLoRA Transfer underperforms Direct Adaptation by 3.3% (\(p=0.014\)), indicating that cross-version delta transfer is not universally immune to performance drops on specialized reasoning tasks.
- Right-Censored Search Boundaries: For 7 of 44 classification settings, the search saturated at the maximum grid ceiling \(\lambda^* = 2.0\), meaning reported ratios between empirical and analytic scales represent conservative lower bounds.
Related Work & Insights¶
- vs Param\(\Delta\) (Cao et al., 2025) & Fine-tuning Transfer (Lin et al., 2025): Prior work established cross-version transfer for deltas trained exclusively on full-precision BF16 bases using fixed \(\lambda = 1.0\); this paper extends the horizon to 4-bit quantized bases and establishes the necessity of tuning \(\lambda\).
- vs Task Arithmetic (Ilharco et al., 2023) & LoRA Soups (Prabhakar et al., 2025): Previous adapter arithmetic methods assumed intra-version homogeneous environments; this work validates non-homogeneous delta reuse across both precision regimes and architectural releases.
- vs LoftQ (Li et al., 2024) & QA-LoRA (Xu et al., 2024): While quantization-aware adaptation designs algorithms to bridge single-model quantization gaps, this paper addresses the orthogonal challenge of downstream portability for already-trained quantized adapters.
Rating¶
- Novelty: ⭐⭐⭐⭐☆ Bridges a vital empirical and conceptual gap in parameter-efficient fine-tuning and model merging.
- Experimental Thoroughness: ⭐⭐⭐⭐⭐ Comprehensive evaluation across 22 tasks, 3 independent LLM families, multi-seed validation, bootstrap significance testing, and VRAM telemetry.
- Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally rigorous experimental design, transparent reporting of negative cases, and insightful theoretical analysis.
- Value: ⭐⭐⭐⭐⭐ Highly actionable guidance for open-source practitioners and resource-constrained researchers navigating rapid foundation model releases.