TensorMutation: Learning Weight-Space Mutations of Deep Learning Models on Neural Graphs

Zaheed Ahmed; Philip Makedonski; Jens Grabowski

cs.SENov 6, 2026

TensorMutation: Learning Weight-Space Mutations of Deep Learning Models on Neural Graphs

Authors: Zaheed Ahmed, Philip Makedonski, Jens Grabowski

Organizations: Institute of Computer Science, University of Göttingen, Germany

Abstract

Mutation testing for deep learning (DL) faces a cost-realism trade-off. The pre-training approach produces realistic but computationally expensive mutants, whereas existing post-training operators modify trained models efficiently but produce mutants that less closely resemble real faults. This study asks whether the trade-off is necessary: can the weight-space effect of a fault-informed pre-training mutation operator be predicted from a trained model, without retraining it? We answer this question with TensorMutation, which represents a trained model as a neural graph and uses a mutation-conditioned graph neural network (GNN) to predict parameter changes. The GNN is trained offline and reused across unseen models so that TensorMutation can generate executable mutants without retraining. We construct a corpus of 25,252 original-mutant pairs from 71 multilayer perceptron (MLP) classifiers using 11 fault-informed pre-training operators, including 25 newly curated subjects. TensorMutation partially reproduces the behavioral effects of pre-training mutations on nine held-out subjects, and direct weight-space supervision significantly improves weight-space fidelity. On 12 real-fault subjects, TensorMutation achieves the highest median behavioral similarity among the evaluated approaches. The improvement over the post-training tool DeepMutation++ is statistically significant (Holm-adjusted p=0.018), whereas the differences from the pre-training tool DeepCrime and the norm-matched random control are not significant. These results show that fault-informed pre-training mutations leave learnable structure in weight space that transfers to unseen models.

Figures & tables

Fig. 1: Mutant-generation workflow of TensorMutation. Given a trained original model $M_{o}$ and a mutation specification $m$ , TensorMutation constructs a mutation-aware neural graph, uses the trained GNN $f_{\phi}$ to predict parameter changes, applies deterministic architectural changes when required, and reconstructs the executable mutant $\widehat{M}_{m}$ without retraining.

Fig. 2: Mutation-aware neural-graph representation used by TensorMutation. Each original–mutant pair yields one supervised graph sample. The original model provides the node and edge input features, while the corresponding parameter differences provide the bias- and weight-delta supervision targets. The mutation specification is associated with the graph and conditions GNN processing through FiLM.

Fig. 3: Offline training of the TensorMutation GNN. The trained GNN is subsequently reused unchanged in the mutant-generation workflow of Fig. 1 .

Fig. 4: Overview of the TensorMutation study design. The study constructs and splits the mutation corpus, trains TensorMutation, evaluates behavioral and weight-space fidelity on held-out subjects, and compares the generated mutants with real faults. TM-B and TM-J denote TensorMutation trained with the behavioral objective only and with the full objective, respectively (Section IV-E ).

ID	Operator	Fault type	Configuration space
HBS	Change batch size	Suboptimal batch size	$\{b/4,\ b/2,\ 2b,\ 4b\}$
HLR	Decrease learning rate	Suboptimal learning rate	$(\sim\!0,\ lr]$ , binary search with precision $lr/10$
HNE	Change number of epochs	Suboptimal number of epochs	$[1,\ e]$ , binary search with precision $e/10$
HDB	Disable data batching	Data batching required	Batch size set to the training-set size; only if batching is used
ACH	Change activation function	Wrong type of activation function	Each layer with a non-linear activation $\times$ each other non-linear activation
ARM	Remove activation function	Missing activation function	Each layer with a non-linear activation

TABLE I: Selected DeepCrime Mutation Operators with Fault Types and Configuration Spaces

	IoU				CS
Subject	TM-B	Rand. B	TM-J	Rand. J	TM-B	TM-J
cleanml018	0.246	0.153	0.232	0.118	0.483	0.483
cleanml019	0.297	0.243	0.299	0.214	0.482	0.441
deepfd030	0.162	0.190	0.179	0.202	0.163	0.229
deepfd050	0.178	0.164	0.248	0.087	0.228	0.316
deeplocalize030	0.091	0.109	0.099	0.085	0.355	0.411

TABLE II: Behavioral similarity to the corresponding DeepCrime mutants on the test subjects

Fig. 5: Operator-level median IoU of TM-B and TM-J with the corresponding DeepCrime mutants across the test subjects. Numbers in parentheses give the number of test subjects to which each operator applies.

	Cosine		Log-norm ratio		Norm. $L_{2}$
Subject	TM-B	TM-J	TM-B	TM-J	TM-B	TM-J
cleanml018	0.146	0.422	0.046	$-0.151$	1.377	1.030
cleanml019	0.282	0.532	$-0.114$	$-0.226$	1.009	0.925
deepfd030	0.020	0.115	0.130	$-0.171$	1.717	1.125
deepfd050	0.080	0.227	$-0.175$	$-0.479$	1.074	0.983
deeplocalize030	0.030	0.156	0.156	$-0.154$	1.777	1.117

TABLE III: Weight-space fidelity to the corresponding DeepCrime deltas on the test subjects

Fig. 6: Operator-level median cosine similarity of TM-B and TM-J with the corresponding DeepCrime deltas across the test subjects. Numbers in parentheses give the number of test subjects to which each operator applies.

		Spearman $\rho$
Subject	Configurations	TM-B	TM-J
cleanml018	53	0.242	0.241
cleanml019	54	0.339	0.202
deepfd030	119	0.404	0.151
deepfd050	87	0.360	0.573
deeplocalize030	120	0.412	0.119

TABLE IV: Spearman correlation between configuration-level cosine similarity and IoU on the test subjects

Fig. 7: IoU with the real fault on the 12 subjects. Boxes show the IoU of individual mutants, and black diamonds show the subject summaries of Table V , on which the tests are based. † Subjects with fewer than 50 real-fault kills.

Subject	TM-J	DC	Rand.	DM++	KI	ceiling
	IoU with real fault					DC
defect4ml038	0.248	0.008	0.117	0.010	486	0.008
defect4ml061	0.172	0.105	0.116	0.131	120	0.131
defect4ml067	0.257	0.175	0.135	0.057	7041	0.250
defect4ml068	0.000	0.000	0.000	0.000	500	0.000
defect4ml069 †	0.000	0.000	0.000	0.000	12	0.000

TABLE V: Real-fault similarity on the 12 subjects

Comparator	W/L/T	Median $\Delta$	$p$	$p_{\mathrm{Holm}}$	$r$
DM++	9/1/2	0.112	0.006	0.018	0.927
Random	9/1/2	0.067	0.084	0.168	0.636
DeepCrime	8/2/2	0.018	0.322	0.322	0.382

TABLE VI: Paired subject-level comparisons of TM-J with each approach

	Test subjects		Real faults
Variant	IoU	Cosine	IoU
TM-J	0.232	0.227	0.181
TM-J-Full	0.223	0.214	0.140
TM-J-Narrow	0.170	0.195	0.155

TABLE VII: Source-context ablation

Mutation Testing

Mutation testing evaluates test suites by injecting controlled faults and measuring whether tests detect them.

See all 19 papers

Quality-Driven Selective Mutation for Deep Learning
Apr 24, 2026Zaheed Ahmed, Emmanuel Charleson Dapaah, Philip Makedonski +1
Institute of Computer Science, University of Göttingen Goettingen, Lower Saxony, Germany
Grounding Deep Learning Framework Testing in Data Flow
Oct 11, 2026Jiatao Ge, Yanzhou Mu, Yinglong Zou +3Automated Software Testing DNN Testing
State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China · UNIST, Ulsan, Republic of Korea
Deep4ge: DNN Training Trajectories for Fault Detection and Diagnosis
Jul 14, 2026Sigma JahanFault Diagnosis
Dalhousie University, Canada

Apr 24, 2026J/K move · Enter open · S save

Quality-Driven Selective Mutation for Deep Learning

Zaheed Ahmed, Emmanuel Charleson Dapaah, Philip Makedonski, Jens Grabowski

Institute of Computer Science, University of Göttingen Goettingen, Lower Saxony, Germany

Open paper AskPDF arXiv

Mutants support testing and debugging in two roles: (i) as test goals and (ii) as substitutes for real faults. Hard-to-kill mutants provide better guidance for test improvement, while realism is essential when mutants are used to simulate real bugs. Building on these roles, selective mutation for deep learning (DL) aims to reduce the cost of mutant generation and execution by choosing operator configurations that yield resistant and realistic mutants. However, the DL literature lacks a unified measure that captures both aspects. This study presents a probabilistic framework to quantify mutant quality along two complementary axes: resistance and realism. Resistance adapts the classical notion of hard-to-kill mutants to the DL setting using statistical killing probabilities, while realism is measured via the generalized Jaccard similarity between mutant and real-fault detectability patterns. The framework enables ranking and filtering of low-quality mutation-operator configurations without assuming a specific use case. We empirically evaluate the approach on four datasets of real DL faults. Three datasets (CleanML, DeepFD, and DeepLocalize) are used to estimate and select high-quality operator configurations, and the held-out defect4ML dataset is used for validation. Results show that quality-driven selection reduces the number of generated mutants by up to 55.6% while preserving typical levels of resistance and realism under baseline-aligned selection thresholds. These findings confirm that dual-objective selection can lower cost without compromising the usefulness of mutants for either role.

Abstract

Mutation Testing

Explore similar work

Quality-Driven Selective Mutation for Deep Learning

Grounding Deep Learning Framework Testing in Data Flow

Deep4ge: DNN Training Trajectories for Fault Detection and Diagnosis