acceptodds
Under review as a conference paper at ICLR 2027

FedCURE: Mitigating Safety Erosion in Fully Benign Federated LLM Fine-Tuning through Cross-Client Update Consensus

Abstract

Fine-tuning aligned large language models (LLMs) on benign data can compromise safety even without malicious intent, but this phenomenon has been studied primarily in centralized settings. We show that safety erosion also arises in federated LLM fine-tuning, even when all participating clients are benign and train exclusively on benign local data. Our analysis reveals a federated-specific signal: erosion-aligned update directions that are consistently shared across clients can become disproportionately represented in the aggregated update. Motivated by this observation, we propose FedCURE, a safety-aware federated fine-tuning framework centered on Erosion Consensus Filtering (ECF). ECF identifies erosion-aligned candidates within individual client updates and filters those with strong cross-client consensus before aggregation. FedCURE further employs offline reference adapters to stabilize local adaptation and correct residual safety degradation in the accumulated global model. Experiments across diverse LLM families and downstream datasets demonstrate that FedCURE substantially reduces safety erosion while retaining effective task performance and general utility.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Federated Learning, LLM Safety, Model Safety

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_4uqoOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper FedCURE is a framework for preserving LLM safety during federated fine-tuning when all clients and their training data are benign. The method identifies update directions that align with an erosion reference and are shared across clients, then filters their contribution to the global update. it also uses reference adapters for local stabilization and global safety restoration. The experiments show substantial safety improvements while generally retaining useful downstream performance.

Critical strengths

I find the focus on safety erosion under fully benign participation worthwhile. The paper goes beyond observing that safety deteriorates, it connects shared, erosion-aligned update components to the aggregated model. The energy-matched interventions in Figure 2 provide useful behavioral evidence that these components matter, rather than merely reflecting the magnitude of the removed updates.

I also appreciate the breadth of the evaluation. The results cover several model families and tasks, with additional experiments on multiple seeds, longer training, different client counts and data distributions, and alternative reference data, and the measurements are particularly helpful. All this make the safety gains of the complete framework convincing and support my positive assessment.

Critical weaknesses

My main concerns the is the comparison with the closest approach. Jailbreak to Protect (Ham et al., 2026), which the paper cites, already combines temporary jailbreaking during fine-tuning with a reinforcing adapter merged afterward, yet a federated adaptation is absent from the experiments. Since ERA and RRA account for substantial gains, I would like a direct comparison to better establish what FedCURE adds beyond bringing this approach into federated training.

I am also not fully convinced that the ablations isolate the benefit of consensus. In Appendix G.2, the variant without consensus introduces an additional cosine-similarity threshold of 0.5, changing the candidate pool as well as the selection rule. In Table 8, adding ECF to ERA–RRA on Dolly reduces harmful score from 2.44 to 2.03 but lowers ROUGE-L from 39.77 to 38.62. The Commonsense results are more favorable, so I view this as a need to clarify when consensus helps, rather than evidence that it is ineffective.

Questions for authors

Could the authors compare FedCURE with a faithful federated adaptation of Ham et al.’s Buffer-and-Reinforce framework using matched reference data and comparable training budgets? This would help me assess the contribution relative to the closest prior approach.

Does ECF retain its advantage when the consensus and no-consensus variants use exactly the same ERA-positive candidate pool? I would find multi-seed safety–task trade-offs for these variants helpful in judging the additional benefit of consensus.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 4uqo (1/2)

Q1&W1: Comparison with Prior Work

We appreciate the comment on the comparison with Buffer-and-Reinforce (BnR) (Ham et al. 2026). We summarize our response as follows.

  1. There are algorithmic differences between ERA/RRA and BnR, particularly in how they are trained and used: Clarified in Section 4 and Appendix B.
  2. Directly applying BnR to federated fine-tuning leads to significant degradation in terms of over-refusal and task performance: Updated in Tables 2,3,6 and figure 5.
  3. Erosion Consensus Filtering (ECF), which leverages cross-client consensus (a signal which is available only in FL), provides consistent gains across diverse settings: Section 2 and Appendix G.4.
  4. FedCURE is not limited to ERA/RRA. Its core idea, ECF, can be integrated with existing methods such as SafeLoRA/SaLoRA, confirming its applicability: Reported in Table 5 and Section 5.2.

Below, we provide our detailed response.


0. Perspective Shift: from Local Intervention to Federated Consensus Filtering

Unlike BnR's model-level safety interventions in a centralized setting, FedCURE targets safety erosion caused by shared client updates accumulating across FL rounds. ECF leverages cross-client consensus to filter erosion-aligned updates at the server before they accumulate in the global adapter. Directly applying BnR to FL is not sufficient, as we describe below.

1. Algorithmic Differences from BnR

Like BnR, FedCURE uses reference adapters, but FedCURE introduces ECF, which applies consensus filtering to clients' marginal updates, and also trains and uses its adapters differently. BnR trains ReinforceLoRA with BufferLoRA active; FedCURE trains RRA independently on frozen W0W_0: LRFedCURE(ϕR)=E(x,y)∼DR[−log⁡pW0+ΔW(ϕR)(y∣x)].\mathcal{L}_R^{\mathrm{FedCURE}}(\phi_R)=\mathbb{E}_{(x,y)\sim\mathcal{D}_R}[-\log p_{W_0+\Delta W(\phi_R)}(y\mid x)]. In federated BnR, each client trains with BufferLoRA frozen and active; the equal-weight merge halves their trained LoRA updates ΔWU,i\Delta W_{U,i} before factor averaging. FedCURE instead applies a convex merge by introducing cc at the server to the ECF-filtered global adapter:

\Delta W_i^{\mathrm{BnR}}&=\mathcal{R}_r[\tfrac12(\Delta W_{U,i}+(I-\alpha P_{U,i})W_R^{\mathrm{BnR}})],\\\\ \Delta W_t^{\mathrm{FedCURE}}&=\mathcal{R}_r[(1-c)\Delta W_t^{\mathrm{stab}}+c(I-\alpha_b P_{T,t})W_R]. \end{aligned}$$ --- ## 2. Applying BnR to Federated Learning Both use the same reference data, training budget, and setup (Llama-3-8B-Instruct; deltas relative to FedIT). |Dataset|Method|HS↓|Task↑|Utility↑|ΔOverall↑| |---|---|---|---|---|---| |Dolly|FedIT|48.67|39.83|58.93|--| ||BnR|0.53|31.81|69.80|+50.99| ||**FedCURE**|2.03|38.62|68.21|**+54.71**| |Commonsense|FedIT|25.10|85.39|49.27|--| ||BnR|0.45|69.48|68.82|+28.29| ||**FedCURE**|1.58|84.36|67.54|**+40.76**| BnR not only loses task performance but also over-refuses on benign prompts, whereas **FedCURE keeps both HS and O-R low while preserving task performance, achieving a higher ΔOverall.** |O-R↓|FedIT|BnR|**FedCURE**| |---|---|---|---| |Dolly|2.80|13.60|**4.40**| |Commonsense|10.40|13.20|**5.60**| --- ## 3. ECF's Gain is Consistent Across Diverse Settings Beyond the differences between ERA/RRA and BnR, we further highlight the core novelty of FedCURE: Erosion Consensus Filtering (ECF). Even in extended federated fine-tuning over round 50, adding ECF to ERA/RRA components leads to consistent gains. ECF raises ΔOverall for every component pairing: |Dataset|Component change|ΔHS↓|ΔTask↑|ΔUtility↑|ΔOverall↑| |---|---|---|---|---|---| |Dolly|ERA-only→ERA–ECF|−2.80|+0.61|+2.90|+6.31| ||RRA-only→ECF–RRA|−2.14|+1.30|+2.46|+5.90| ||ERA–RRA→FedCURE|−0.87|−0.56|+2.18|+2.49| |Commonsense|ERA-only→ERA–ECF|−0.03|−0.48|+2.72|+2.27| ||RRA-only→ECF–RRA|−1.54|−0.25|+4.64|+5.93| ||ERA–RRA→FedCURE|−1.80|−0.43|+2.98|+4.35| --- ## 4. ECF is Plug-and-Play: Alone or on Top of Other Defenses Finally, we would like to highlight that ECF can be integrated with other safety tuning methods (e.g., SafeLoRA, SaLoRA) without requiring client-side ERA conditioning or RRA correction. ERA serves only as the server reference. When added to other defenses on Dolly, ECF lowers HS and raises ΔOverall against FedIT in all three cases: |Method|HS↓|Task↑|Utility↑|ΔOverall↑| |---|---|---|---|---| |SafeLoRA|45.67|39.63|60.31|+4.18| |**SafeLoRA+ECF**|11.83|38.17|65.67|**+41.93**| |SaLoRA|34.80|39.04|60.85|+15.00| |**SaLoRA+ECF**|26.87|38.67|62.32|**+24.03**| |Multi-Krum|44.89|39.19|60.29|+4.50| |**Multi-Krum+ECF**|6.83|37.99|67.61|**+48.68**| --- Overall, we believe that (i) the algorithmic differences between FedCURE and BnR, (ii) the novelty and effectiveness of ECF in FL, and (iii) ECF's compatibility with existing methods, collectively demonstrate the advantages of FedCURE. We hope these contributions provide insights into how to preserve LLM safety during federated fine-tuning with fully benign clients, a setting that has received limited attention.
AuthorsOpenReview ↗

Response to Reviewer 4uqo (2/2)

Q2&W2: Isolating the Benefit of Consensus

We thank the reviewer for this question, which lets us clarify the benefit of consensus. We agree that the cosine threshold in our original no-consensus variant changes the candidate pool as well as the selection rule. We therefore repeat the comparison with the pool held fixed.

  • Results are presented in Appendices G.4 and G.6.

1. Consensus vs. No-Consensus on the Same ERA-positive Pool

Both variants start from exactly the same ERA-positive candidate pool and differ only in what they remove: No-consensus removes the entire pool (no additional threshold), while Consensus (ECF) removes only the components shared across clients. The original cosine-thresholded variant (similarity to ERA ≥0.5) is kept for reference. Client-side ERA conditioning and RRA correction are off; ERA remains the server reference on Dolly, Dirichlet α=0.1\alpha=0.1, 32 fully participating clients over three seeds.

VariantPoolConsensusHS↓Task↑Utility↑ΔOverall↑
FedIT––39.73±6.9839.23±0.2462.98±1.52--
CosinethresholdedNo29.30±20.1338.52±0.9364.38±2.88+11.11±18.08
No-consensussameNo2.88±0.0825.32±0.1265.41±0.17+25.36±8.58
Consensus (ECF)sameYes5.68±2.4435.96±1.0770.20±0.95+38.00±6.91

On the same ERA-positive pool, the no-consensus variant suffers a large task score drop, severely degrading its safety–task trade-off. In contrast, with consensus, ECF achieves a competitive HS while largely preserving task performance and also improving utility, yielding the highest ΔOverall. This result is presented in Appendix G.6.


2. When ECF Helps

ECF is designed to help when clients’ marginal updates share erosion-aligned directions, using cross-client consensus to identify which ERA-positive components to remove. Even with ERA or ERA–RRA, we found that these components are not fully eliminated, leaving room for ECF to provide additional gains. As shown below, adding ECF reduces harmful scores and improves ΔOverall across all evaluated component pairings:

DatasetComponent changeΔHS↓ΔTask↑ΔUtility↑ΔOverall↑
DollyERA-only→ERA–ECF−2.80+0.61+2.90+6.31
RRA-only→ECF–RRA−2.14+1.30+2.46+5.90
ERA–RRA→FedCURE−0.87−0.56+2.18+2.49
CommonsenseERA-only→ERA–ECF−0.03−0.48+2.72+2.27
RRA-only→ECF–RRA−1.54−0.25+4.64+5.93
ERA–RRA→FedCURE−1.80−0.43+2.98+4.35

To summarize, ECF complements ERA/RRA by filtering residual ERA-aligned components in each round’s aggregated marginal update. Its additional gains when combined with SafeLoRA, SaLoRA, and Multi-Krum further demonstrate its broader applicability, as shown in our response to Q1&W1. The conditions under which ECF provides additional gains are discussed in Appendix G.4, with supporting component comparisons in Table 14.


We appreciate your time and effort in reviewing our paper. We would be glad to answer any remaining concerns the reviewer might have.

AuthorsOpenReview ↗

Follow-up to Reviewer 4uqo

Dear Reviewer 4uqo,

Thank you again for your comments on our paper. With limited time remaining in the author-reviewer discussion period, we wanted to follow up and ask whether you have had a chance to check our response. We have added new results including (i) comparison with prior work (Ham et al., 2026) and (ii) clear ablations to isolate the benefit of consensus.

If you have any additional comments, we would be happy to address them before the discussion period ends.

Sincerely, Authors of Paper 55503

Reviewer_TZ25OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses safety erosion in federated LoRA fine-tuning of aligned LLMs when all clients are benign, and proposes FedCURE, which has three parts:

  • ERA: a frozen adapter trained in advance to comply with harmful requests. Clients train with it mounted, and it is removed after training.
  • ECF: on the server, each client's update is searched for directions that point the same way as ERA. Among these, directions shared by many clients are removed from the aggregated update.
  • RRA: an adapter trained on refusal data and general instruction data. After its overlap with the task directions is reduced, it is mixed into the global adapter a little each round.
Critical strengths
  1. ECF is a new idea that uses the structure of federated learning, and it works on its own.

ECF alone substantially reduces safety erosion with little loss in task performance. The filtering-policy comparisons are also well designed. Removing random directions, ignoring the degree of cross-client sharing, or ignoring ERA alignment all produce worse trade-offs between safety and task performance. These results support the claim that the key is the combination of being “aligned with ERA” and “shared across multiple clients.”

  1. Broad experiments.

Results are consistent across models, tasks, seeds, levels of data heterogeneity, and numbers of clients, and the code is released.

Critical weaknesses
  1. Much of the performance improvement can already be achieved using components similar to those in prior work.

ERA/RRA are substantially similar to BufferLoRA/ReinforceLoRA from Ham et al. (ICML 2026), both in their training-data construction and in how they are used. However, the paper cites this work only as motivation and does not explain the specific similarities and differences between the two pairs of adapters.

The paper’s own component ablations show two things:

On Dolly, ERA and RRA without ECF already achieve nearly the same results as the full method. On both tasks covered by the component ablations, RRA alone outperforms all evaluated external baselines in terms of the overall score. The paper’s distinct contribution is therefore essentially ECF, but the incremental gain it provides, particularly on Dolly, is very small.

  1. The additional impact of federated learning and the fairness of the comparison protocol are not well established.

(1) There is no comparison with centralized training. This makes it difficult to determine the additional impact of federated learning and repeated aggregation on safety degradation.

(2) The overall score may reflect benefits from additional data. It simply sums the changes in safety, task performance, and general utility. The general instruction data (Alpaca) used to train RRA may contribute to the utility gains, while the baselines do not receive this data. The amount of safety data also differs across methods, and some baselines exhibit collapsed task performance, warranting closer examination of their tuning settings.

Questions for authors
  1. How do ERA/RRA differ from BufferLoRA/ReinforceLoRA which are from Ham et al. (ICML 2026)? Could you state the relationship to that prior work clearly?
  2. Training with ERA mounted may already reduce the erosion-related directions in client updates. In that setting, what does ECF actually remove, and how much?
  3. How does erosion behave under centralized training on pooled data? Could you show the same analysis for shared directions that are unrelated to safety?
  4. Could you report the overall score without the utility term, or give the baselines the same general instruction data? How were the baselines tuned?
  5. Minor detail. the scores of the original model before fine-tuning which is missing from the paper.
4 comments
AuthorsOpenReview ↗

Response to Reviewer TZ25 (1/3)

Thank you for the opportunity to clarify FedCURE's relation to Buffer-and-Reinforce (BnR) (Ham et al., 2026) and the role of Erosion Consensus Filtering (ECF). We summarize our response as follows.

  • (Q1) There are algorithmic differences between ERA/RRA and BnR, particularly in how they are trained and used, and directly applying BnR to federated fine-tuning leads to significant degradation in terms of over-refusal and task performance: Algorithmic differences are in Section 4 and Appendix B, and Comparison with BnR is presented in Section 2, Tables 2,3 and 6 and figure 5.
  • (Q1) FedCURE is not limited to ERA/RRA. Its core idea, ECF, can be integrated with existing methods such as SafeLoRA/SaLoRA, confirming its applicability: Reported in Table 5 and Section 5.2.
  • (Q2) ECF removes the residual erosion-aligned components that ERA conditioning does not eliminate, providing consistent additional gains: Reported in Appendix G.4.
  • (Q3) Safety erosion occurs under both centralized and federated fine-tuning, and removing unrelated to shared safety directions does not reduce HS the most: Centralized training is presented in Appendix G.1 and Shared direction is in Appendix G.7.
  • (Q4) FedCURE remains the best method without the utility term and Alpaca dataset: Presented in Appendix G.8.
  • (Q5) We report the original-model scores: FedCURE improves overall scores while keeping HS near the original level: Presented in Appendix G.1.

Below, We provide our detailed response.


Q1&W1: Comparison with Prior Work

We appreciate the comment. Below, we describe the novelty of our FedCURE compared to BnR.

0. Perspective Shift: from Local Intervention to Federated Consensus Filtering

Unlike BnR's model-level safety interventions in a centralized setting, FedCURE targets safety erosion caused by shared client updates accumulating across FL rounds. ECF leverages cross-client consensus to filter erosion-aligned updates at the server before they accumulate in the global adapter.


1. Algorithmic Differences from BnR

Unlike BnR, FedCURE not only introduces ECF, but also trains and uses its adapters differently. BnR trains ReinforceLoRA with BufferLoRA active; FedCURE trains RRA independently on frozen W0W_0: LRFedCURE(ϕR)=E(x,y)∼DR[−log⁡pW0+ΔW(ϕR)(y∣x)].\mathcal{L}_R^{\mathrm{FedCURE}}(\phi_R)=\mathbb{E}_{(x,y)\sim\mathcal{D}_R}[-\log p_{W_0+\Delta W(\phi_R)}(y\mid x)]. In federated BnR, each client trains with BufferLoRA frozen and active; the equal-weight merge halves their trained LoRA updates ΔWU,i\Delta W_{U,i} before factor averaging. FedCURE instead applies a convex merge by introducing cc at the server to the ECF-filtered global adapter:

\Delta W_i^{\mathrm{BnR}}&=\mathcal{R}_r[\tfrac12(\Delta W_{U,i}+(I-\alpha P_{U,i})W_R^{\mathrm{BnR}})],\\\\ \Delta W_t^{\mathrm{FedCURE}}&=\mathcal{R}_r[(1-c)\Delta W_t^{\mathrm{stab}}+c(I-\alpha_b P_{T,t})W_R]. \end{aligned}$$ Like BufferLoRA, ERA is mounted during local training; it also serves as the server-side erosion reference for ECF. --- ## 2. Federated BnR Lowers HS but Over-refuses and Loses Task Performance Both use the same reference data, training budget, and setup (Llama-3-8B-Instruct; deltas relative to FedIT). |Dataset|Method|HS↓|Task↑|Utility↑|ΔOverall↑| |---|---|---|---|---|---| |Dolly|FedIT|48.67|39.83|58.93|--| ||BnR|0.53|31.81|69.80|+50.99| ||**FedCURE**|2.03|38.62|68.21|**+54.71**| |Commonsense|FedIT|25.10|85.39|49.27|--| ||BnR|0.45|69.48|68.82|+28.29| ||**FedCURE**|1.58|84.36|67.54|**+40.76**| BnR not only loses task performance but also over-refuses on benign prompts, whereas **FedCURE keeps both HS and O-R low while preserving task performance, achieving a higher ΔOverall.** |O-R↓|FedIT|BnR|**FedCURE**| |---|---|---|---| |Dolly|2.80|13.60|**4.40**| |Commonsense|10.40|13.20|**5.60**| --- ## 3. ECF is Plug-and-Play on Top of Other Defenses Finally, our key component, ECF, can be integrated with other safety tuning methods (e.g., SafeLoRA, SaLoRA) without requiring client-side ERA conditioning or RRA correction. ERA serves only as the server reference. When added to other defenses on Dolly, ECF lowers HS and raises ΔOverall against FedIT in all three cases: |Method|HS↓|Task↑|Utility↑|ΔOverall↑| |---|---|---|---|---| |SafeLoRA|45.67|39.63|60.31|+4.18| |**SafeLoRA+ECF**|11.83|38.17|65.67|**+41.93**| |SaLoRA|34.80|39.04|60.85|+15.00| |**SaLoRA+ECF**|26.87|38.67|62.32|**+24.03**| |Multi-Krum|44.89|39.19|60.29|+4.50| |**Multi-Krum+ECF**|6.83|37.99|67.61|**+48.68**| Overall, we believe that (i) the algorithmic differences between FedCURE and BnR, (ii) the novelty and effectiveness of ECF in FL, and (iii) ECF's compatibility with existing methods, collectively demonstrate the advantages of FedCURE. We hope these contributions provide insights into how to preserve LLM safety during federated fine-tuning with fully benign clients, a setting that has received limited attention.
AuthorsOpenReview ↗

Response to Reviewer TZ25 (2/3)

Q2&W1: What ECF Removes under ERA Conditioning

Thank you for the comment. Below, we describe the role and benefit of ECF even with ERA mounted.

  • Results are reported in Appendix G.4.

1. ERA Reduces ERA-Positive Consensus, but a Residual Remains

As the reviewer anticipated, mounting ERA already reduces erosion-aligned shared components, but it does not eliminate them all. We repeat the Section 3 analysis on an ERA-only run:

Consensus modesRunServer energy(%)Client energy(%)Commitment
ERA-positive (Q+)FedIT62.816.60.78
ERA-only15.64.40.74
ERA-negative (Q-)FedIT0.50.20.47
ERA-only53.413.90.79

With ERA mounted, the energy share of ERA-positive consensus modes falls in both client and server marginals. The residual is still committed at nearly the same rate as without ERA: ERA makes this component smaller but about as shared across clients, so aggregation does not cancel it. This residual is what ECF removes under ERA conditioning. Most consensus energy instead lies in ERA-negative modes; these are similarly well shared but negatively aligned with ERA, so ECF, which selects candidates by positive ERA alignment, leaves them untouched.


2. With ERA Mounted, ECF Still Filters the Residual

ECF is designed to help when clients’ marginal updates share erosion-aligned directions, using cross-client consensus to identify which ERA-positive components to remove. Removed energy denotes the per-round fraction of aggregate marginal energy removed by ECF, averaged through the evaluated round (Dolly: round 50; Commonsense: round 10). Even with ERA or ERA–RRA, we found that these components are not fully eliminated, leaving room for ECF to provide additional gains. As shown below, adding ECF reduces harmful scores and improves ΔOverall across all evaluated component pairings:

DatasetComponent changeRemoved energy (%)ΔHS↓ΔTask↑ΔUtility↑ΔOverall↑
DollyERA-only→ERA–ECF7.25−2.80+0.61+2.90+6.31
RRA-only→ECF–RRA39.62−2.14+1.30+2.46+5.90
ERA–RRA→FedCURE17.94−0.87−0.56+2.18+2.49
CommonsenseERA-only→ERA–ECF18.75−0.03−0.48+2.72+2.27
RRA-only→ECF–RRA28.63−1.54−0.25+4.64+5.93
ERA–RRA→FedCURE23.71−1.80−0.43+2.98+4.35

To summarize, even with ERA/RRA, ECF filters the residual ERA-aligned components in each round’s aggregated marginal update. Its additional gains when combined with SafeLoRA, SaLoRA, and Multi-Krum further demonstrate its broader applicability, as shown in our response to Q1&W1.


Q3&W2: Erosion under Centralized Training & Shared Directions Analysis

  • Results for centralized training and shared direction analysis are presented in Appendices G.1 and G.7 respectively.

1. Centralized Training on Pooled Data

As shown below, centralized SFT also exhibits safety erosion on all datasets, and its difference from federated training varies across datasets, under the same training budget, data, and setup (Llama-3-8B-Instruct, α=0.5\alpha=0.5). Thus erosion is not unique to FL, and safety erosion in centralized settings has been extensively studied.

DatasetMethodHS↓O-R↓Task↑Utility↑
DollyCentralized SFT74.621.2038.5758.94
FedIT48.672.8039.8358.93
CommonsenseCentralized SFT15.8711.2085.0837.00
FedIT25.1010.4085.3949.27
WizardLMCentralized SFT6.4320.806.68864.57
FedIT3.6819.606.87965.23

The unique/key part of FL is to mitigate this erosion while training data remain distributed across clients. Simply removing all ERA-aligned update components can degrade task performance, motivating a more selective approach. ECF uses cross-client consensus to select which ERA-positive components to remove, targeting those consistently shared across clients.


2. Shared Directions Unrelated to Safety

As shared directions exist only in the context of FL (where datasets are distributed across clients), we filter shared directions negatively aligned with ERA (ERA-Negative) or positively aligned with a gaussian-noise reference unrelated to safety (Gaussian-Positive) throughout training with ERA & RRA off on Dolly. The Gaussian reference provides a control constructed without safety information.

Filtered ModesHS↓Task↑Utility↑ΔOverall↑
FedIT37.9639.4464.05--
ERA-Negative37.8238.9063.52−0.93
Gaussian-Positive25.7038.6366.02+13.42
ECF3.2634.9971.30+37.50

Both controls yield shared directions, but removing them does not restore safety. ECF achieves a substantially larger reduction, suggesting that the choice of shared directions matters for safety mitigation.

AuthorsOpenReview ↗

Response to Reviewer TZ25 (3/3)

Q4&W2: Overall Score without Utility and Baseline Settings

We appreciate the comment. We organize our response into three parts.

  • Results are reported in Appendix G.8.

1. Overall Score without the Utility Term

As the reviewer suggested, we drop the utility term: ΔOverall(−U)=−ΔHS+ΔTask\Delta_{\mathrm{Overall}}^{(-U)}=-\Delta_{\mathrm{HS}}+\Delta_{\mathrm{Task}}. FedCURE still exceeds the best baseline across all datasets on Llama:

DatasetFedCUREBest baselineGap
Dolly45.43BnR 40.12+5.31
Commonsense22.49SafeLoRA 12.91+9.58
WizardLM3.482SaLoRA 1.181+2.301

2. Without Alpaca in RRA Training

To test whether the gains depend on Alpaca, we retrain RRA using only 2.5K harmful prompt–refusal pairs and rerun FedCURE on Commonsense. Without Alpaca, FedCURE still outperforms the compared baselines under both ΔOverall and ΔOverall(−U).

MethodHS↓Task↑Utility↑ΔOverall(−U)↑ΔOverall↑
Booster18.9084.6552.33+5.46+8.52
BnR0.4569.4868.82+8.74+28.29
SaLoRA13.8086.0145.45+11.92+8.10
SafeLoRA11.7484.9449.59+12.91+13.23
FedCURE1.5884.3667.54+22.49+40.76
FedCURE w/o Alpaca2.3684.0565.72+21.40+37.85

FedCURE's gains persist without Alpaca: without Alpaca, FedCURE still exceeds on both scores and improves utility over FedIT on Commonsense.


3. Baseline Data and Tuning

Safety data are drawn from the same sources wherever a method requires: Circuit Breakers, as for ERA and RRA, plus BeaverTails where a second source is needed. The amount of data follows each method's original setting. Common FL training hyperparameters are identical across methods.

MethodSourceCount
SaLoRACircuit Breakers refusal pairs + pooled client task samples256+256(sampling budget)
GuardSpaceBeaverTails + Circuit Breakers harmful prompts, with no overlap520+520
BoosterCircuit Breakers refusal pairs + BeaverTails harmful-completion pairsUp to 5K each
FedLLM-AttackReleased Mistral-generated defense data: harmful refusal pairs + ordinary harmless compliant pairs500+500

We use the common FL protocol and method-specific settings based on the reference implementations (Appendices A–B). Especially for GuardSpace on Llama-3-8B, the null-space tolerance was adjusted to permit effective updates and then held fixed across downstream datasets.


Q5: Original Models Before Fine-Tuning

We report the scores of the original model before fine-tuning: HS, and Utility are Dolly/Commonsense/WizardLM.

  • Results are presented in Appendix G.1.
ModelMethodHS↓Dolly↑Commonsense↑WizardLM↑Utility↑
Llama-3-8BOriginal3.1224.6433.607.50664.96
FedCURE2.03/1.58/0.3038.6284.366.98168.21/67.54/67.89
Qwen2.5-7BOriginal4.3221.3878.357.79475.28
FedCURE3.23/2.86/1.0533.3684.337.68875.94/75.33/76.32
Gemma-2-9BOriginal1.1623.4673.257.95671.59
FedCURE1.84/1.16/0.3039.6287.807.42575.20/74.52/76.54

Overall, we believe that (i) the centralized comparison and the shared-direction controls and (ii) FedCURE's gains without the utility term and Alpaca clarify the impact of FedCURE.


Thank you again for your time and effort in reviewing our paper. We would be happy to address any remaining questions or concerns the reviewer might have.

AuthorsOpenReview ↗

Follow-up to Reviewer TZ25

Dear Reviewer TZ25,

Thank you again for your comments on our paper. With limited time remaining in the author-reviewer discussion period, we wanted to follow up and ask whether you have had a chance to check our response. We have added new results including (i) relationship and comparison with prior work (Ham et al., 2026), (ii) what ECF actually removes, (iii) comparison with centralized training and (iv) overall score without the utility term.

If you have any additional comments, we would be happy to address them before the discussion period ends.

Sincerely, Authors of Paper 55503

Reviewer_6i8FOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies safety degradation during federated LoRA fine-tuning when all clients perform benign task adaptation. It observes that update directions aligned with an independently trained erosion reference and shared across clients can account for a disproportionate fraction of the aggregated update. Based on this observation, FedCURE combines three components: local training conditioned on a frozen erosion reference adapter, Erosion Consensus Filtering (ECF) of shared erosion-aligned marginal directions, and server-side correction using a safety-reinforcement adapter. The server operations are implemented through low-rank factorizations. Experiments cover several model families and downstream tasks, extended training, different client counts and data heterogeneity, reference-data substitutions, and over-refusal. The complete framework substantially reduces measured harmful-response rates, although task-performance trade-offs remain in some settings.

Critical strengths
  1. FedCURE substantially reduces harmful-response rates across multiple model families and tasks, with consistent results across three random seeds. Reference-data and over-refusal evaluations further support the robustness of these improvements.

  2. The geometric analysis is supported by behavioral interventions: removing shared erosion-aligned directions improves safety more than energy-matched control directions. The low-rank implementation keeps the additional computational cost manageable.

Critical weaknesses
  1. The paper lacks a direct comparison with the closely related Buffer-and-Reinforce framework despite similarities in reference-adapter conditioning and safety restoration. On Dolly, ERA–RRA without ECF already achieves 2.44 HS versus 2.03 for the full method, while retaining higher ROUGE-L (39.77 versus 38.62), leaving ECF’s incremental value unclear.

  2. The no-consensus ablation introduces an additional cosine-similarity threshold of 0.5 that ECF does not use, changing both candidate selection and consensus filtering. Table 11 therefore does not cleanly isolate the contribution of cross-client consensus.

Questions for authors
  • Closest-method comparison: How does FedCURE compare with a federated adaptation of Ham et al.’s Buffer-and-Reinforce framework under matched reference data and training budgets? Please also report uncertainty for the incremental ECF gains over ERA–RRA.

  • Controlled consensus ablation: Does ECF retain its advantage when the no-consensus variant uses exactly the same positively aligned candidate spaces, without the additional cosine threshold? Please control or report removed energy and rank for the compared policies.

  • Scope of the federated mechanism: Do the commitment pattern and ECF benefit persist under effective-weight averaging followed by rank truncation, rather than factor-wise averaging? A comparison with centralized training at matched data exposure would also clarify which part of the observed erosion is specifically attributable to federation.

  • Reference-model controls: What are the safety, downstream-task, and general-utility scores of the original aligned model and the model with only the reinforcement reference adapter? These controls would help distinguish preservation of learned task capability from restoration toward the reference model.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 6i8F (1/4)

Q1&W1: Closest-Method Comparison (1/2)

We appreciate the thoughtful comment on the comparison with Buffer-and-Reinforce (BnR) (Ham et al., 2026). We summarize our first response as follows.

    1. There are algorithmic differences between ERA/RRA and BnR, particularly in how they are trained and used: Clarified in Section 4 and Appendix B.
    1. Directly applying BnR to federated fine-tuning leads to significant degradation in terms of over-refusal and task performance: Updated in Tables 2,3 and 6 and figure 5.
    1. Erosion Consensus Filtering (ECF), which leverages cross-client consensus (a signal which is available only in FL), provides consistent gains across diverse settings: Reported in Appendix G.4.
    1. FedCURE is not limited to ERA/RRA. Its core idea, ECF, can be integrated with existing methods such as SafeLoRA/SaLoRA, confirming its applicability: Reported in Table 5 and Section 5.2.

We provide our detailed responses below.


0. Perspective Shift: from Local Intervention to Federated Consensus Filtering

Unlike BnR's model-level safety interventions in a centralized setting, FedCURE targets safety erosion caused by shared client updates accumulating across FL rounds. ECF leverages cross-client consensus to filter erosion-aligned updates at the server before they accumulate in the global adapter.


1. Algorithmic Differences from BnR

Unlike BnR, FedCURE not only introduces ECF, but also trains and uses its adapters differently. BnR trains ReinforceLoRA with BufferLoRA active; FedCURE trains RRA independently on frozen W0W_0: LRFedCURE(ϕR)=E(x,y)∼DR[−log⁡pW0+ΔW(ϕR)(y∣x)].\mathcal{L}_R^{\mathrm{FedCURE}}(\phi_R)=\mathbb{E}_{(x,y)\sim\mathcal{D}_R}[-\log p_{W_0+\Delta W(\phi_R)}(y\mid x)]. In federated BnR, each client trains with BufferLoRA frozen and active; the equal-weight merge halves their trained LoRA updates ΔWU,i\Delta W_{U,i} before factor averaging. FedCURE instead applies a convex merge by introducing cc at the server to the ECF-filtered global adapter:

\Delta W_i^{\mathrm{BnR}}&=\mathcal{R}_r[\tfrac12(\Delta W_{U,i}+(I-\alpha P_{U,i})W_R^{\mathrm{BnR}})],\\\\ \Delta W_t^{\mathrm{FedCURE}}&=\mathcal{R}_r[(1-c)\Delta W_t^{\mathrm{stab}}+c(I-\alpha_b P_{T,t})W_R]. \end{aligned}$$ Like BufferLoRA, ERA is mounted during local training; it also serves as the server-side erosion reference for ECF. --- ## 2. Federated BnR Lowers HS but Over-refuses and Loses Task Performance Both use the same reference data, training budget, and setup (Llama-3-8B-Instruct; deltas relative to FedIT). |Dataset|Method|HS↓|Task↑|Utility↑|ΔOverall↑| |---|---|---|---|---|---| |Dolly|FedIT|48.67|39.83|58.93|--| ||BnR|0.53|31.81|69.80|+50.99| ||**FedCURE**|2.03|38.62|68.21|**+54.71**| |Commonsense|FedIT|25.10|85.39|49.27|--| ||BnR|0.45|69.48|68.82|+28.29| ||**FedCURE**|1.58|84.36|67.54|**+40.76**| BnR not only loses task performance but also over-refuses on benign prompts, whereas **FedCURE keeps both HS and O-R low while preserving task performance, achieving a higher ΔOverall.** |O-R↓|FedIT|BnR|**FedCURE**| |---|---|---|---| |Dolly|2.80|13.60|**4.40**| |Commonsense|10.40|13.20|**5.60**|
AuthorsOpenReview ↗

Response to Reviewer 6i8F (2/4)

Q1&W1: Closest-Method Comparison (2/2)

3. Consistent Gain of ECF Across Diverse Settings

We thank the reviewer for raising this issue. To address this concern, we assess the uncertainty of ECF's consistent gain by examining whether it holds across component pairings, datasets, training lengths, and model families.

  • We present the results in Appendix G.4.

ECF Still Filters the Residual Even With ERA Mounted

ECF is designed to help when clients’ marginal updates share erosion-aligned directions, using cross-client consensus to identify which ERA-positive components to remove. Removed energy denotes the per-round fraction of aggregate marginal energy removed by ECF, averaged through the evaluated round (Dolly: round 50; Commonsense: round 10). Even with ERA or ERA–RRA, we found that these components are not fully eliminated, leaving room for ECF to provide additional gains. Adding ECF reduces harmful scores and improves ΔOverall across all evaluated component pairings:

DatasetComponent changeRemoved energy (%)ΔHS↓ΔTask↑ΔUtility↑ΔOverall↑
DollyERA-only→ERA–ECF7.25−2.80+0.61+2.90+6.31
RRA-only→ECF–RRA39.62−2.14+1.30+2.46+5.90
ERA–RRA→FedCURE17.94−0.87−0.56+2.18+2.49
CommonsenseERA-only→ERA–ECF18.75−0.03−0.48+2.72+2.27
RRA-only→ECF–RRA28.63−1.54−0.25+4.64+5.93
ERA–RRA→FedCURE23.71−1.80−0.43+2.98+4.35

In Other Model Families, ECF Reduces the Over-Refusal of ERA–RRA

On Qwen2.5-7B-Instruct and Gemma-2-9B-it, ERA–RRA reaches a low HS but over-refuses on benign prompts (O-R 18.40/12.40 and 22.80/19.20 on Dolly/Commonsense, vs. 4.00 and 16.00 for the original models; Appendix G.1). Adding ECF lowers O-R by 4.8–12.0 points in all four settings, at a small cost in HS. As over-refusal matters here, ΔOverall in this table also includes it (−ΔHS−ΔO-R+ΔTask+ΔUtility; round 10):

ModelDatasetComponent changeΔHS↓ΔO-R↓ΔTask↑ΔUtility↑ΔOverall(OR)↑
Qwen2.5-7BDollyERA–RRA→FedCURE+1.99−12.00+2.78+0.12+12.91
CommonsenseERA–RRA→FedCURE+1.47−6.00−0.18+0.49+4.84
Gemma-2-9BDollyERA–RRA→FedCURE+1.35−7.60+0.53−0.65+6.13
CommonsenseERA–RRA→FedCURE+0.75−4.80−0.31+8.37+12.11

Overall, the HS gain of ECF may be small in some settings, but a consistent gain is observed across diverse settings. The gain of ECF is not confined to a single setting but appears as lower HS where residual erosion remains and as reduced over-refusal where the reference adapters over-correct.


4. ECF is Plug-and-Play on Top of Other Defenses

  • Results are reported in Table 5 and Section 5.2.

Finally, we would like to highlight that our key component, ECF, can be integrated with other safety methods (e.g., SafeLoRA, SaLoRA) without requiring client-side ERA conditioning or RRA correction; ERA serves only as the server reference. When added to other defenses on Dolly, ECF lowers HS and raises ΔOverall against FedIT in all three cases:

MethodHS↓Task↑Utility↑ΔOverall↑
SafeLoRA45.6739.6360.31+4.18
SafeLoRA+ECF11.8338.1765.67+41.93
SaLoRA34.8039.0460.85+15.00
SaLoRA+ECF26.8738.6762.32+24.03
Multi-Krum44.8939.1960.29+4.50
Multi-Krum+ECF6.8337.9967.61+48.68

Overall, we believe that (i) the algorithmic differences between FedCURE and BnR, (ii) the novelty and effectiveness of ECF in FL even with ERA/RRA, and (iii) ECF's compatibility with existing methods, collectively demonstrate the advantages of FedCURE. We hope these contributions provide insights into how to preserve LLM safety during federated fine-tuning with fully benign clients, a setting that has received limited attention.

AuthorsOpenReview ↗

Response to Reviewer 6i8F (3/4)

We summarize our responses to the remaining questions below.

  • (Q2) On the same positive aligned candidate space, consensus gives a better safety-task trade-off than no-consensus variant and preserves more task performance than the energy-matched control: Presented in Appendix G.6.
  • (Q3) The commitment pattern and benefit of ECF persist under effective weight averaging. While safety erosion also occurs in centralized training, ECF exploits cross-client consensus available in the federated setting to guide selective filtering: Centralized training is presented in Appendix G.1 and Effective-weight is in Appendix G.2.

Q2&W2: Controlled Consensus Ablation

  • Results are presented in Appendix G.6.

We agree that the cosine threshold in our original no-consensus variant changes candidate selection as well as consensus filtering. We fix the pool and report removed energy: ERA-Full removes the entire pool, and ERA-Uniform attenuates the union to match ECF's removed energy per round. ERA & RRA are off; ERA remains the server reference on Dolly, Dirichlet α=0.1\alpha=0.1, 32 fully participating clients.

PolicyConsensusRemoved Energy(%)HS↓Task↑UtilityΔOverall↑
FedIT----37.9639.4464.05--
Cosine ≥0.5No18.917.2437.4467.31+31.98
ERA-UniformNo54.242.5928.3569.19+29.43
ERA-FullNo92.552.9725.2265.45+22.17
ECFYes54.243.2634.9971.30+37.50

On the same positive aligned candidate space, no-consensus variant incurs a clear loss in task performance, degrading its safety-task trade-off, including energy-matched removal. In contrast, with consensus, ECF achieves a competitive HS while largely preserving task performance, yielding the highest ΔOverall. Within the ERA-positive space, consensus provides a criterion for selecting which components to filter while largely preserving task performance.


Q3: Scope of the Federated Mechanism

  • Centralized training and effective weight results are in Appendices G.1 and G.2 respectively.

1. Comparison with Centralized Training at Matched Data

As shown below, centralized SFT also exhibits safety erosion on all three datasets, and its difference from federated training varies across datasets, under the same training budget, data, and setup (Llama-3-8B-Instruct, α=0.5\alpha=0.5). Thus erosion is not unique to FL, and safety erosion in centralized settings has been extensively studied.

The unique/key part of FL is to mitigate this erosion while training data remain distributed across clients. Simply removing all ERA-aligned update components can degrade task performance, motivating a more selective approach. ECF uses cross-client consensus to select which ERA-positive components to remove, targeting those consistently shared across clients.

DatasetMethodHS↓O-R↓Task↑Utility↑
DollyCentralized SFT74.621.2038.5758.94
FedIT48.672.8039.8358.93
CommonsenseCentralized SFT15.8711.2085.0837.00
FedIT25.1010.4085.3949.27
WizardLMCentralized SFT6.4320.806.68864.57
FedIT3.6819.606.87965.23

2. Commitment Pattern and ECF Benefit Persist under Effective-weight Averaging

For effective-weight averaging, the server averages clients' effective updates ΔWi=(αLoRA/r)BiAi\Delta W_i=(\alpha_{\mathrm{LoRA}}/r)B_iA_i weighted by their training-sample counts, then rebuilds the global LoRA via rank-rr truncated SVD. The commitment pattern is not specific to factor-wise averaging: under effective-weight averaging, shared ERA-positive modes (Q+) still take a larger share of server-marginal than client-marginal energy (36.2% vs. 22.2%), with almost no gap for ERA-negative modes (Q-).

AggregationCand.Shared modes(%)Server energy(%)Client energy(%)Commit.
FactorQ+1.5262.816.60.78
Q-0.230.50.20.47
EffectiveQ+8.3736.222.20.43
Q-7.3010.910.50.27

To examine whether ECF remains beneficial under effective-weight averaging, we train all methods using this aggregation scheme. Together with the commitment analysis on Dolly, the results suggest that the observed commitment pattern and the benefits of ECF are not specific to factor-wise aggregation on both datasets:

DatasetMethodHS↓Task↑Utility↑ΔOverall↑
DollyFedIT38.7839.5160.54--
SafeLoRA35.6339.8861.62+4.60
SaLoRA34.0138.8962.18+5.79
ECF17.6639.3967.74+28.20
CommonsenseFedIT23.6985.4345.10--
SafeLoRA10.0684.4656.84+24.40
SaLoRA26.2985.1945.82−2.12
ECF7.7085.2557.86+28.57

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response to AC and All Reviewers

We thank the reviewers for their constructive feedback.


Core Idea

FedCURE targets safety erosion caused by shared client updates accumulating across FL rounds, even when all clients are benign. Erosion Consensus Filtering (ECF) leverages cross-client consensus, a signal available only in FL, to filter erosion-aligned updates at the server before they accumulate in the global adapter.


Our Contributions

Erosion Consensus Filtering (ECF).

A server-side filter that identifies ERA-aligned candidates in each client's marginal update and removes only those with strong cross-client consensus. It is effective alone and on top of existing methods (Table 5).

Complementary Reference Adapters.

ERA conditions local training and serves as ECF's erosion reference; RRA corrects the accumulated global adapter. Together, FedCURE keeps both the harmful score (HS) and over-refusal low while preserving task performance.


We revised the manuscript accordingly; purple marks revised and added text. The common questions are summarized below.

1. Comparison with Prior Work (6i8F, TZ25, 4uqo)

  • Answer. Beyond introducing ECF, FedCURE trains RRA independently on the frozen base model and merges it at the server with a convex scale. Buffer-and-Reinforce (Ham et al., 2026) under matched data and budgets reaches a low HS (0.53/0.45 on Dolly/Commonsense) but over-refuses (13.60/13.20 vs. 4.40/5.60) and loses task performance (31.81/69.48 vs. 38.62/84.36).
  • Revision. Sections 2 and 4, Appendix B; Tables 2, 3, and 6 and Figure 5.

2. What ECF adds beyond ERA/RRA (6i8F, TZ25, 4uqo)

  • Answer. Even with ERA mounted, ECF still removes 7.25–23.71% of the marginal energy, after which a consistent gain is observed across component pairings and datasets. On Qwen2.5-7B-Instruct and Gemma-2-9B-it, it lowers the over-refusal of ERA–RRA by 4.8–12.0 points. ECF is also compatible with existing methods: combined with other methods it lowers HS and raises overall scores.
  • Revision. Appendix G.4; Table 5 and Section 5.2.

3. Isolating the Benefit of Consensus (6i8F, 4uqo)

  • Answer. On exactly the same ERA-positive candidate pool, without the cosine threshold, removal without consensus sharply lowers task performance (25.22 R-L for full and 28.35 for energy-matched removal vs. 34.99 for ECF) at comparable HS.
  • Revision. Appendix G.6.

4. Federated vs. Centralized Training (6i8F, TZ25)

  • Answer. Centralized SFT on pooled data also erodes safety (HS 74.62/15.87/6.43 on Dolly/Commonsense/WizardLM vs. 48.67/25.10/3.68 for FedIT). Erosion is not unique to FL; the cross-client consensus that ECF uses is.
  • Revision. Appendix G.1.

5. Original and Reference-only Models (6i8F, TZ25)

  • Answer. FedCURE keeps HS near the original model (2.03/1.58 vs. 3.12 on Llama-3-8B-Instruct) while retaining learned task capability (38.62 R-L and 84.36 accuracy vs. 36.12 and 48.39 for RRA alone).
  • Revision. Appendices G.1 and G.3.

Reviewer-Specific Additions. Effective-weight averaging (6i8F): Appendix G.2. Shared directions unrelated to safety (TZ25): Appendix G.7. Overall score without utility, RRA without Alpaca, and baseline settings (TZ25): Appendix G.8.


For other details, please refer to our responses corresponding to each reviewer. In case there are remaining questions/concerns, it would be grateful if we can have an opportunity to further answer them.

Best, Authors

Dear Reviewers,

With less than 24 hours remaining in the author-reviewer discussion period, we would appreciate any feedback on our responses and the additional experiments.

If you have any remaining questions or concerns, we would be happy to address them before the discussion period ends. Thank you again for your comments on our paper.

Sincerely, Authors of Paper 55503

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.