Neuron-Guided Fine-Tuning: Unlocking Efficient Alignment Mechanisms for Large Language Models

Zeyu Wu; Junchao Wu; Shudong Liu; Runzhe Zhan; Xin Chen; Shu Yang; Yichao Du; Longyue Wang

cs.CLSep 5, 2026

Neuron-Guided Fine-Tuning: Unlocking Efficient Alignment Mechanisms for Large Language Models

Authors: Zeyu Wu, Junchao Wu, Shudong Liu, Runzhe Zhan, Xin Chen, Shu Yang, Yichao Du, Longyue Wang,

Organizations: NLP2CT Lab, Faculty of Information Science and Computing, University of Macau · Alibaba Group · Nanjing University · KAUST ·

Abstract

Existing Supervised Fine-Tuning paradigms, particularly Full Parameter Fine-Tuning are often plagued by parameter redundancy, inconsistent data quality, and catastrophic forgetting, which current methods typically address in isolation and lack a unified optimization signal to bridge data selection, parameter updates, and knowledge preservation. To address this, we propose Neuron-Guided Fine-Tuning (NGFT), a holistic framework that leverages neuron activation patterns as a universal proxy to unify the fine-tuning lifecycle. NGFT operates via three synergistic mechanisms: (1) Adaptive Task-Specific Neuron Selection, which identifies essential neurons in a single forward pass to concentrate updates and reduce redundancy; (2) Activation-Based Data Selection, which prioritizes information-dense samples that maximize contribution to key neurons; and (3) Neuron Activation Alignment, a novel loss function that anchors activations to pre-trained states, deepening representation learning and preserving general knowledge. Experimental results across three models across both domain-specific and general benchmarks demonstrate that NGFT significantly outperforms existing mainstream fine-tuning methods in both efficiency and performance, while effectively mitigating catastrophic forgetting.

Figures & tables

Figure 1: Overview of the proposed Ngft framework. It consists of three components: (I) Task-Knowledge Neuron Selection, which identifies sparse, stably activated neurons via coefficient of variation analysis; (II) Activation-Based Data Selection, which filters representative samples based on neuron activation intensity; and (III) Neuron-Level Alignment, which updates only critical neurons using gradient masking and incorporates a Neuron Alignment Loss to achieve deep alignment, thereby improving task performance while mitigating catastrophic forgetting.

			GSM8K				DialogSum				BioInstruct
Model	Method	Metric	1%	5%	10%	Full	1%	5%	10%	Full	1%	5%	10%	Full
Qwen2.5-7B-Instruct	Zero-shot	Test	42.91				23.27				30.68
	Zero-shot	CF	–				68.62				–
	FPFT	Test	77.48	78.99	80.81	84.91	34.49	35.17	36.68	37.78	36.12	37.53	36.92	40.21
	FPFT	CF	57.24	56.25	56.31	58.09	65.96	64.33	64.33	57.88	62.18	65.85	62.30	58.10
	Strong Baseline	Test	80.81	83.01	84.22	85.75	33.70	36.24	36.98	38.24	36.45	38.34	38.96	41.21

Table 1: Performance Comparison of Ngft with Zero-Shot, FPFT, and Strong Baseline (Data Whisperer + NCFT + CE Loss, represent the most SOTA baseline for data selection, neuron selection, and loss function optimization) when training with top 1%, 5%, 10%, and full GSM8K, DialogSum, and BioInstruct data. The Test Metric measures in-domain performance, while the CF Metric assesses average performance on MMLU, BBH, and TyDiQA to reflect forgetting mitigation. Bold and underline indicates best Test and CF performance.

	GSM8K			DialogSum			BioInstruct
Method	1%	5%	10%	1%	5%	10%	1%	5%	10%
	Qwen2.5-7B-Instruct
Random	77.48 / 55.95	78.99 / 55.91	80.81 / 55.44	34.49 / 65.72	35.17 / 62.61	36.68 / 65.02	36.12 / 63.60	37.53 / 63.52	36.92 / 62.33
CCS	80.14 / 55.53	81.20 / 56.31	82.11 / 56.89	34.44 / 66.21	36.00 / 63.28	36.74 / 64.63	37.23 / 64.02	37.99 / 62.87	38.51 / 61.68
Nuggets	78.47 / 55.44	80.51 / 55.34	82.87 / 57.06	34.65 / 66.34	36.02 / 62.12	35.99 / 65.37	36.96 / 63.83	36.75 / 63.20	37.67 / 62.91
LESS	80.14 / 55.89	81.05 / 56.11	82.03 / 56.27	34.96 / 65.98	35.78 / 63.99	36.62 / 65.12	36.88 / 64.55	37.21 / 63.77	38.09 / 62.15

Table 2: Performance (Test / CF) Comparison of Ngft with Data Selection Baselines when training with top 1%, 5%, and 10% of the GSM8K, DialogSum, and BioInstruct data. The Test Metric measures in-domain performance, while the CF Metric assesses average performance on MMLU, BBH, and TyDiQA to reflect the model’s ability to mitigate forgetting. Bold indicates the best Test performance, and underline indicates the best CF performance.

Method	GSM8K	DialogSum	BioInstruct
	Qwen2.5-7B-Instruct
Random	62.17 / 63.31	29.87 / 63.08	32.12 / 62.90
MathNeuro	82.87 / 67.05	36.47 / 62.01	38.61 / 61.78
NeFT	83.57 / 65.09	37.93 / 61.84	40.47 / 62.52
NCFT	85.75 / 67.20	38.24 / 62.39	41.21 / 62.37
$\textbf{{{Ngft}}}_{\text{NS}}$	87.57 / 67.28	38.76 / 68.29	41.69 / 66.68

Table 3: Performance (Test / CF) Comparison of

\text{{Ngft}}_{\text{NS}}

with Neuron Selection Baselines on full data.

Figure 2: Impact of Neuron Alignment Loss on model performance and ablation on intervention timing.

Figure 3: Impact of neuron analysis scope (all FFN neurons or Task-Knowledge Neurons) on data selection performance.

Figure 4: Comparison of anchor activation vector effectiveness for optimizing Neuron Alignment Loss across three data selection methods: random selection, NGFT selection, and the full dataset.

Appendix figures & tables19 assets

Supplementary material from the paper’s appendix.

Appendix

Model	BioInstruct	DialogSum	GSM8K
Qwen2.5-7B-Instruct	7.20%	9.33%	14.67%
Llama3.1-8B-Instruct	5.28%	8.23%	12.18%
Mistral-7B-Instruct-v0.3	1.19%	14.42%	17.03%

Appendix

Table 4: Neuron selection ratios of

\textsc{Ngft}_{\mathrm{NS}}

across different models and tasks.

Model	BioInstruct	DialogSum	GSM8K
Qwen2.5-7B-Instruct	41.83 ( $\uparrow$ 0.61 )	38.45 ( $\uparrow$ 0.21 )	86.81 ( $\uparrow$ 1.05 )
Llama3.1-8B-Instruct	40.95 ( $\uparrow$ 0.11 )	38.77 ( $\uparrow$ 0.19 )	80.60 ( $\uparrow$ 2.96 )
Mistral-7B-Instruct-v0.3	38.80 ( $\downarrow$ 0.04 )	36.96 ( $\uparrow$ 0.09 )	58.41 ( $\uparrow$ 0.56 )

Appendix

Table 5: Experimental results of NCFT method on three models with neuron selection ratios from

\text{{Ngft}}_{\mathrm{NS}}

using full dataset.

Method	GSM8K	DialogSum	BioInstruct
Nugget	1.73	2.83	2.35
CCS	0.93	1.16	1.08
LESS	1.54	2.07	2.21
Data Whisperer	0.41	0.64	0.81
Ngft	0.08	0.11	0.12
Speedup	21.6 $\times$	25.72 $\times$	19.58 $\times$

Appendix

Table 6: STR metrics of different data selection methods on the Qwen2.5-7B-Instruct model. A lower STR indicates lower time complexity for the method.

Method	GSM8K		DialogSum		BioInstruct
Method	Accuracy	CF	ROUGE-L	CF	ROUGE-L	CF
Full Parameter	84.91	58.09	37.78	57.88	40.21	58.10
LoRA	83.40	65.83	37.61	63.31	39.57	64.74
QLoRA	82.79	66.41	37.10	64.10	38.94	65.12
DoRA	85.90	66.47	38.25	63.73	41.15	65.01
Ngft	87.64	67.10	39.41	64.21	42.27	65.34

Appendix

Table 7: Performance comparison between Ngft and PEFT methods on Qwen2.5-7B-Instruct.

Figure 5: Performance comparison of neurons selected by different Coefficient of Variation thresholds on Qwen2.5-7B-Instruct and Llama3.1-8B-Instruct.

Figure 6: Performance comparison across different intervention timings of Neural Alignment Loss on Qwen2.5-7B-Instruct and Llama3.1-8B-Instruct.

Section	GSM8K	DialogSum	BioInstruct
	Only Cross-Entropy
Task-Knowledge Neurons	$2.98\times 10^{-4}$	$4.54\times 10^{-4}$	$6.81\times 10^{-4}$
Non-Task-Knowledge Neurons	$2.64\times 10^{-4}$	$4.22\times 10^{-4}$	$6.67\times 10^{-4}$
	Cross-Entropy + Neuron Alignment Loss
Task-Knowledge Neurons	$2.71\times 10^{-4}$	$4.26\times 10^{-4}$	$6.36\times 10^{-4}$
Non-Task-Knowledge Neurons	$1.89\times 10^{-4}$	$3.77\times 10^{-4}$	$4.48\times 10^{-4}$

Appendix

Table 8: Comparison of changes in Task-Knowledge and Non-Task-Knowledge Neurons of Qwen2.5-7B-Instruct after fine-tuning with different loss functions.

Figure 7: Analysis of In-Domain Performance and Time Overhead of the Ngft Framework on Qwen2.5-7B-Instruct Under Different Configurations. Specifically,

\text{{Ngft}}_{\text{NS}}

fine-tunes only the Task-Knowledge Neurons;

\text{{Ngft}}_{\text{NS+NL}}

incorporates a Neuron Alignment Loss on this basis;

\text{{Ngft}}_{\text{NS+DS(10\%)}}

fine-tunes the aforementioned neurons using only 10% of the selected data; while

\text{{Ngft}}_{\text{NS+NL+DS(10\%)}}

combines both the 10% data selection and the Neuron Alignment Loss.

Method	GSM8K	DialogSum	BioInstruct
FPFT	84.91 / 1.00	37.78 / 1.00	40.21 / 1.00
$\text{{Ngft}}_{\text{NS}}$	87.57 / 1.11	38.82 / 1.13	41.69 / 1.15
$\text{{Ngft}}_{\text{NS+NL}}$	87.64 / 1.87	39.41 / 1.92	42.27 / 1.97
$\text{NGFT}_{\text{NS+DS}(10\%)}$	85.06 / 0.21	37.37 / 0.20	40.43 / 0.21
$\text{{Ngft}}_{\text{NS+NL+DS}(10\%)}$	86.13 / 0.72	38.16 / 0.75	40.97 / 0.81

Appendix

Table 9: Comparison of in-domain performance and time cost (Score / Normalized Time Cost) for Qwen2.5-7B-Instruct under Different Ngft Configurations.

			GSM				DialogSum				BioInstruct
Model	Method	Metric	1%	5%	10%	Full	1%	5%	10%	Full	1%	5%	10%	Full
Qwen2.5-7B-Instruct	Zero-shot	Test	42.91				23.27				30.68
		MMLU	–				74.14				–
		BBH	–				65.74				–
		TyDiQA	–				65.99				–
	FPFT	Test	77.48	78.99	80.81	84.91	34.49	35.17	36.68	37.78	36.12	37.53	36.92	40.21

Appendix

Table 10: Performance Comparison of Ngft with Zero-Shot, FPFT, and Strong Baselines (Data Whisperer + NCFT + CE Loss represent the SOTA baselines for data selection, neuron selection, and loss function optimization. ) when training with top 1%, 5%, 10%, and full GSM8K, DialogSum, and BioInstruct data.

			GSM				DialogSum				BioInstruct
Model	Method	Metric	1%	5%	10%	Full	1%	5%	10%	Full	1%	5%	10%	Full
Qwen2.5-0.5B-Instruct	Zero-shot	Test	24.03				15.95				17.22
	Zero-shot	CF	–				39.88				–
	FPFT	Test	31.18	32.67	35.92	39.30	28.95	30.72	31.58	34.88	28.39	29.68	31.71	36.40
	FPFT	CF	32.07	31.39	31.53	30.17	28.51	27.08	25.94	18.01	32.40	26.99	24.57	11.82
	Strong Baseline	Test	40.96	42.79	44.55	47.56	32.46	33.70	34.00	36.40	32.24	34.37	34.78	39.10

Appendix

Table 11: Performance (Test / CF) Comparison on Qwen2.5-0.5B-Instruct, Qwen2.5-3B-Instruct, and Qwen2.5-14B-Instruct. The Test Metric measures in-domain performance, while the CF Metric assesses average performance on MMLU, BBH, and TyDiQA to reflect the model’s ability to mitigate forgetting. Bold indicates the best Test performance, and underline indicates the best CF performance.

			GSM				DialogSum				BioInstruct
Model	Method	Metric	1%	5%	10%	Full	1%	5%	10%	Full	1%	5%	10%	Full
Qwen2.5-0.5B-Instruct	Zero-shot	Test	24.03				15.95				17.22
		MMLU	–				46.87				–
		BBH	–				32.31				–
		TyDiQA	–				40.48				–
	$\textsc{Ngft}_{\text{DS}}+\text{FPFT}$	Test	32.62	34.62	37.56	39.30	30.51	32.25	32.89	34.88	30.22	31.75	33.08	36.40

Appendix

Table 12: Comparison of performance improvements on the Qwen2.5-0.5B-Instruct model using different Ngft components individually and in combination.

			GSM				DialogSum				BioInstruct
Model	Method	Metric	1%	5%	10%	Full	1%	5%	10%	Full	1%	5%	10%	Full
Qwen2.5-3B-Instruct	Zero-shot	Test	37.44				21.70				27.50
		MMLU	–				66.36				–
		BBH	–				56.29				–
		TyDiQA	–				55.44				–
	$\textsc{Ngft}_{\text{DS}}+\text{FPFT}$	Test	62.03	64.30	65.24	67.24	34.45	36.13	36.60	37.19	37.51	38.68	39.13	41.14

Appendix

Table 13: Comparison of performance improvements on the Qwen2.5-3B-Instruct model using different Ngft components individually and in combination.

			GSM				DialogSum				BioInstruct
Model	Method	Metric	1%	5%	10%	Full	1%	5%	10%	Full	1%	5%	10%	Full
Qwen2.5-7B-Instruct	Zero-shot	Test	42.91				23.27				30.68
		MMLU	–				74.14				–
		BBH	–				65.74				–
		TyDiQA	–				65.99				–
	$\textsc{Ngft}_{\text{DS}}+\text{FPFT}$	Test	80.41	81.68	83.26	84.91	33.97	36.05	37.14	37.78	38.35	38.60	39.64	40.21

Appendix

Table 14: Comparison of performance improvements on the Qwen2.5-7B-Instruct model using different Ngft components individually and in combination.

			GSM				DialogSum				BioInstruct
Model	Method	Metric	1%	5%	10%	Full	1%	5%	10%	Full	1%	5%	10%	Full
Qwen2.5-14B-Instruct	Zero-shot	Test	54.81				34.07				38.93
		MMLU	–				79.80				–
		BBH	–				69.25				–
		TyDiQA	–				67.48				–
	$\textsc{Ngft}_{\text{DS}}+\text{FPFT}$	Test	85.75	85.82	87.42	86.43	34.27	36.53	38.00	38.82	38.45	38.62	38.86	41.52

Appendix

Table 15: Comparison of performance improvements on the Qwen2.5-14B-Instruct model using different Ngft components individually and in combination.

Model	Method	Metric	1%	5%	10%	Full	1%	5%	10%	Full	1%	5%	10%	Full
			GSM				DialogSum				BioInstruct
Mistral-7B-Instruct-v0.3	Zero-shot	Test	25.24				20.46				21.40
		MMLU	–				61.67				–
		BBH	–				65.74				–
		TyDiQA	–				54.29				–
	$\textsc{Ngft}_{\text{DS}}+\text{FPFT}$	Test	40.96	43.98	45.85	57.24	30.45	33.38	34.09	36.42	25.42	30.32	32.62	38.14

Appendix

Table 16: Comparison of performance improvements on the Mistral-7B-Instruct model using different Ngft components individually and in combination.

			GSM				DialogSum				BioInstruct
Model	Method	Metric	1%	5%	10%	Full	1%	5%	10%	Full	1%	5%	10%	Full
Llama3.2-1B-Instruct	Zero-shot	Test	12.51				14.02				22.59
		MMLU	–				46.00				–
		BBH	–				38.42				–
		TyDiQA	–				43.69				–
	$\textsc{Ngft}_{\text{DS}}+\text{FPFT}$	Test	32.15	33.20	34.41	38.66	31.70	33.32	34.52	36.92	31.96	34.63	35.52	38.50

Appendix

Table 17: Comparison of performance improvements on the Llama3.2-1B-Instruct model using different Ngft components individually and in combination.

			GSM				DialogSum				BioInstruct
Model	Method	Metric	1%	5%	10%	Full	1%	5%	10%	Full	1%	5%	10%	Full
Llama3.2-3B-Instruct	Zero-shot	Test	29.45				15.12				26.88
		MMLU	–				60.45				–
		BBH	–				58.61				–
		TyDiQA	–				62.50				–
	$\textsc{Ngft}_{\text{DS}}+\text{FPFT}$	Test	62.25	67.40	68.46	70.97	33.31	36.13	37.09	38.21	35.47	37.56	38.53	41.49

Appendix

Table 18: Comparison of performance improvements on the Llama3.2-3B-Instruct model using different Ngft components individually and in combination.

			GSM				DialogSum				BioInstruct
Model	Method	Metric	1%	5%	10%	Full	1%	5%	10%	Full	1%	5%	10%	Full
Llama3.1-8B-Instruct	Zero-shot	Test	36.22				16.66				27.53
		MMLU	–				68.51				–
		BBH	–				69.90				–
		TyDiQA	–				67.37				–
	$\textsc{Ngft}_{\text{DS}}+\text{FPFT}$	Test	63.50	72.32	73.26	77.71	34.74	35.62	36.64	38.10	34.02	36.96	37.76	40.57

Appendix

Table 19: Comparison of performance improvements on the Llama3.1-8B-Instruct model using different Ngft components individually and in combination.

LLM Fine-Tuning

LLM fine-tuning adapts pretrained large language models to tasks, domains, languages, and behaviors through additional training.

See all 754 papers

GAST: Gradient-aligned Sparse Tuning of Large Language Models with Data-layer Selection
Mar 10, 2026Kai Yao, Zhenghan Song, Kaixin Wu +5Sparse Fine-Tuning Selective Fine-Tuning
1Ant Group · 2Cornell University · University of Liverpool +2
Estimating and Orthogonalizing Unknown Pre-training Gradients for Continual Fine-tuning of Large Language Models
Sep 25, 2026Bing Wang, Changchun Li, Xin-Qiang Cai +4Continual Learning for LLMs Catastrophic Forgetting
College of Computer Science and Technology, Jilin University · Key Laboratory of Symbolic Computation and Knowledge Engineering, MoE, Jilin University · RIKEN Center for Advanced Intelligence Project +2
APE: Selective Fine-tuning with Acceptance Criteria for Language Model Adaptation
May 26, 2025Javier MarínLLM Fine-Tuning Selective Fine-Tuning

Mar 10, 2026J/K move · Enter open · S save

GAST: Gradient-aligned Sparse Tuning of Large Language Models with Data-layer Selection

Kai Yao, Zhenghan Song, Kaixin Wu, Mingjie Zhong+4

1Ant Group · 2Cornell University · University of Liverpool+2

Sparse Fine-Tuning Selective Fine-Tuning Parameter-Efficient Fine-Tuning Training Data Selection

Open paper AskPDF arXiv

Parameter-Efficient Fine-Tuning (PEFT) has become a key strategy for adapting large language models, with recent advances in sparse tuning reducing overhead by selectively updating key parameters or subsets of data. Existing approaches generally focus on two distinct paradigms: layer-selective methods aiming to fine-tune critical layers to minimize computational load, and data-selective methods aiming to select effective training subsets to boost training. However, current methods typically overlook the fact that different data points contribute varying degrees to distinct model layers, and they often discard potentially valuable information from data perceived as of low quality. To address these limitations, we propose Gradient-aligned Sparse Tuning (GAST), an innovative method that simultaneously performs selective fine-tuning at both data and layer dimensions as integral components of a unified optimization strategy. GAST specifically targets redundancy in information by employing a layer-sparse strategy that adaptively selects the most impactful data points for each layer, providing a more comprehensive and sophisticated solution than approaches restricted to a single dimension. Experiments demonstrate that GAST consistently outperforms baseline methods, establishing a promising direction for future research in PEFT strategies.

Abstract

LLM Fine-Tuning

Explore similar work

GAST: Gradient-aligned Sparse Tuning of Large Language Models with Data-layer Selection

Estimating and Orthogonalizing Unknown Pre-training Gradients for Continual Fine-tuning of Large Language Models

APE: Selective Fine-tuning with Acceptance Criteria for Language Model Adaptation