acceptodds
Under review as a conference paper at ICLR 2027

Training-Free Protection of LLMs against Unauthorized Model Merging

Abstract

Model merging allows the specialized capabilities of a released language model to be incorporated into other models without authorization. We propose a training-free defense that adjusts model weights using parameter differences from compatible models with weaker target-domain performance. The guiding idea is to move the specialized model's parameters toward the boundary of a high-performing region in its loss landscape, retaining useful standalone performance while reducing the specialized capability preserved after subsequent merging. Global processing combines auxiliary-model weight differences to define the overall displacement. Local refinement applies alternating task-vector edits while preserving task-critical weights selected using calibration data. The procedure requires no additional gradient-based optimization or architectural changes. Experiments on mathematical and medical tasks show that the processed models retain most target-domain performance when used on their own, while contributing less capability through the evaluated merging algorithms.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
model merging

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_EomcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how to protect specialized LLMs against unauthorized model merging while preserving their standalone capabilities. The authors propose a training-free method that combines global perturbations derived from weaker auxiliary models with local weight refinement to reduce capability transfer. Experiments on mathematical and medical tasks show that the method retains most standalone performance while substantially reducing the specialized capabilities transferred through model merging.

Critical strengths
  1. The paper is trying to address unauthorized capability transfer through model merging, an important challenge for protecting the IP of open-source LLMs.
  2. The proposed method operates directly on model weights without additional gradient-based optimization or architectural modifications. This makes it potentially efficient and applicable to already fine-tuned models.
Critical weaknesses
  1. The paper discusses several relevant unmergeability defenses,such as PaRaMS, MergeLock, and MergeGuard, but does not provide direct experimental comparisons. Comparisons with existing training-free methods are essential to establish the proposed method's effectiveness.
  2. The proposed method introduces a noticeable performance cost to the protected model. Specifically, the combined method reduces GSM8K accuracy from 71.87% to 65.07% in Table 3. This performance degradation compromises the standalone utility of the protected model, potentially limiting its practical applicability.
  3. The effectiveness of the proposed method may depend on the downstream recipient model. For example, BioMistral-7B is used both to construct the global perturbation and as a merging recipient, where the method achieves a substantially higher drop ratio (1.10) than with Mistral-Instruct (0.55) under TA merging. This raises concerns about the generalizability of the method to unseen recipient models.
Questions for authors
  1. Could the authors further distinguish between merging-specific protection and general parameter fragility? For example, the combined model shows a substantially larger accuracy drop than the original model under Gaussian perturbations (0.36 vs. 0.18). It would be helpful to clarify whether the proposed method specifically disrupts model merging or primarily increases sensitivity to parameter changes.
  2. As a broader open question, to what extent can model capability and model mergeability be systematically disentangled in the parameter space?
10 comments

We sincerely thank Reviewer Eomc for the thoughtful review and constructive feedback. We appreciate your recognition of the problem and the operational appeal of protecting an already trained checkpoint. We address your concerns below.


Q1. The paper discusses several relevant unmergeability defenses,such as PaRaMS, MergeLock, and MergeGuard, but does not provide direct experimental comparisons. Comparisons with existing training-free methods are essential to establish the proposed method's effectiveness.

A1. Thank you for emphasizing the importance of direct comparisons. To represent existing training-free, parameter-level unmergeability defenses in our direct evaluation, we select PaRaMS as a representative baseline. PaRaMS operates on an already trained checkpoint through parameter transformations without protection-specific fine-tuning, making it the closest representative of this class to our setting.

We implement a Mistral-compatible adaptation of PaRaMS and evaluate it together with our method under the same model family, target task, recipient models, merging protocol, and flexible-extraction metric. Because strong suppression under a fixed merging recipe may disappear after parameter realignment, we compare both methods after the same class of public-base repair attack.

Protection methodStandalone GSM8K (%)Bio TA after repair (%)Instruct TA after repair (%)Measured processing time (s)
Our method65.0733.6250.71241.02
PaRaMS adaptation68.9153.8358.26845.67

The corresponding unprotected-merge references are 54.06% on Bio and 59.00% on Instruct. PaRaMS preserves standalone utility better, achieving 68.91% compared with 65.07% for our method. However, after public-base repair, PaRaMS recovers to within only 0.23 and 0.74 percentage points of the unprotected Bio and Instruct references, indicating that its protection is almost completely removed. In addition, PaRaMS requires 845.67 seconds for protection processing in our implementation, whereas the measured subtotal of our pipeline is 241.02 seconds. PaRaMS therefore requires approximately 3.5 times more processing time.


A2. Our method is designed as a modular protection framework rather than requiring every model owner to use the combined method. The global and local components provide different operating points:

Protection configurationStandalone GSM8K (%)Standalone loss (pp)Bio TA (%)Instruct TA (%)
Unprotected specialized model71.870.0054.0659.00
Local only68.693.1851.0856.79
Global only67.324.5534.1253.15
Combined65.076.8032.7450.46

A model owner with a strict standalone-utility requirement can use the local or global component alone, whereas the combined configuration provides the strongest protection at a larger utility cost. For example, the global-only configuration retains 67.32% GSM8K accuracy—93.7% of the original standalone accuracy—while substantially reducing Bio TA transfer from 54.06% to 34.12%. The choice can therefore be made according to the owner’s acceptable utility budget and desired protection strength.

The standalone degradation is also concentrated on the protected mathematical capability rather than indicating broad model collapse. On MMLU, the combined model achieves 58.94% average accuracy compared with 58.33% for the unprotected specialized model, showing no general-capability degradation on that benchmark.

As discussed in A1, the representative PaRaMS comparison further illustrates why standalone accuracy should be considered together with recovery resistance. PaRaMS retains more standalone utility, but public-base repair restores its post-merge performance close to the unprotected references. Our method incurs a larger standalone cost but remains substantially protected after the same class of repair. It also requires lower measured protection-processing time and introduces no additional inference-time architecture or runtime overhead after release.

Thus, our advantage is not zero-cost protection. Rather, the method provides configurable global, local, and combined operating points, preserves broader model utility, and offers a stronger recovery-aware protection profile at lower measured processing cost.

Q3. The effectiveness of the proposed method may depend on the downstream recipient model. For example, BioMistral-7B is used both to construct the global perturbation and as a merging recipient, where the method achieves a substantially higher drop ratio (1.10) than with Mistral-Instruct (0.55) under TA merging. This raises concerns about the generalizability of the method to unseen recipient models.

A3. To evaluate whether protection requires the recipient to participate in construction, we additionally test OpenOrca and OpenHermes. Neither model is used as a global auxiliary, for G1 construction, for critical-weight selection, or for protection-parameter tuning.

We apply the same protected checkpoint and the same TA configuration to every recipient, without recipient-specific reconstruction or tuning.

RecipientUsed in protection construction?Unprotected TA accuracy (%)Protected TA accuracy (%)Reduction (pp)
BioMistral-7BYes54.0632.7421.32
OpenOrcaNo59.9945.8314.16
OpenHermesNo56.3646.999.37

As expected, the strongest effect occurs for BioMistral, which participates in protection construction. However, protection consistently transfers to both construction-held-out recipients, reducing post-merge accuracy by 14.16 percentage points on OpenOrca and 9.37 points on OpenHermes without any recipient-specific tuning.

These results demonstrate that the method does not require the downstream recipient to participate in protection construction.


Q4. Could the authors further distinguish between merging-specific protection and general parameter fragility? For example, the combined model shows a substantially larger accuracy drop than the original model under Gaussian perturbations (0.36 vs. 0.18). It would be helpful to clarify whether the proposed method specifically disrupts model merging or primarily increases sensitivity to parameter changes.

A4. Thank you for highlighting this important distinction. We introduce same-source Gaussian controls to test whether generic parameter fragility or standalone degradation is sufficient to reproduce the observed protection. Starting from the same original specialized checkpoint, we apply Gaussian perturbations that reduce standalone utility below the corresponding protected checkpoints, and then evaluate all models using the same merging protocol.

CheckpointStandalone accuracy (%) ↑Bio TA accuracy (%) ↓Instruct TA accuracy (%) ↓
Global67.3234.1253.15
Global + Gaussian control62.7543.6856.29
Combined65.0732.7450.46
Combined + Gaussian control61.8941.7653.78

Here, higher standalone accuracy indicates better source-model utility, whereas lower post-merge accuracy indicates stronger protection against capability transfer.

The Gaussian controls damage standalone utility more severely than our protection operations, yet they remain substantially more mergeable. Relative to Global, its Gaussian control loses 4.57 points of standalone accuracy but increases Bio and Instruct TA accuracy by 9.56 and 3.14 points, respectively. Similarly, the Gaussian control for Combined loses 3.18 points of standalone accuracy but increases Bio and Instruct TA accuracy by 9.02 and 3.32 points.

This conservative comparison rules out generic parameter fragility or weaker standalone capability as a sufficient explanation: although the Gaussian controls suffer greater standalone degradation, they retain substantially higher post-merge accuracy and transfer more target-domain capability. The protection is therefore better explained by the structured interaction between our global–local parameter configuration and task-vector merging, rather than merely by broad sensitivity to arbitrary weight changes.

Q5. As a broader open question, to what extent can model capability and model mergeability be systematically disentangled in the parameter space?

A5. Thank you for raising this great question. Our results suggest that capability and mergeability can be partially separated. Standalone capability concerns a model’s own behavior, whereas mergeability depends on how its parameters interact with a recipient model and merging operator. Our training-free construction provides one concrete example of this distinction: the protected checkpoint remains useful on its own, but transfers substantially less capability through merging. The same-source controls help explain why this is more than ordinary model degradation: Gaussian controls lose more standalone accuracy yet remain more mergeable, and task-vector shrinkage preserves higher utility without reproducing our protection. Recovery tests further examine whether this separation survives attempts to repair the checkpoint. The global and local variants also illustrate different utility–protection trade-offs. Together, these results suggest that standalone accuracy alone does not determine capability transfer, and that checkpoint design can influence the relationship between them. How far this separation can be pushed at a given utility budget, across recipients, merging operators, and recovery resources, remains open. Our construction and experiments provide a practical starting point for investigating that frontier.


We thank the reviewer again for the constructive suggestions. We hope that the additional experiments and clarifications have addressed your concerns, and we would be grateful if you could reconsider your assessment in light of the new evidence presented in our rebuttal.

Reviewer_6e8cOpenReview ↗
Rating
3(+1)
Confidence
1(−2)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies how to retain a released specialized LLM’s standalone utility while reducing capability transfer through subsequent model merging. Its training-free procedure combines a global displacement toward compatible auxiliary models with alternating local task-vector edits, preserving weights selected by intersected Wanda rankings on calibration data. On MetaMath-Mistral-7B, the combined model scores 65.07% on GSM8K versus 71.87% before protection. At the default merging settings, it removes more than half of the accuracy gain over the matched base-model merge across three operators and two recipients. A separate Llama-3-8B experiment on MedQA provides evidence beyond the main model family and domain. Component ablations, MMLU evaluation, weaker unprotected controls, merging-coefficient variation, Gaussian perturbations, and one post-merge LoRA recovery experiment characterize the observed trade-off.

Critical strengths
  1. The method addresses a clearly defined release-time objective through direct weight processing, without additional gradient-based optimization or architectural changes. Separating global displacement, local edits, and critical-weight preservation makes the construction understandable and allows their contributions to be examined in Tables 1–3. Proposition 3.1 is presented with explicit assumptions and appropriately qualified as geometric motivation, rather than a guarantee for the implemented defense.
  2. The experiments provide useful evidence that reduced post-merge performance is not explained by standalone source accuracy alone. In Table 7, the protected model has higher standalone GSM8K accuracy than both weaker unprotected controls but lower accuracy after every evaluated merge. MMLU shows no consistent degradation on that benchmark, and the separate 8B medical experiment broadens the evidence. These controls support the empirical effect within the evaluated settings.
Critical weaknesses
  1. Robustness to recovery before merging remains insufficiently established. Section 4.3 varies the downstream coefficient, and Section 4.2 tests one post-merge LoRA procedure, but neither tests an adversary that first modifies the released checkpoint. A concrete concern follows from Eq. (4): global processing is an affine map, exactly invertible when the auxiliary checkpoints and coefficients are known and their sum is not 1. For the reported coefficients, the original specialized weights equal (Theta_global - 0.3 Theta_1 - 0.05 Theta_2) / 0.65. This observation applies to the global-only variant, and it does not establish recovery of the combined model, whose local edits can discard information. It makes an evaluation of auxiliary-based recovery against the combined checkpoint particularly relevant. The 100-example LoRA result (32.74% to 1.90%) establishes failure of that configuration, but does not resolve this concern. The paper acknowledges that its recovery experiment has limited scope, and broader practical protection still needs evidence.
  2. The additional protection achieved at a fixed utility and computational cost is unclear. The experiments compare components and weaker unprotected checkpoints, but do not directly compare against the existing defenses discussed in the paper or simple perturbation/shrinkage controls matched for standalone target accuracy. Table 7 is useful evidence, yet different source checkpoints do not isolate the effect of the proposed construction at equal utility. This matters because the combined method costs 6.80 percentage points on GSM8K, and its strongest main results use BioMistral, which also contributes to protection construction. Results for the unseen Mistral-Instruct recipient are encouraging but weaker. Protection-parameter sensitivity and selection are not reported; Section 4.3 varies the merging coefficient, not the protection coefficients. Missing calibration details, the coefficient used to construct G1, repeated-run uncertainty for stochastic steps, and processing cost further limit reproducibility and assessment of the trade-off.
Questions for authors
  1. Recovery before merging: with the auxiliary checkpoints and protection coefficients available, how effective is inverse-global processing or an auxiliary-based approximation when applied to the released combined model before merging? Please distinguish the global-only variant from the complete method, state the attacker’s assumed resources, and report a recovery comparison under a defined budget. This would materially affect my assessment of robustness.
  2. Post-merge recovery: can you reproduce the 32.74% to 1.90% LoRA collapse across training seeds and examine whether validation-selected training settings or checkpoints recover useful accuracy? Section 4.2 offers a hypothesis, but does not validate it. Loss curves and generation examples would help distinguish consistent recovery resistance from optimization or evaluation failure.
  3. Comparative benefit: how does the complete method compare with a representative compatible prior defense and simple weight-editing controls at similar standalone accuracy? Please report absolute standalone and post-merge accuracies, alongside drop ratios, preferably including a recipient not used in construction. This would clarify the method’s incremental contribution.
  4. Parameter selection and reproducibility: how were the protection coefficients, Wanda selection fraction, and group size chosen, and was selection separated from the evaluation sets? Please specify calibration sample count/split and the TA coefficient used for G1, and provide a focused sensitivity comparison with uncertainty for stochastic runs and processing cost. The paper states that the local-refinement configuration is reused for MedQA. Please additionally report the global coefficients and calibration setup for that experiment.
12 comments

We sincerely thank Reviewer 6e8c for the detailed review and concrete suggestions. We appreciate your recognition of the direct weight-processing approach, the qualified geometric motivation, and the controls separating standalone accuracy from transfer. We address your recovery, comparison, and reproducibility questions below.


Q1. Robustness to recovery before merging remains insufficiently established. Section 4.3 varies the downstream coefficient, and Section 4.2 tests one post-merge LoRA procedure, but neither tests an adversary that first modifies the released checkpoint. A concrete concern follows from Eq. (4): global processing is an affine map, exactly invertible when the auxiliary checkpoints and coefficients are known and their sum is not 1. For the reported coefficients, the original specialized weights equal (Theta_global - 0.3 Theta_1 - 0.05 Theta_2) / 0.65. This observation applies to the global-only variant, and it does not establish recovery of the combined model, whose local edits can discard information. It makes an evaluation of auxiliary-based recovery against the combined checkpoint particularly relevant. The 100-example LoRA result (32.74% to 1.90%) establishes failure of that configuration, but does not resolve this concern. The paper acknowledges that its recovery experiment has limited scope, and broader practical protection still needs evidence. (part1/2)

A1. Thank you for highlighting this important recovery concern. In our intended release-time threat model, the attacker receives the protected checkpoint but not the instance-specific auxiliary checkpoints and coefficients used by the model owner. These choices can differ across protected releases and need not be disclosed. Therefore, exact auxiliary-informed inversion is not ordinarily available to an attacker.

Nevertheless, we first evaluate this stronger-than-intended white-box attack by granting the attacker the exact auxiliary checkpoints and coefficients. The attacker applies the algebraic inverse directly to the final protected model before merging. Although the global affine component is exactly reversible in isolation, the final released model additionally contains coordinate-wise local refinement that is not canceled by the global inverse.

Source checkpoint used for merging with BioMistralTA accuracy (%)TIES accuracy (%)DARE-TIES accuracy (%)
Unprotected specialized model54.0640.7950.34
Our protected model32.7418.9531.92
Auxiliary-informed inverse of our protected model36.8623.7133.68

The auxiliary-informed inverse provides only partial recovery: +4.12 points under TA, +4.76 points under TIES, and +1.76 points under DARE-TIES. The recovered models remain 17.20, 17.08, and 16.66 percentage points below the corresponding unprotected references. Equivalently, the attack recovers only 19.3%, 21.8%, and 9.6% of the performance gap introduced by protection.

These results show the complementary roles of the two protection stages. The inverse cancels the known global affine displacement, but it does not remove the subsequent non-affine local edits. Consequently, knowledge of the exact auxiliary mixture remains insufficient to reconstruct the mergeability of the original specialized model.

We further broaden the practical recovery evaluation beyond this white-box inverse attack. First, we expand the original post-merge fine-tuning experiment across all three merging algorithms, three recovery-data budgets, and both parameter-efficient and full-parameter updates. All recovery experiments start from the corresponding merge produced using our complete protection method.

Recovery methodRecovery examplesUpdated parametersTA accuracy (%)TIES accuracy (%)DARE-TIES accuracy (%)
No recovery0None32.7418.9531.92
LoRA SFT100Rank-16 LoRA1.901.752.02
LoRA SFT200Rank-16 LoRA1.860.971.77
LoRA SFT300Rank-16 LoRA2.181.821.46
Full-parameter SFT300All parameters1.951.531.62

Increasing the LoRA recovery set from 100 to 200 and 300 examples does not restore mathematical capability under any merging algorithm. Moreover, full-parameter SFT on 300 examples for one epoch also remains at low accuracy. Thus, the failed recovery is not specific to one merging algorithm, one 100-example dataset, or the limited capacity of a low-rank adapter.

Q1. Robustness to recovery before merging remains insufficiently established. Section 4.3 varies the downstream coefficient, and Section 4.2 tests one post-merge LoRA procedure, but neither tests an adversary that first modifies the released checkpoint. A concrete concern follows from Eq. (4): global processing is an affine map, exactly invertible when the auxiliary checkpoints and coefficients are known and their sum is not 1. For the reported coefficients, the original specialized weights equal (Theta_global - 0.3 Theta_1 - 0.05 Theta_2) / 0.65. This observation applies to the global-only variant, and it does not establish recovery of the combined model, whose local edits can discard information. It makes an evaluation of auxiliary-based recovery against the combined checkpoint particularly relevant. The 100-example LoRA result (32.74% to 1.90%) establishes failure of that configuration, but does not resolve this concern. The paper acknowledges that its recovery experiment has limited scope, and broader practical protection still needs evidence. (part 2/2)

A1. Second, we evaluate a pre-merge repair attack using only the released protected checkpoint and the publicly available base model. The attacker uses the observable parameter relationship between these accessible checkpoints to realign the protected model before applying model merging. This attack requires neither the original specialized checkpoint nor the private auxiliary models, coefficients, calibration data, or critical-weight mask.

Source used for merging with BioMistralTA accuracy (%)TIES accuracy (%)DARE-TIES accuracy (%)
Unprotected specialized model54.0640.7950.34
Our protected model32.7418.9531.92
Protected model + public-base alignment33.6819.7532.36

Public-base alignment provides only marginal recovery: +0.94 points for TA, +0.80 points for TIES, and +0.44 points for DARE-TIES. The repaired models remain 20.38, 21.04, and 17.98 points below the corresponding unprotected references.

Together, these experiments cover three distinct recovery routes: auxiliary-informed white-box inversion before merging, public-base repair before merging, and LoRA or full-parameter recovery after merging. None restores substantial target-domain capability under the evaluated information and optimization budgets. The conclusion also holds consistently across TA, TIES, and DARE-TIES. These results provide substantially broader evidence that the complete global–local protection is robust to both pre-merge and post-merge recovery attempts, rather than being fragile under only one fixed merging recipe or one lightweight fine-tuning configuration.


Q2. The additional protection achieved at a fixed utility and computational cost is unclear. The experiments compare components and weaker unprotected checkpoints, but do not directly compare against the existing defenses discussed in the paper or simple perturbation/shrinkage controls matched for standalone target accuracy. Table 7 is useful evidence, yet different source checkpoints do not isolate the effect of the proposed construction at equal utility. Results for the unseen Mistral-Instruct recipient are encouraging but weaker. Protection-parameter sensitivity and selection are not reported; Section 4.3 varies the merging coefficient, not the protection coefficients. Missing calibration details, the coefficient used to construct G1, repeated-run uncertainty for stochastic steps, and processing cost further limit reproducibility and assessment of the trade-off. (part 1/3)

A2. We address them through a same-source shrinkage control, a representative existing training-free defense, construction-held-out recipients, direct sensitivity analyses of the protection parameters, explicit calibration details, repeated protection constructions, and measured processing cost.

Same-source weight-editing control: Table 7 compares different source checkpoints and therefore does not isolate the effect of the proposed construction at similar source utility. We additionally apply task-vector shrinkage directly to the same original specialized model. We use a shrinkage coefficient of 0.60, which retains higher standalone accuracy than our method and therefore provides a conservative utility control.

Protection methodStandalone GSM8K (%)Bio TA (%)Instruct TA (%)
Unprotected specialized model71.8754.0659.00
Our method65.0732.7450.46
Task-vector shrinkage, coefficient 0.6067.6353.3758.53

Task-vector shrinkage retains 2.56 points more standalone accuracy than ours. However, its post-merge accuracy is 20.63 points higher on Bio and 8.07 points on Instruct. Thus, simply shrinking the original task vector does not reproduce the protection achieved by our global–local construction, even though the control retains better standalone utility.

Q2. The additional protection achieved at a fixed utility and computational cost is unclear. The experiments compare components and weaker unprotected checkpoints, but do not directly compare against the existing defenses discussed in the paper or simple perturbation/shrinkage controls matched for standalone target accuracy. Table 7 is useful evidence, yet different source checkpoints do not isolate the effect of the proposed construction at equal utility. This matters because the combined method costs 6.80 percentage points on GSM8K, and its strongest main results use BioMistral, which also contributes to protection construction. Results for the unseen Mistral-Instruct recipient are encouraging but weaker. Protection-parameter sensitivity and selection are not reported; Section 4.3 varies the merging coefficient, not the protection coefficients. Missing calibration details, the coefficient used to construct G1, repeated-run uncertainty for stochastic steps, and processing cost further limit reproducibility and assessment of the trade-off. (part 2/3)

A2. Comparison with a representative existing defense: We select PaRaMS as a representative training-free baseline because it is also a parameter-level defense applied directly to an already trained checkpoint and requires no protection-specific fine-tuning. Because strong suppression under a fixed merging recipe may disappear after parameter realignment, we compare both methods after public-base repair.

Protection methodStandalone GSM8K (%)Bio TA after repair (%)Instruct TA after repair (%)Measured processing time (s)
Our method65.0733.6250.71241.02
PaRaMS adaptation68.9153.8358.26845.67

The corresponding unprotected-merge references are 54.06% on Bio and 59.00% on Instruct. PaRaMS preserves standalone utility better, losing 2.96 points relative to the 71.87% unprotected specialized model, whereas our method loses 6.80 points. However, after public-base repair, PaRaMS recovers to within 0.23 and 0.74 points of the unprotected Bio and Instruct references. In contrast, our method remains 20.44 and 8.29 points below those references.

Thus, PaRaMS provides better standalone utility, whereas our method provides substantially stronger resistance to public-base recovery. The measured PaRaMS processing time is approximately 3.5 times the measured subtotal of our protection pipeline. Because the timing scopes are not completely identical, we report the absolute measurements rather than claiming a hardware-independent speedup.

Generalization beyond construction-involved recipients: To test whether the strongest result depends on BioMistral’s participation in protection construction, we additionally evaluate OpenOrca and OpenHermes. Neither recipient is used as an auxiliary model, for critical-weight selection, or for protection-parameter tuning. We apply the same protected checkpoint and TA configuration without recipient-specific reconstruction or tuning.

RecipientUsed in protection construction?Unprotected TA accuracy (%)Protected TA accuracy (%)Reduction (pp)
OpenOrcaNo59.9945.8314.16
OpenHermesNo56.3646.999.37

Protection therefore transfers to both construction-held-out recipients, reducing post-merge accuracy by 14.16 and 9.37 points without recipient-specific tuning. These results show that the method does not require the downstream recipient to participate in protection construction, although the protection magnitude remains recipient dependent.

Protection-parameter sensitivity and stability: We clarify that Section 4.3 varies the downstream merging coefficient, whereas the following experiments vary the parameters used to construct the protected checkpoint itself. All other construction, merging, and evaluation settings are held fixed. Each configuration is independently constructed and evaluated over three runs; the tables report the three-run means, and we will include the corresponding per-run values and sample standard deviations in the revised appendix.

We first vary the global auxiliary coefficients while fixing the BioMistral coefficient at 0.30:

Bio coefficientTrendyol coefficientStandalone GSM8K (%)Bio TA (%)Instruct TA (%)
0.300.0567.3234.1253.15
0.300.1067.2133.9953.11
0.300.2066.3733.8652.98

Increasing the Trendyol coefficient from 0.05 to 0.20 changes standalone accuracy by 0.95 points, Bio TA accuracy by 0.26 points, and Instruct TA accuracy by 0.17 points. The protection effect therefore remains stable across the evaluated global-coefficient range, while the stronger displacement introduces a modest additional standalone cost.

Reviewer_pC3EOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a training-free defense against unauthorized merging of specialized language models. It combines a global displacement toward weaker auxiliary models with alternating task-vector edits, while preserving task-important weights identified through calibration-based Wanda scores. Experiments on mathematical and medical tasks evaluate standalone utility and capability transfer through Task Arithmetic, TIES, and DARE-TIES. The combined model reduces transferred mathematical capability across the tested configurations, at a 6.80-point standalone GSM8K accuracy cost. Additional experiments examine component contributions, merging coefficients, lightweight recovery fine-tuning, and general capability preservation.

Critical strengths

The problem is well motivated, and the global–local framework is straightforward to follow. Applying protection directly to an already trained checkpoint is operationally appealing. Component comparisons, MMLU evaluation, and weaker-model controls provide useful evidence that reduced transferability is not simply explained by lower standalone accuracy. The additional medical-domain experiment also offers initial evidence beyond the primary mathematical setting.

Critical weaknesses
  1. The fundamental concern is whether the method provides protection against adaptive merging rather than fragility under fixed merging recipes. Attacker knowledge and recovery resources are insufficiently specified. The global transformation is algebraically reversible when its auxiliary weights and coefficients are known; this does not establish a bypass of the combined defense, but motivates explicit recovery tests. Three merging algorithms, a limited coefficient sweep, and one 100-example LoRA recovery configuration do not adequately address this threat.

  2. The advantage over existing defenses remains unclear. Although PaRaMS, MergeLock, MergeBarrier, TRAP2, and MergeGuard are discussed, none is included in the experimental comparisons. Training-free operation alone therefore does not establish the contribution. Given the 6.80-point standalone GSM8K loss, comparisons at matched utility budgets, together with protection cost, are needed to determine whether the proposed method offers a favorable protection–utility trade-off.

  3. The evidence for generalization is limited. BioMistral participates in both global construction and critical-weight selection, yet also serves as the recipient producing the strongest protection results. The unseen Mistral-Instruct recipient shows substantially weaker protection, while the medical extension tests only Task Arithmetic. More held-out recipients and clearer reporting of calibration splits and coefficient selection would help establish whether the approach generalizes beyond configurations closely related to its construction.

Questions for authors
  1. What does the attacker know, and how does the combined defense withstand attempts to undo auxiliary mixing or repair the checkpoint before merging?

  2. How does the method compare with existing defenses at matched standalone accuracy and protection cost?

6 comments

We sincerely thank Reviewer pC3E for the thoughtful review and constructive feedback. We appreciate your recognition of the release-time objective and the global–local framework. Your comments highlight the importance of adaptive attacks, direct comparisons, and generalization beyond construction-related recipients. We address each point below.


Q1. The fundamental concern is whether the method provides protection against adaptive merging rather than fragility under fixed merging recipes. What does the attacker know, and what resources are available for coefficient search and recovery? The global transformation is algebraically reversible when its auxiliary weights and coefficients are known; this does not establish a bypass of the combined defense, but motivates explicit recovery tests.

A1. Thank you for raising this important question. In our intended release-time threat model, only the processed checkpoint is published: the attacker may inspect its weights and adaptively choose recipient models, merging algorithms, coefficients, and recoveryprocedures, but has no access to the original unprotected specialized model. For each protected release, the model owner can also keep the exact auxiliary models, global coefficients, calibration data, and local-refinement configuration private; these choices may differ across protected models and need not be disclosed. Therefore, an attacker cannot directly apply the algebraic inverse without knowing the instance-specific protection configuration.


Q2. Three merging algorithms, a limited coefficient sweep, and one 100-example LoRA recovery configuration do not adequately address this threat.

A2. We expanded the original recovery experiment across three merging algorithms, three recovery-data budgets, and both parameter-efficient and full-parameter updating. All recovery experiments start from the corresponding merge produced using our complete protection method.

Recovery methodRecovery examplesUpdated parametersTA accuracy (%)TIES accuracy (%)DARE-TIES accuracy (%)
No recovery0None32.7418.9531.92
LoRA SFT100Rank-16 LoRA1.901.752.02
LoRA SFT200Rank-16 LoRA1.860.971.77
LoRA SFT300Rank-16 LoRA2.181.821.46
Full-parameter SFT300All parameters1.951.531.62

Increasing the LoRA recovery set from 100 to 200 and 300 examples does not restore the protected model’s mathematical capability under any of the three merging algorithms. More importantly, updating all model parameters on 300 examples for one epoch also remains low accuracy. Therefore, the failed recovery is not specific to one merging algorithm, one 100-example dataset, or the limited capacity of a low-rank adapter.

We additionally evaluated a pre-merge repair attack using only the released protected checkpoint and the publicly available base model.

The attacker uses the observable parameter difference between these two accessible checkpoints to realign the protected checkpoint before applying model merging. This attack does not require the original unprotected specialized checkpoint, the private auxiliary models or coefficients, the calibration data, or the critical-weight mask.

Source used for merging with BioMistralTA accuracy (%)TIES accuracy (%)DARE-TIES accuracy (%)
Unprotected specialized model54.0640.7950.34
Our protected model32.7418.9531.92
Protected model + public-base alignment33.6819.7532.36

Public-base alignment provides only marginal recovery across all three merging algorithms: +0.94 percentage points for TA, +0.80 points for TIES, and +0.44 points for DARE-TIES. After the repair attempt, the resulting accuracies remain 20.38, 21.04, and 17.98 points below their corresponding unprotected-merge references, respectively. Thus, the parameter relationship observable from the released protected checkpoint and public base model is insufficient to reconstruct the protected target-domain capability.

Together, our experiments evaluate two distinct recovery routes: post-merge recovery through LoRA or full-parameter SFT, and pre-merge repair using the publicly available base model. Neither route restores substantial mathematical capability under the evaluated information and optimization budgets, and the conclusion holds consistently across TA, TIES, and DARE-TIES.

Q3. The advantage over existing defenses remains unclear. Although PaRaMS, MergeLock, MergeBarrier, TRAP2, and MergeGuard are discussed, none is included in the experimental comparisons. Training-free operation alone therefore does not establish the contribution. Given the 6.80-point standalone GSM8K loss, comparisons at matched utility budgets, together with protection cost, are needed to determine whether the proposed method offers a favorable protection–utility trade-off.

A3. We select PaRaMS as a representative baseline because it is a parameter-level defense applied directly to an already trained checkpoint, requires no protection-specific fine-tuning, and is therefore operationally the closest to our setting. We implemented a Mistral-compatible adaptation of PaRaMS and evaluated it together with our method under the same flexible-extraction protocol.

We compare the two methods after a public-base repair attack. The attacker uses a publicly available base checkpoint as an alignment anchor to restore the neuron and attention-component correspondence of the released protected model before applying the same TA merging operation. This attack requires neither the original unprotected specialized checkpoint nor private protection information.

All results below are obtained from our common flexible-extraction reevaluation protocol.

Released source modelStandalone GSM8K (%)Bio TA after public-base repair (%)Instruct TA after public-base repair (%)
Unprotected specialized model71.8754.0659.00
Our protected model65.0733.6250.71
PaRaMS adaptation68.9153.8358.26

PaRaMS preserves standalone utility better than our method: its standalone accuracy is 68.91%, corresponding to a 2.96-point loss relative to the unprotected specialized model, whereas our protected model achieves 65.07%, corresponding to a 6.80-point loss. However, the comparison changes substantially after public-base repair.

On Bio, public-base repair restores the PaRaMS-protected merge to 53.83%, only 0.23 percentage points below the unprotected reference of 54.06%. On Instruct, it restores PaRaMS to 58.26%, only 0.74 points below the unprotected reference of 59.00%. Thus, PaRaMS retains better standalone utility, but its protection is almost completely removed by the evaluated public-base repair.

In contrast, our protected model remains at 33.62% on Bio and 50.71% on Instruct after the same class of repair. These results remain 20.44 and 8.29 percentage points below the corresponding unprotected references. Compared directly with repaired PaRaMS, our method yields 20.21-point and 7.55-point lower post-merge accuracies on Bio and Instruct, respectively.

The comparison therefore reveals a clear trade-off. PaRaMS better preserves the released model’s standalone capability, whereas our method incurs a larger utility cost but provides substantially stronger resistance to public-base recovery. Our advantage lies not in fixed-recipe suppression or standalone utility, but in maintaining reduced capability transfer after the attacker attempts to realign the protected checkpoint using publicly available information.

We also measured the processing cost of the two protection pipelines:

Protection methodMeasured processing time (s)
Our method241.02
PaRaMS adaptation845.67

Under our implementations, the recorded PaRaMS processing time is approximately 3.5 times the measured subtotal of our protection pipeline. Because the measured scopes are not completely identical, we report the absolute costs rather than claiming a hardware-independent speedup.

Overall, the representative PaRaMS comparison shows that the two methods optimize different aspects of the protection–utility trade-off. PaRaMS can be almost completely removed by public-base realignment. Our method remains substantially less mergeable after the same recovery attempt and has a lower measured processing cost. This recovery-aware comparison clarifies our contribution beyond the training-free property alone.

Q4. The evidence for generalization is limited. BioMistral participates in both global construction and critical-weight selection, yet also serves as the recipient producing the strongest protection results. The unseen Mistral-Instruct recipient shows substantially weaker protection, while the medical extension tests only Task Arithmetic. More held-out recipients and clearer reporting of calibration splits and coefficient selection would help establish whether the approach generalizes beyond configurations closely related to its construction.

A4. To test whether protection depends on the construction-involved BioMistral recipient, we additionally evaluated two recipients that do not participate in any stage of protection construction: OpenOrca and OpenHermes. Neither model is used as an auxiliary model, for Wanda-based critical-weight selection, or for protection-parameter tuning. We apply exactly the same protected checkpoint and TA configuration to both recipients, without recipient-specific reconstruction or protection tuning.

RecipientUsed in protection construction?Unprotected TA accuracy (%)Protected TA accuracy (%)Reduction (pp)
OpenOrcaNo59.9945.8314.16
OpenHermesNo56.3646.999.37

Protection consistently transfers to both construction-held-out recipients, reducing post-merge accuracy by 14.16 percentage points on OpenOrca and 9.37 points on OpenHermes. These reductions are obtained without using either recipient during protection construction or parameter selection. The results therefore demonstrate that the method does not require the downstream recipient to participate in protection construction.

We also clarify the separation between protection construction, attack selection, and evaluation. The evaluated configuration uses 128 GSM8K training examples for Wanda calibration and critical-weight selection. A disjoint set of 512 GSM8K training examples is used for validation and downstream merging or recovery configuration selection.

Data splitNumber of examplesPurpose
Protection calibration128Wanda statistics and critical-weight selection
Validation512Merging and recovery configuration selection

The global coefficients, Wanda selection fraction, and local group size are fixed before evaluating the held-out recipients, and the same settings are shared across BioMistral, OpenOrca, and OpenHermes. Thus, the additional recipient results do not arise from test-recipient-tuning.

Finally, the medical extension changes the model family, protected capability, recipient model, auxiliary models, and evaluation benchmark. It therefore provides complementary evidence that the protection mechanism transfers beyond the original Mistral-based mathematical setting. Together, the held-out-recipient and medical-domain results support both recipient-level and cross-family generalization under the evaluated merging configurations.


Q5. What does the attacker know, and how does the combined defense withstand attempts to undo auxiliary mixing or repair the checkpoint before merging?

A5. As clarified in A1, the attacker has full access to the released protected checkpoint, the publicly available base and recipient models, and the general protection algorithm. The attacker may adaptively choose the merging algorithm, merging coefficients, and recovery procedure. However, the attacker does not have access to the original unprotected specialized checkpoint or the instance-specific protection information, including the exact auxiliary models and coefficients, calibration data, and critical-weight mask. These private choices can differ across protected releases.

Under the intended release-time threat model, we evaluate two recovery routes. First, post-merge recovery using 100–300 examples with rank-16 LoRA, as well as full-parameter SFT on 300 examples, fails to restore mathematical capability across TA, TIES, and DARE-TIES: all recovered accuracies remain approximately 1%–2%. Second, a pre-merge repair based only on the released checkpoint and publicly available base model provides marginal recovery. Across TA, TIES, and DARE-TIES, public-base alignment improves accuracy by only 0.94, 0.80, and 0.44 percentage points, reaching 33.68%, 19.75%, and 32.36%, respectively. These values remain 20.38, 21.04, and 17.98 points below the corresponding unprotected references.

Thus, neither post-merge SFT nor pre-merge public-base alignment substantially restores target-domain capability within the tested budgets. We also grant exact auxiliaries and coefficients, which are not assumed available in our threat model. With this information, inversion of the complete checkpoint yields only partial recovery because local edits remain. These results support resistance to the tested attacks, not arbitrary white-box recovery or unlimited training.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Rebuttal Summary

Dear Program Chairs, Senior Area Chairs, Area Chairs, and Reviewers,

We sincerely appreciate the efforts of the Program Chairs, Senior Area Chairs, and especially the Area Chair in coordinating the review process. We also thank all reviewers for their detailed reviews and constructive suggestions.

We provide this summary to assist the Area Chair in assessing the additional evidence and clarifications presented during the discussion. The main concerns centered on recovery before and after merging, comparisons with existing defenses, the standalone-utility trade-off, generalization beyond construction-related recipients, and reproducibility.

Additional experiments

We conducted the following experiments in response to these concerns:

  • Pre-merge recovery with auxiliary information and public checkpoints. Following Reviewers pC3E and 6e8c, we evaluated auxiliary-informed inversion of the complete protected checkpoint, granting the attacker the exact auxiliary models and global coefficients. This attack recovers only 9.6%–21.8% of the protection-induced performance gap across TA, TIES, and DARE-TIES. We also evaluated public-base alignment, which improves post-merge accuracy by only 0.94, 0.80, and 0.44 percentage points, respectively. These tests distinguish the algebraically reversible global component from the complete protection method and evaluate concrete recovery routes under defined information and optimization budgets.

  • Broader post-merge recovery evaluation. We expanded the original recovery experiment to three merging algorithms, recovery sets of 100, 200, and 300 examples, and both rank-16 LoRA and full-parameter SFT. Recovery configurations were independently trained with three seeds, and intermediate checkpoints were selected using held-out validation performance. Reported recovered accuracies remain between 0.97% and 2.18%. We additionally inspected decreasing, finite training losses and answer extractability, addressing concerns that the low accuracy might result from an obvious optimization or evaluation failure. These experiments respond to Reviewers pC3E and 6e8c.

  • Direct comparisons with PaRaMS and MergeGuard. We implemented a Mistral-compatible PaRaMS adaptation and, following Reviewer Eomc’s follow-up, reproduced MergeGuard. Comparisons use a common evaluation protocol and include public-base repair. The original global-only variant achieves 35.71% Bio TA and 52.89% Instruct TA after repair, compared with 52.63% and 57.18% for MergeGuard, and 53.83% and 58.26% for PaRaMS. These results provide direct evidence of stronger post-repair protection in the evaluated setting, alongside explicit reporting of standalone utility.

  • An additional utility-preserving operating point. To address Reviewer Eomc’s remaining utility concern, we adjusted the global coefficients to 0.20 for BioMistral and 0.05 for Trendyol. Standalone GSM8K accuracy increases from 67.32% to 68.75%, placing it 0.16 percentage points below PaRaMS and 0.78 points above MergeGuard. The same setting achieves 35.37% after TA merging, corresponding to a normalized post-merge drop ratio of 0.96 relative to the matched base-model merge. This demonstrates that stronger standalone utility can be retained while preserving substantial merging suppression.

  • Same-source shrinkage and Gaussian controls. Following Reviewers 6e8c and Eomc, we added controls derived from the same original specialized checkpoint. Task-vector shrinkage retains higher standalone accuracy but removes only 4% and 3% of the specialized-model gain after Bio and Instruct TA merging. The Gaussian controls incur greater standalone degradation yet remain more mergeable than the protected checkpoints. These results support the contribution of the structured protection beyond simple shrinkage or generic parameter damage.

  • Construction-held-out recipients. Following all three reviewers, we evaluated OpenOrca and OpenHermes, neither of which participates in auxiliary selection, critical-weight selection, or protection-parameter tuning. Applying the same protected checkpoint and TA configuration reduces post-merge accuracy by 14.16 and 9.37 percentage points, respectively. These results support transfer to recipients outside protection construction, while documenting that protection strength remains recipient dependent.

  • Protection-parameter sensitivity and repeated constructions. We varied the global auxiliary coefficients, Wanda selection fraction, and local group size, independently constructing and evaluating each configuration over three runs. The tested Trendyol-coefficient sweep changes standalone accuracy by at most 0.95 percentage points; the local-refinement sweep changes it by at most 0.44 points. These experiments directly vary protection-construction parameters, supplementing the original downstream merging-coefficient sweep.

  • Measured protection-processing costs. We reported processing measurements of 97.53 s for global-only processing, 241.02 s for the complete pipeline’s measured subtotal, 717.16 s for MergeGuard, and 845.67 s for PaRaMS. We specified the measurement scope and acknowledged differences in timing coverage, presenting these as implementation-specific costs rather than a hardware-independent speedup or a perfectly matched utility–cost frontier.

Additional clarifications

We also provided the following explanations and reporting details:

  • Release-time threat model. We specified access to the released protected checkpoint, public base and recipient models, and the general algorithm, while distinguishing these from the original specialized checkpoint and instance-specific protection information. We explicitly acknowledged that the global-only affine transformation is reversible when its exact auxiliary mixture is known. Recovery claims for the complete method are limited to the evaluated attacks and budgets.

  • Calibration, validation, and construction settings. We specified 128 GSM8K training examples for Wanda calibration and a disjoint 512-example validation set for configuration selection. We detailed G1 construction at TA coefficient 0.10, the 20% Wanda-ranking intersection, and 2% local grouping, with settings fixed before held-out-recipient evaluation.

  • Medical-domain reproducibility. We supplied the Llama-3 MedQA configuration, including the auxiliary coefficients 0.30 and 0.10, 128-example calibration set, G1 construction, and reused local-refinement settings. This complements the mathematical experiments with a different model family and protected domain.

  • Standalone capability versus mergeability. We explained the configurable utility–protection trade-off and the role of the same-source controls. The original MMLU evaluation remains within 58.28%–59.30% across the reported protected variants, compared with 58.33% for the specialized reference. These results support preservation of broader utility on that benchmark, without implying zero target-domain cost or unconditional protection against arbitrary recovery.

Discussion outcome and scores

  • Reviewer 6e8c thanked us for the clarifications and increased the rating from 2 to 3.

  • Reviewer Eomc acknowledged that the rebuttal addressed parts of their concerns and that the MergeGuard comparison “provides a better comparison with existing baselines.” The reviewer retained the rating of 2, citing standalone utility and practical usability. We subsequently supplied the additional 68.75% utility-preserving configuration, its 0.96 TA drop ratio, and the consolidated usability evidence described above. We have not received further feedback on this final supplement.

  • Reviewer pC3E has not provided follow-up feedback on our responses. We supplied additional experiments addressing the principal recovery, comparative-benefit, and held-out-recipient concerns, together with point-by-point responses to the remaining comments.

We believe the expanded recovery tests, direct defense comparisons, additional utility operating point, and held-out-recipient results substantially strengthen the empirical evaluation. They provide a clearer account of the method’s protection–utility–recovery trade-off and its measured processing cost, while making the threat assumptions and reproducibility details explicit. We respectfully ask the Area Chair to consider this additional evidence and the discussion feedback alongside the initial reviews.

Thank you once again for your time, careful consideration, and dedication to the community.

Best regards,

The Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.