DRIVE: Fine-tuning via Data Contribution- and Diversity-aware Weighting with Prior Regularization

Qing Liu; Xinrui Chen; Yi Du; Ou Wu; Weiyao Zhu

cs.LGOct 23, 2026

DRIVE: Fine-tuning via Data Contribution- and Diversity-aware Weighting with Prior Regularization

Authors: Qing Liu, Xinrui Chen, Yi Du, Ou Wu, Weiyao Zhu

Organizations: Hangzhou Institute for Advanced Study, University of Chinese Academy of Sciences, Hangzhou, China · Computer Network Information Center, Chinese Academy of Sciences, Beijing, China · School of Mathematics, Tianjin University, Tianjin, China

Abstract

Single-stage fine-tuning is pervasive, yet catastrophic forgetting persists when the source pre-training data is inaccessible. Data weighting affords a natural lever to modulate sample influence. Yet existing weighting methods lack a principled mechanism to explicitly arbitrate knowledge retention against new-task adaptation, and provide no unified differentiable objective to accommodate auxiliary signals such as sample diversity and sample characteristics. This is because current schemes privilege one goal over the other, and no explicit gradient-compatible encoding exists for such signals. To address this, we propose DRIVE, a principled and efficient weight assignment framework. It derives per-sample weights via a unified differentiable objective where three coupled signals co-evolve: Shapley additivity splits contribution into anti-forgetting and adaptation components, which are reshaped by a covariance-based diversity term and stabilized by a sample-characteristic prior. Far from a mere combination, this yields an end-to-end gradient-based scheme balancing stability against plasticity. Theoretically, we prove that DRIVE enjoys superior stability-plasticity guarantees over existing weighting schemes and enhances parameter estimation robustness. Empirically, it attains superior overall performance, faster convergence, and stronger robustness. Code at https://github.com/luciabcd/DRIVE.

Figures & tables

Figure 1: Overview of DRIVE’s unified differentiable data weighting objective.

Case	Weight form	Meaning
$\eta=0$	$w_{i}^{*}\propto\exp[(-\lambda l_{i}^{\theta_{\mathrm{pre}}}+(1-\lambda)l_{i}^{\theta})/\tau]$	Without diversity
- $\lambda=1$	$w_{i}^{*}\propto\exp(-l_{i}^{\theta_{\mathrm{pre}}}/\tau)$	Anti-forgetting [ 20 ]
- $\lambda=0$	$w_{i}^{*}\propto\exp(l_{i}^{\theta}/\tau)$	Task adaptation [ 16 ]
- $\lambda=0.5$	$\bm{\theta}\!=\!\bm{\theta}_{\mathrm{pre}}\Rightarrow w_{i}^{*}\propto 1/N$	Standard FT (init.)
$\eta>0$	$w_{i}^{(t)}\propto\exp[(c_{i}+\eta\,\bar{\bm{x}}_{i}^{\top}(\bm{\Sigma}_{w}^{(t-1)})^{-1}\bar{\bm{x}}_{i})/\tau]$	Full DRIVE

Table 1: Special cases of DRIVE under a uniform prior

Figure 2: Impact of Uniform Prior and Determinant Regularization on Weight Distribution.

Method	Gen Acc.	Target Acc.	Overall Avg.
$\ell_{2}$ -Reg.	55.58 (+0.00)	73.21 (+0.00)	64.40 (+0.00)
+ FLOW	56.08 (+0.50)	74.80 (+1.59)	65.44 (+1.04)
+ DRIVE	56.92 (+1.34)	76.19 (+2.98)	66.56 (+2.16)
LoRA	56.90 (+0.00)	74.50 (+0.00)	65.70 (+0.00)
+ FLOW	57.52 (+0.62)	74.91 (+0.41)	66.22 (+0.52)
+ DRIVE	57.54 (+0.64)	76.41 (+1.91)	66.98 (+1.28)

Table 4: DRIVE acts as a plug-and-play module, boosting performance in

\ell_{2}

-Reg. and LoRA.

Method	General	Target	Avg
DRIVE	57.51	77.03	67.27
- w/o Task-adaptation ( $\lambda=1$ )	58.42	73.19	65.81
- w/o Anti-forgetting ( $\lambda=0$ )	52.24	79.83	66.04
- w/o Diversity ( $\eta=0$ )	56.18	74.78	65.48
- w/o KL Prior	53.96	70.61	62.29

Table 5: Ablation study of DRIVE components.

Panel of Figure 3 — Figure 3: Sensitivity analysis of $\tau$ and $\eta$ on ResNet-50.

Setting	General	Target	Overall
DRIVE	57.48	77.10	67.29
Shuffled	56.31	73.92	65.12
Reversed	50.02	70.43	60.23

Table 9: Effect of score perturbations.

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Metric	Anti-forgetting	Task adaptation	Unified score $c_{i}$
Quantile Monotonicity Score $\uparrow$	0.89	0.98	0.92
Spearman $\rho$ $\uparrow$	0.76	0.87	0.83
Top-50% Overlap $\uparrow$	0.87	0.95	0.89
Pearson $r$ $\uparrow$	0.65	0.87	0.81

Appendix

Table 10: Agreement between loss-based proxy contributions and validation-based Shapley reference scores on ResNet-50 fine-tuned on Flowers102, using an ImageNet subset as the retain validation set.

Datasets	CIFAR10	CIFAR100	Flowers	Caltech	Dogs	Cars	Food	Beans	EuroSAT
Epochs	20	25	25	30	30	30	30	20	20
ResNet18	1e-3	5e-3	5e-2	1e-2	5e-3	1e-2	1e-2	1e-2	1e-3
ResNet50	5e-4	1e-3	1e-2	1e-2	5e-3	1e-2	1e-2	1e-2	1e-3

Appendix

Table 12: Hyperparameter configurations for DRIVE across various datasets.

	Method	Caltech101		Dogs		Cars		Food101		Beans		EuroSAT
	Method	Gen	Tgt	Gen	Tgt	Gen	Tgt	Gen	Tgt	Gen	Tgt	Gen	Tgt
ResNet18	Linear probing	69.76	88.63	69.76	78.55	69.76	50.21	69.76	60.03	69.76	85.78	69.76	91.35
	Standard FT	35.79	94.89	58.01	80.30	1.84	87.64	0.23	74.03	48.95	97.66	24.79	96.79
	$\ell_{2}$ -Regularization	45.16	94.89	60.11	81.33	20.79	83.79	10.20	75.69	40.77	96.88	39.09	95.42
	WiSE-FT	62.96	90.60	66.89	75.42	32.57	65.46	20.79	70.30	44.87	96.41	50.75	90.18
	DEFT-UCS	59.69	89.59	59.98	79.62	28.54	60.31	10.77	70.94	41.97	95.31	50.32	90.01

Appendix

Table 13: Combined performance comparison across six datasets. For each dataset, Tgt denotes the target task accuracy (higher is better for adaptation), and Gen denotes the ImageNet-1K accuracy (higher is better for anti-forgetting) (Accuracy in %). Best results are bold , second-best results are underlined .

Model	Method	Caltech101	Dogs	Cars	Food101	Beans	EuroSAT
ResNet18	Linear probing	79.20	74.16	59.99	64.90	77.77	80.56
	Standard FT	65.34	69.16	44.74	37.13	73.31	60.79
	$\ell_{2}$ -Regularization	70.03	70.72	52.29	42.95	68.83	67.26
	WiSE-FT	76.78	71.16	49.02	45.55	70.64	70.47
	DEFT-UCS	74.64	69.80	44.43	40.86	68.64	70.17
	FLOW	78.09	74.27	59.75	63.61	79.93	81.13

Appendix

Table 14: Comparison of Average Accuracy (%) across six downstream vision datasets. This metric reflects the overall balance between representation preservation and task adaptation. Best results are bold , second-best results are underlined .

Method	IN-1K	Food101	Avg.
Pre-trained	82.95	-	-
Standard FT	58.42	92.24	75.33
Linear Probe	82.95	83.93	83.44
$l_{2}$ -Reg.	57.94	91.78	74.86
WiSE-FT	55.38	90.64	73.01
DEFT-UCS	67.35	90.15	78.75

Appendix

Table 15: Performance comparison (Accuracy in %) of ViT-B/16 on Food101. Best in bold , second in underlined .

$\lambda$ value	General Capability Acc. (%)				Target Acc. (%)	Overall (%)
$\lambda$ value	CS	MMLU	MBPP	Avg	GSM8K	Avg
$\lambda=0.2$	42.23	55.55	39.20	45.66	63.81	54.74
$\lambda=0.5$	54.95	55.64	40.40	50.33	60.82	55.58
$\lambda=0.8$	55.79	55.86	41.64	51.10	60.29	55.70

Appendix

Table 17: Comparison of different

\lambda

on Llama-3.2-3B. Higher

\lambda

favors general capabilities, while lower

\lambda

benefits target adaptation.

\lambda=0.8

achieves the best overall balance.

Model	General Capability Acc. (%)				Target Acc. (%)	Overall (%)
Model	CS	MMLU	MBPP	Avg	GSM8K	Avg
Gemma-2-2B	56.57	52.59	29.00	46.05	63.79	54.92
Llama-3.2-3B	54.28	54.23	37.00	48.50	63.12	55.81
Llama-3.1-8B	58.48	62.29	45.80	55.52	77.98	66.75

Appendix

Table 18: Performance on GSM8K and general capabilities using a linearly increasing schedule for

\lambda

. All accuracies are reported in %.

$\tau$ -Percentile(%)	General Capability Acc. (%)				Target Acc. (%)	Overall (%)
$\tau$ -Percentile(%)	CS	MMLU	MBPP	Avg	GSM8K	Avg
10	55.92	55.71	41.00	50.88	64.50	57.69
30	55.19	55.80	38.80	49.93	61.90	55.92
50	54.91	54.34	38.80	49.35	60.80	55.08
70	55.06	55.65	39.60	50.10	59.14	54.62
90	53.13	55.90	40.00	49.68	58.01	53.85

Appendix

Table 19: Comparison of different

\tau

percentiles on Llama-3.2-3B. All accuracies are reported in %.

$\eta$ value	General Capability Acc. (%)				Target Acc. (%)	Overall (%)
$\eta$ value	CS	MMLU	MBPP	Avg	GSM8K	Avg
0.25	55.08	56.30	40.88	50.75	59.67	55.21
0.50	54.56	56.67	41.52	50.92	58.65	54.79
0.75	55.67	56.79	41.00	51.15	60.61	55.88
1.00	55.92	55.71	41.00	50.88	64.50	57.69
1.25	55.19	55.80	38.80	49.93	61.90	55.92

Appendix

Table 20: Comparison of different

\eta

on Llama-3.2-3B.

Method	CIFAR-100			Flowers102
Method	IN-1K	Target Acc.	Avg.	IN-1K	Target Acc.	Avg.
DRIVE w/o Diversity ( $\eta=0$ )	65.79	75.76	70.78	77.93	88.93	83.43
DRIVE w/o KL Prior	55.93	69.71	62.82	66.94	87.52	77.23
DRIVE w/o anti-forgetting ( $\lambda=0$ )	57.49	86.40	71.95	71.98	93.67	82.83
DRIVE w/o task-adaptation ( $\lambda=1$ )	63.49	80.91	72.20	76.26	90.73	83.50
DRIVE	61.64	83.03	72.34	75.02	92.67	83.85

Appendix

Table 21: Ablation study of DRIVE components on ResNet-50 (Accuracy in %).

Model	Method	GPU Time / Epoch (s)	Overhead (%)
ResNet-18	Standard FT	280	-
	FLOW	293	+4.64%
	DRIVE (Ours)	313	+11.79%
ResNet-50	Standard FT	389	-
	FLOW	402	+3.34%
	DRIVE (Ours)	446	+14.65%

Appendix

Table 22: Training efficiency on CIFAR-100, reported as GPU seconds per epoch.

Model	Method	GPU Time / Epoch (h)	Overhead (%)
Gemma-2-2B	Standard FT	5.45	-
	FLOW	6.00	+10.09%
	DRIVE (Ours)	6.38	+17.06%
Llama-3.1-8B	Standard FT	19.42	-
	FLOW	21.36	+9.99%
	DRIVE (Ours)	22.75	+17.15%

Appendix

Table 23: Training efficiency on MetaMathQA, reported as GPU hours per epoch.

Model	Baseline	Baseline Overall	DRIVE Overall	$\Delta$	95% CI	Raw $p$	Holm-adjusted $p$	Hedges’ $g_{z}$
Gemma-2-2B	TALR	53.87	55.55	+1.68	$[0.85,\,2.51]$	0.0049	0.0148	2.00
Llama-3.2-3B	TALR	56.26	57.69	+1.43	$[0.46,\,2.42]$	0.0152	0.0152	1.45
Llama-3.1-8B	LoRA	65.70	67.27	+1.57	$[0.77,\,2.37]$	0.0055	0.0148	1.95

Appendix

Table 24: Paired statistical analysis of DRIVE against the strongest baseline for each LLM, using five runs with matched random seeds and data orders. Overall scores are percentages; differences and confidence intervals are expressed in percentage points. Positive

\Delta

favors DRIVE.

Subset	Target (%)	Retain (%)
Pre-trained	-	80.10
Conflicting	84.85	38.42
Synergistic	83.76	79.15

Appendix

Table 25: Impact of Gradient Alignment. Synergistic samples prevent forgetting despite high loss.

Figure 7: Scatter plot of normalized Anti-forgetting (

\widetilde{a}_{i}

) vs. Task Adaptation (

\widetilde{t}_{i}

) contributions.

Method	CIFAR-100			Flowers102
Method	Target	IN-1K	Avg.	Target	IN-1K	Avg.
Single Proj. ( $M=1$ )	83.03	61.64	72.34	92.67	75.02	83.85
Ensemble ( $M=5$ )	83.08	61.67	72.38	92.65	75.09	83.87
Ensemble ( $M=10$ )	83.10	61.65	72.38	92.71	75.05	83.88

Appendix

Table 26: Performance of DRIVE with varying ensemble sizes on CIFAR-100 and Flowers102.

Figure 8: Loss Dynamics on

\mathcal{D}_{hard}

. FLOW exhibits high-variance stalling (jagged), while DRIVE achieves stable adaptation (smooth).

$k$	Gen.	Target	Overall	Time
32	46.72	54.18	50.45	1.00
64	47.01	57.60	52.31	1.30
128	47.64	63.46	55.55	1.70
256	47.69	63.79	55.74	2.50
512	47.71	64.32	56.02	4.10

Appendix

Table 27: Sensitivity to

k

on Gemma-2-2B / MetaMathQA. Relative time is normalized by

k=32

Fine-Tuning

Fine-tuning adapts pretrained models to new tasks, domains, and behaviors by updating some or all of their parameters.

See all 250 papers

Weight Updates as Activation Shifts: A Principled Framework for Steering
Feb 28, 2026Dyah Adila, John Cooper, Alexander Yun +2Activation Steering Parameter-Efficient Fine-Tuning
Department of Computer Science, University of Wisconsin-Madison
PriFT: Guiding Supervised Fine-Tuning with Pretrained Priors
Jun 8, 2026Ke Wang, Shuangqi Li, Mathieu Salzmann +1Supervised Fine-Tuning Fine-Tuning
EPFL, Lausanne, Switzerland
Fisher-Guided Progressive Parameter Selection for Adaptive Fine-Tuning
Jun 8, 2026Ghodsiyeh Rostami, Po-Han Chen, Mahdi S. HosseiniFine-Tuning Vision Foundation Model Adaptation
Concordia University · Concordia University Mila - Québec AI Institute

Feb 28, 2026J/K move · Enter open · S save

Weight Updates as Activation Shifts: A Principled Framework for Steering

Dyah Adila, John Cooper, Alexander Yun, Avi Trost+1

Department of Computer Science, University of Wisconsin-Madison

Activation Steering Parameter-Efficient Fine-Tuning

Open paper AskPDF arXiv

Activation steering promises to be an extremely parameter-efficient form of adaptation, but its effectiveness depends on critical design choices -- such as intervention location and parameterization -- that currently rely on empirical heuristics rather than a principled foundation. We establish a first-order equivalence between activation-space interventions and weight-space updates, deriving the conditions under which activation steering can replicate fine-tuning behavior. This equivalence yields a principled framework for steering design and identifies the post-block output as a theoretically-backed and highly expressive intervention site. We further explain why certain intervention locations outperform others and show that weight updates and activation updates play distinct, complementary functional roles. This analysis motivates a new approach -- joint adaptation -- that trains in both spaces simultaneously. Our post-block steering achieves accuracy within 0.2%-0.9%$ of full-parameter tuning, on average across tasks and models, while training only 0.04% of model parameters. It consistently outperforms prior activation steering methods such as ReFT and PEFT approaches including LoRA, while using significantly fewer parameters. Finally, we show that joint adaptation often surpasses the performance ceilings of weight and activation updates in isolation, introducing a new paradigm for efficient model adaptation.

Abstract

Fine-Tuning

Explore similar work

Weight Updates as Activation Shifts: A Principled Framework for Steering

PriFT: Guiding Supervised Fine-Tuning with Pretrained Priors

Fisher-Guided Progressive Parameter Selection for Adaptive Fine-Tuning