acceptodds
Under review as a conference paper at ICLR 2027

Router-Free Modular Storage for Knowledge Unlearning in Large Language Models

Abstract

Machine unlearning aims to remove the influence of selected training data sources, such as individual users or domains, without retraining a model from scratch. This is difficult because information from many sources is entangled in shared model parameters, so that removing one source harms utility on others. Unlearning by construction sets up training to avoid entanglement by confining information on individual data sources to separate modules of a trained model. However, this causes deployment overhead because it requires a router and source labels for every data point at inference time to determine which module is responsible. To overcome this limitation, we propose Modular Unlearning via Self-Routing (MUSR), which places one small module per source in parallel with the MLP layers and passes every input through all of them. During training, each module learns to contribute on its own source and to stay near zero on every other. Thereby, the modules gate themselves and inference is a sum of their outputs with no router and source labels. Our empirical evaluation across four different language models and across pretraining and fine-tuning setups highlights that MUSR matches the removal performance of the modular baselines while retaining 99.7% of its pre-unlearning model utility on average. Beyond single deletions, MUSR supports repeated deletion requests and the addition of new sources after training. Overall, MUSR makes unlearning by construction more practical for standard model inference.

Primary area:
foundation or frontier models, including LLMs
Keywords:
unlearn, LLM

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_xa59OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MUSR, a router-free method for knowledge unlearning. The method add small modules per source parallel to MLP layers and modifies the MLP layers output to be a sum of the original MLP and of all introduced modules. During training a loss term pushes each module to activate only on its dedicated source and to stay zero on other sources. To unlearn a source the dedicated module’s W0 matrix is set to zero and all other modules stay intact. The method therefore requires no router.

Critical strengths

MUSR is a router-free method that wraps around the base model, so the initial model weights remain untouched for fine-tuning setups The approach is tested across different model families and training regimes, pre-training and fine-tuning, and is compared against a wide selection of baselines Almost all of the post-MUSR model’s utility is maintained after unlearning a source (Figures 3, 4) MUSR preserves model’s utility better than GRAM and APA on TOFU (Table 2)

Critical weaknesses
  1. If a model doesn't produce the facts in generation, it's not clear if the knowledge is still stored in the model, but just has a smaller probability of being decoded. The experiment in Appendix, Figure 8, shows that probes on the retained modules’ outputs reach 0.911 AUC, versus 0.562 for the model trained without forget10. The relearning experiment, Figure 9, suggests the facts are not recoverable, but a white-box attacker could still determine which authors were unlearned.

  2. Generalization to broader domains, for instance, unlearning the entire organic chemistry to a separate module, remains untested on large models (the paper covers a pre-training stage on a 200M model with a small K=4)

  3. The storage overhead / parameter count and FLOPs scale linearly with the amount of sources. Efficiency claims (Table 16) rely on batch_size=1 tests and show that at K=200 MUSR's extra FLOPs is 12.73% vs 0.06% of GRAM

  4. In the pretraining experiments, MUSR’s cross-entropy on the unlearned source after unlearning overshoots the retraining reference (3.002 vs 2.709), while GRAM (2.700) matches it more closely, which with retraining set as a standard indicates over-unlearning. The paper shows that most of the gap comes from LISP (Table 7), so the overshoot is concentrated in the source most different from most of the data. GRAM shows a similar pattern with smaller absolute changes in cross-entropy

  5. Table 12 shows that on LUME with Llama-1B, MMLU drops from frozen base’s 0.455 to 0.376 with modules attached, and rises to 0.427 when the forget modules are removed. It seems that self-routing doesn’t fully suppress modules on OOD inputs, and deletion changes the model’s behavior on non-source prompts

Because of weaknesses 4 and 5 I am leaning to rejection.

Update: the authors have provided experiments resolving weaknesses 4 and 5

Questions for authors
  1. How do you check retention of facts that are partially overlapping with unlearned knowledge? If a fact is duplicated for two different authors, would unlearning one of the authors do anything with the other?

  2. Could you provide more detailed analysis on whether modules fire on non-source queries, for instance MMLU questions?

  3. Could you report the per-source retraining cross-entropy for Table 1, to verify that the overshoot relative to retraining comes mainly from LISP?

  4. Extension to safety: RMU is used as a baseline, but it would be curious to know how MUSR compares against RMU on WMDP benchmark

4 comments

We thank the reviewer for the detailed feedback. We appreciate the recognition of MUSR's retained utility and the evaluation across models and training settings.

W1: Residual traces after unlearning

The white-box membership probe detects a residual trace after deletion, but requires access to internal module activations and labeled examples for training the probe. Our black-box membership inference evaluations consider a common deployment setting with output-only access and show no significant evidence of leakage after unlearning (Figure 3(d) and the per-model results). This difference in results arises because the white-box probe examines the effect of negative samples on the retained modules. During training, these modules learn to remain quiet on forget-source examples, and that pattern persists after the corresponding source modules are removed. To test whether this trace helps recover the deleted answers, we additionally ran two relearning attacks that fine-tune the model on forget data after unlearning. Neither attack recovers the answers faster than the corresponding control trained without forget10.

W2: Generalization to broader domains and larger models

We agree that domain breadth and model scale are evaluated separately in our experiments. The 200M-parameter pretraining experiment assigns LISP, biology, nuclear science, and cybersecurity to separate removable modules, with an average cross-entropy change of only 0.0020.002on retained data after deletion. Our larger-model experiments evaluate source-level removal on Llama-3.1-8B and Qwen2.5-7B with 200200 modules on TOFU. These experiments together demonstrate domain-level removal during pretraining and source-level removal at larger model sizes.

W3: Inference efficiency beyond batch size one

To address the reviewer’s comment that our efficiency measurements used only batch size 1, we ran additional experiments with batch sizes up to 256 on Llama-1B and Llama-8B. At K=200K=200, MUSR has lower generation latency than GRAM at every tested batch size from 11 to 256256 on both Llama backbones. At batch size 256256, MUSR takes 6.4256.425 seconds versus 6.6776.677 for GRAM on Llama-1B, and 23.64723.647 versus 24.61624.616 on Llama-8B. Dense evaluation avoids the module-selection operations of routed evaluation. We have added the measurements in Appendix F, Table 26.

W4/Q3: Pretraining overshoot on LISP and per-source comparisons to retraining

We addressed the pretraining overshoot by training the backbone alone on a fraction of the shared core batches. Previously, the backbone was always trained with the source modules active. We now switch off all modules on 8080% of shared core batches, reducing the backbone's dependence on their outputs. Source batches retain the same module-specific training rule.

With this change, MUSR reaches an average unlearned-source cross-entropy of 2.764±0.0102.764 \pm 0.010, compared with 2.711±0.0022.711 \pm 0.002 for retraining (Table 1). The per-source comparison is reported in Table 8 in Appendix C.2 and summarized below.

SourceRetrainingMUSR after unlearning
LISP1.605±0.0051.605 \pm 0.0051.684±0.0331.684 \pm 0.033
Biology2.904±0.0032.904 \pm 0.0032.949±0.0042.949 \pm 0.004
Nuclear3.220±0.0033.220 \pm 0.0033.266±0.0053.266 \pm 0.005
Cyber3.114±0.0013.114 \pm 0.0013.157±0.0073.157 \pm 0.007

The gap to retraining is now 0.0790.079 for LISP. Source learning remains strong, with pre-unlearning cross-entropy of 2.4682.468, compared with 2.5192.519 for GRAM. The revised training therefore brings unlearning closer to retraining while preserving source learning. We have updated Table 1 and the per-source heatmap in Table 7 with these results. The training change and its ablation are described in Appendix C.2, Table 8 of the revised manuscript.

W5/Q2: Module activity on MMLU

The MMLU decrease also happens under standard adaptation to LUME. On Llama-3.2-1B-Instruct, joint fine-tuning reaches 0.348±0.0100.348 \pm 0.010, retraining reaches 0.330±0.0080.330 \pm 0.008, and GRAM reaches 0.395±0.0120.395 \pm 0.012, compared with 0.4550.455 for the frozen base. To address this for MUSR, we extended the output penalty with general-text negatives from Alpaca [1], reducing module output norms on MMLU from 18.8% to 5.3% of their own-source norms.

TrainingMMLU beforeMMLU afterRetain afterForget after
Frozen backbone0.4550.455---
Fine-tuning / retraining0.348±0.0100.348 \pm 0.0100.330±0.0080.330 \pm 0.0080.983±0.0010.983 \pm 0.0010.130±0.0010.130 \pm 0.001
Original MUSR0.369±0.0120.369 \pm 0.0120.428±0.0030.428 \pm 0.0030.983±0.0000.983 \pm 0.0000.130±0.0010.130 \pm 0.001
MUSR with Alpaca negatives0.447±0.0020.447 \pm 0.0020.448±0.0010.448 \pm 0.0010.983±0.0000.983 \pm 0.0000.130±0.0030.130 \pm 0.003

This brings MMLU close to the frozen base while preserving Retain and Forget performance, without additional inference computation. We have added the analysis and results in Appendix D.3, Table 18.

Q1: Unlearning facts shared across sources

We evaluated both naturally overlapping information and explicitly shared facts on TOFU. In the original data, we identified 555555 retained questions whose answers share an entity with forget10, such as a country mentioned for both a retained and an unlearned author. After unlearning, their recall remains at 0.9890.989, decreasing by only 0.0050.005.

We also constructed 2020 cases where retained and unlearned authors share the same attribute value. For example, if authors A and B both speak Spanish, deleting A should preserve “B speaks Spanish” while unlearning “A speaks Spanish”. After unlearning, recall on these facts is 0.9730.973 for retained authors and 0.3740.374 for unlearned authors, compared with 1.0001.000 and 0.3490.349 for retraining. These results support MUSR’s ability to unlearn author-specific associations while preserving shared information about retained authors. We have added the results in Appendix E.9, Table 24.

Q4: Comparison with RMU on WMDP

To address the reviewer’s request for a comparison with RMU on hazardous knowledge, we ran additional experiments in our realistic pretraining setting. WMDP tests knowledge already present in pretrained models, whereas our fine-tuning setup removes information learned through added modules around a frozen backbone. Our pretraining setting instead introduces the domain data during modular training. We use it as a controlled proxy for WMDP because it includes cybersecurity, biology, and nuclear-science sources relevant to hazardous knowledge, while providing a retrained reference for each removed source.

We therefore compare MUSR with RMU, GradDiff, NPO, and SimNPO. RMU preserves retained performance but raises cross-entropy on the unlearned sources to 5.1535.153, compared with 2.7112.711 for retraining. NPO reaches 2.7312.731 after 10001000 optimization steps per deletion request. MUSR reaches 2.7642.764 through module deletion alone, with an average cross-entropy change of only 0.0020.002 on retained data. We have added these comparisons in Appendix C.3, Table 9.

We hope that the revised pretraining results and the MMLU analysis address the two concerns that led to the rating, and we would appreciate it if the reviewer would consider raising the score in that case. Please let us know if further clarification would be helpful.

References: [1] Stanford Alpaca: An Instruction-following LLaMA model, Taori et al., 2023

I thank the authors for additional experiements. I believe my main concerns have been addressed (MUSR no longer overshoots retraining by a large margin and it's MMLU performance is closer to frozen). I will raise my score

We thank the reviewer for the constructive feedback and for raising the score. Your comments helped us improve the pretraining removal results and the preservation of MMLU performance. We are glad these changes address your main concerns and are happy to answer any further questions.

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

This paper studies machine unlearning by construction through Modular Unlearning via Self-Routing (MUSR). It assigns source-specific modules to selected MLP layers and evaluates all modules on every input. During training, each sample's task loss updates its assigned module, while an output penalty suppresses the contributions of non-assigned modules. Source labels are used during training but are not required for ordinary inference. Unlearning a specified source zeros its module's output projections. Experiments cover pretraining and fine-tuning settings. On TOFU, MUSR retains approximately 99.7% of its pre-unlearning utility on average across four backbones. The pretraining experiments also report increased cross-entropy on removed sources with small changes on retained sources; removal performance must be interpreted relative to the retraining reference.

Critical strengths
  1. A straightforward method with a useful component ablation. The method is easy to understand, and the roles of the source-assigned task loss and the non-assigned output penalty are clearly specified. The ablation over the output-penalty weight helps explain why suppressing non-assigned outputs matters when all modules participate in inference. This makes the proposed mechanism and its behavior easier to assess.
Critical weaknesses
  1. The practical scope and significance of the proposed deployment advantage are not sufficiently established. MUSR assumes training-time source labels and source-to-module assignments, so it addresses removal from a model prepared for source-level deletion. Its applicability to sensitive information discovered after that information has already been incorporated into a pretrained backbone is less clear. Furthermore, although MUSR removes query-dependent module selection, it still requires source-to-module bookkeeping to identify the modules affected by a deletion request. These are different operations, but the introduction does not sufficiently establish the practical advantage over a router or retrieval mechanism that can exclude deleted modules. The assumptions governing source granularity and the management of the remaining module set need a clearer treatment. This limits the significance I assign to the problem formulation and the claimed simplification.

  2. Scalability to large numbers of fine-grained sources is insufficiently demonstrated. Every remaining module is evaluated for every input, and module storage and dense computation grow linearly with the number of modules. Appendix F provides useful measurements, but the reported configurations do not establish the practicality of this design when the number of independently removable sources becomes much larger. The relevant trade-off is the cost of evaluating all modules versus the selection overhead and sparse computation of routed alternatives. The fine-tuning evaluation is also confined to Llama-3.1/3.2 and Qwen2.5, leaving the generality of the findings beyond these backbone families unestablished. These limitations reduce my confidence in the broader deployment claims.

Questions for authors

The following are clarification requests that could materially affect my assessment:

  1. What assumptions are made about source identifiers and deletion granularity? In particular, how is a request handled if it concerns only part of a predefined source, or content distributed across several sources?
  2. Could you distinguish the source-to-module registry and deletion bookkeeping retained by MUSR from the inference-time selection mechanism it removes? What deployment overhead is avoided relative to a router or retrieval index that supports excluding deleted modules and adding new ones?
  3. What range of source counts is the method intended to support? Based on the measurements in Appendix F, what are the implications of substantially finer source granularity for total storage, active computation, and end-to-end inference latency?
5 comments

We thank the reviewer for the detailed feedback. We appreciate the recognition of the method’s clarity and the output-penalty ablation.

W1: Applicability to sensitive information discovered after training

We agree that MUSR requires preparation during training and does not remove sensitive information already stored in an unprepared pretrained backbone. Like other unlearning-by-construction methods, MUSR studies how to prepare a model so that later removal is inexpensive. The source labels are needed during training, but the deletion targets can be identified afterward. For example, a document can be assigned to a module during training and if sensitive content is discovered later, the assigned module can be removed or retrained without the affected information.

W1 (cont’d), Q2: Deployment advantage despite source bookkeeping

We keep a source-to-module registry to identify the modules affected by a deletion request. This lookup occurs when processing the request. A router or retrieval index can also exclude deleted modules, but still needs to select among the remaining modules for every query. GRAM and NULL rely on oracle routing for this and require the correct source label for each query. APA instead computes query embeddings and searches a retrieval index. MUSR avoids both requirements by using the same active module set for every input. We have clarified this distinction in the revised manuscript.

Q1: Partial-source requests and information distributed across modules

For deletion within a shared module, we reinitialize and retrain that module on its remaining samples, keeping all other module weights frozen. We tested deleting one author from modules containing four or ten authors on TOFU.

Authors per moduleDeleted-author recall afterSame-module retained recall afterUtility beforeUtility after
440.329±0.0070.329 \pm 0.0071.000±0.0001.000 \pm 0.0000.583±0.0050.583 \pm 0.0050.581±0.0090.581 \pm 0.009
10100.289±0.0150.289 \pm 0.0151.000±0.0011.000 \pm 0.0010.582±0.0150.582 \pm 0.0150.581±0.0130.581 \pm 0.013

The retained authors sharing the module remain available while utility stays close to its pre-deletion value. If a source spans several modules, we retrain all affected modules on their remaining samples. With random assignment, forget10 touches 177177 modules. MUSR reaches utility 0.5540.554 and forget ROUGE 0.3430.343 after retraining, compared with 0.4280.428 and 0.4090.409 for GRAM. Table 2 reports both experiments.

W2/Q3: Source counts, storage and inference cost

We extended the generation measurements in Appendix F to batch sizes 11 through 256256, with up to 20002000 modules on Llama-1B and 200200 on Llama-8B. These measurements use randomly initialized modules. Since MUSR evaluates every active module regardless of its output magnitude, they perform the same operations as trained modules of the same size and count. Selected latencies in seconds are shown below.

BackboneModulesBatch sizeFrozen base (s)MUSR (s)Routed GRAM (s)
Llama-1B200200112.7812.7813.2143.2143.3103.310
Llama-1B2002002562566.1886.1886.4256.4256.6776.677
Llama-1B2000200032322.8652.8653.3183.3183.3653.365
Llama-1B200020002562566.1886.1888.0018.0016.7166.716
Llama-8B20020025625622.81422.81423.64723.64724.61624.616

Most generation time is spent on the base model, while MUSR processes the small source modules together. MUSR has lower latency at every tested batch size for K≤200K\leq200 on both backbones, and up to 20002000 modules at batches up to 3232 on Llama-1B. At batches 128128 and 256256, routed GRAM becomes faster from 10001000 modules.

Storage and dense computation grow linearly with module count. At batch 256256, peak memory for MUSR and GRAM is 5.445.44 and 5.425.42 GiB at K=200K=200, and 16.9916.99 and 8.068.06 GiB at K=2000K=2000. We have added the measurements and benchmark details in Appendix F, Table 26.

W2 (cont’d): Generalization across model families

We have evaluated MUSR in pretraining with 26M- and 200M-parameter Transformers and in fine-tuning with Llama-3.2-1B-Instruct, Llama-3.1-8B-Instruct, Qwen2.5-1.5B-Instruct and Qwen2.5-7B-Instruct. We have now extended the fine-tuning evaluation to OLMo-2-0425-1B-Instruct and Gemma-3-1B-it. This brings the evaluation to two pretraining models and six fine-tuning backbones across four model families. The additional experiments use the same source-specific modules, training objective and module-deletion operation, with backbone-specific training configurations.

BackboneMethodUtility beforeUtility afterForget ROUGE after
OLMo-2-1BJoint fine-tuning0.578±0.0060.578 \pm 0.006----
OLMo-2-1BRetraining--0.578±0.0050.578 \pm 0.0050.375±0.0040.375 \pm 0.004
OLMo-2-1BMUSR0.690±0.0080.690 \pm 0.0080.692±0.0080.692 \pm 0.0080.403±0.0150.403 \pm 0.015
Gemma-3-1BJoint fine-tuning0.614±0.0070.614 \pm 0.007----
Gemma-3-1BRetraining--0.614±0.0110.614 \pm 0.0110.367±0.0040.367 \pm 0.004
Gemma-3-1BMUSR0.599±0.0210.599 \pm 0.0210.596±0.0100.596 \pm 0.0100.387±0.0030.387 \pm 0.003

MUSR preserves its pre-unlearning utility on both additional backbones.

On OLMo, the higher utility is reflected in both source learning and general performance. Retained-author ROUGE reaches 0.985±0.0030.985 \pm 0.003, compared with 0.949±0.0040.949 \pm 0.004 for fine-tuning, while world-fact truth ratio reaches 0.696±0.0190.696 \pm 0.019, compared with 0.532±0.0080.532 \pm 0.008. These evaluations use the same question-answering format but relate to different sources.

For Gemma, we select λout=0.3\lambda_{\mathrm{out}}=0.3 through a hyperparameter sweep. Gemma’s hidden states have larger values than Llama’s, so we use a smaller penalty weight to balance source learning and output suppression. The training objective and unlearning operation remain unchanged. Together, these results support MUSR’s applicability across varying model families and both pretraining and fine-tuning settings. We have added the results to the revised manuscript in Table 14.

We hope these results address the reviewer's concerns, and we would appreciate it if the reviewer would consider raising the score in that case. Please let us know if further clarification would be helpful.

AuthorsOpenReview ↗

The discussion phase is coming to an end

As the rebuttal period is coming to an end, thank you again for your constructive feedback. We would like to highlight the additions addressing your concerns.

  1. Deployment distinctions. We clarified that MUSR uses its source-to-module registry only when processing deletion requests. Unlike oracle-routed GRAM and NULL, MUSR does not require identifying the correct source of each incoming query, and unlike methods like APA, it needs no retrieval step. All queries use the same remaining modules.

  2. Deletion granularity. Retraining affected modules removes an author from a shared module or a source spread across modules while preserving retained information. Under random assignment, MUSR retains higher utility than the compared baselines (Table 2).

  3. Scaling and inference cost. We tested batch sizes up to 256256 and module counts up to 20002000. At K=200K=200, MUSR generates faster than oracle-routed GRAM at every tested batch size on both Llama backbones, despite evaluating all modules (Appendix F).

  4. Additional model families. Adding OLMo and Gemma extends evaluation to six fine-tuning backbones across four families. Both learn the source information and preserve utility after deletion, with forget recall returning close to the frozen-base level (Table 14).

We would be grateful if you could review these additions and consider increasing your rating if they address your concerns. We are happy to answer any remaining questions.

Reviewer_bfY9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces Modular Unlearning via Self-Routing (MUSR), a framework for knowledge unlearning in large language models that eliminates the need for external routers or source labels during inference. MUSR adds source-specific gated MLP modules in parallel with selected layers of the backbone. During training, the gradients from each task are limited to the module assigned to that data source, and an output penalty pushes the modules that are not assigned to produce outputs close to zero. As a result, all modules can be evaluated together during inference without any explicit routing. To carry out unlearning, the method zeros the output projections of the modules associated with the target source, while keeping the backbone and the other modules unchanged.

The authors evaluate MUSR in both pretraining and fine-tuning regimes, running experiments on TOFU and LUME across multiple Llama and Qwen models. The results suggest that MUSR can achieve retained utility and forgetting behavior that are similar to existing modular and post-hoc unlearning techniques. They also carry out extra studies on the effect of different module designs, repeated unlearning, the impact of adding new sources over time, the amount of residual information left behind, and the extra inference cost incurred.

Critical strengths

1.MUSR addresses a key limitation of current modular unlearning methods by getting rid of the reliance on source labels or an external router during inference. It keeps distinct, source-specific modules, yet still lets their outputs be merged directly, which makes choosing the correct module during inference straightforward. 2.The paper tests MUSR under both pretraining and fine-tuning conditions, using a range of Llama and Qwen models, along with the TOFU and LUME benchmarks. It also includes ablation studies and analyses on repeated unlearning, continual source addition, residual information, and inference overhead. Overall, the findings show strong retained utility and effective forgetting.

Critical weaknesses

1.The approach is largely aligned with GRAM, as both methods use source-specific auxiliary modules and gradient routing to confine the knowledge localization. The main distinction is the addition of an output penalty that suppresses any modules that were not assigned, which then allows their outputs to be aggregated without explicit routing. Even though this change is practically helpful, it still represents a relatively modest methodological advance compared with earlier modular unlearning approaches. 2.The output penalty used for self-routing creates cross-source dependencies, because samples from one source can end up updating modules that weren’t assigned to them. In Appendix E.6 (Figure 8), the authors demonstrate that a membership probe applied to the retained module outputs reaches an AUC of 0.911 after unlearning, whereas a model trained without the forget set only achieves an AUC of 0.562. While the authors do acknowledge this drawback and run relearning experiments, the analyses do not include direct comparisons to modular baselines such as GRAM under the same evaluation conditions. As a result, it is still not clear whether removing explicit routing leads to weaker source isolation, or to less robust unlearning, compared with existing modular methods.

Questions for authors

1.Because GRAM has already shown that it can perform inference with all auxiliary modules enabled, could the authors also evaluate MUSR against this full GRAM configuration on TOFU? In particular, how much extra improvement does the proposed output penalty yield beyond just combining all GRAM modules directly? 2.Appendix E.6 suggests that detectable, source-specific remnants may still persist in the retained modules even after unlearning, despite the relearning experiments indicating only a small impact on knowledge recovery. The authors should therefore clarify whether these residual traces introduce any additional privacy concerns. If feasible, they should also report results that are directly comparable—e.g., residual-information assessments or relearning outcomes—for modular baselines such as GRAM. This would help evaluate whether the proposed self-routing mechanism incurs extra costs for source isolation.

4 comments

We thank the reviewer for the feedback and for recognizing the practical benefit of router-free inference and the breadth of our evaluation.

W1/Q1: Distinction from GRAM

Both GRAM and MUSR use source-specific modules and gradient routing during training. While gradient routing does confine source learning to the assigned modules, it does not ensure that the other modules remain inactive on that source’s inputs. GRAM uses oracle routing during inference to handle this separation. MUSR instead uses the output penalty to train non-assigned modules to remain quiet, allowing all remaining modules to be evaluated together. This removes the requirement to identify the correct source for each incoming query, which may be unavailable in deployment. Our contribution is therefore a training objective that makes source-specific learning compatible with inference without source labels or external module selection.

Following the reviewer’s suggestion, we evaluated GRAM with all active modules summed instead of using oracle routing, both before and after deleting the forget-source modules. We tested all four TOFU backbones to determine whether GRAM can preserve source learning and retained utility without the correct source label for each query. The table below reports results for Llama-3.2-1B-Instruct.

MethodUtility beforeForget beforeUtility afterForget after
GRAM, oracle router0.557±0.0010.557 \pm 0.0011.000±0.0001.000 \pm 0.0000.557±0.0010.557 \pm 0.0010.362±0.0000.362 \pm 0.000
GRAM, all modules summed0.000±0.0000.000 \pm 0.0000.000±0.0000.000 \pm 0.0000.000±0.0000.000 \pm 0.0000.000±0.0000.000 \pm 0.000
MUSR0.589±0.0220.589 \pm 0.0220.966±0.0030.966 \pm 0.0030.579±0.0160.579 \pm 0.0160.369±0.0160.369 \pm 0.016

Summed GRAM reaches zero utility on all four backbones, before and after unlearning (Appendix E.8, Table 22). In pretraining, summing GRAM’s modules also increases source cross-entropy relative to oracle routing, with the gap growing with source count (Figure 2(b)). These results show that removing oracle selection requires changing how the modules are trained. MUSR’s output penalty enables joint evaluation while preserving source learning and retained utility. Figure 4 further isolates its effect by varying λout\lambda_{\mathrm{out}}.

W2/Q2: Residual information after module deletion

The white-box membership probe from Figure 8 tests whether the retained modules carry a trace of the deleted sources. It requires access to internal module activations and labeled examples for training the probe. Our black-box membership inference evaluations instead use model outputs and show performance close to retraining, with a Privacy score of −0.83-0.83 against the retraining reference of zero (Figure 3(d)). The difference comes from the output penalty. During training, Lout\mathcal{L}_{\mathrm{out}} teaches non-assigned modules to remain quiet on other sources’ examples, including forget10. Deleting the forget10 modules leaves this learned quietness in the retained modules, which the white-box probe detects. We additionally test whether this trace helps recover the deleted answers through relearning.

Q2 (cont’d): Comparable residual-information and relearning evaluations for GRAM

Following the reviewer’s suggestion, we added matched probes and relearning experiments for GRAM on Llama-3.2-1B-Instruct. For each method, we compare the unlearned model with a control trained without forget10. The probe tests whether retained-module outputs distinguish forget10 from unseen authors. The relearning score measures the change in answer loss after ten fine-tuning steps on forget10, adjusted by the corresponding change on unseen authors. More negative values indicate greater recovery of forget10 answers. Appendix E.6 provides the full protocol.

ModelRetained-output probe AUCRetained-module relearning
MUSR after unlearning0.911±0.0310.911 \pm 0.031−0.004±0.035-0.004 \pm 0.035
MUSR without forget100.562±0.0300.562 \pm 0.030−0.044±0.069-0.044 \pm 0.069
GRAM after unlearning0.648±0.0210.648 \pm 0.021−0.026±0.021-0.026 \pm 0.021
GRAM without forget100.642±0.0200.642 \pm 0.020−0.031±0.012-0.031 \pm 0.012

For both MUSR and GRAM, fine-tuning the retained modules after deletion does not recover forget10 answers faster than fine-tuning a model trained without forget10. We also tested fine-tuning MUSR’s backbone. After ten steps, both MUSR after deletion and the model trained without forget10 reach an answer loss of 1.801.80. Neither tested attack therefore finds faster answer recovery from MUSR’s prior exposure to forget10. We have added the matched GRAM comparison and the relearning results in Appendix E.6.

We hope that our responses address the reviewer's concerns. If so, we would be grateful if the reviewer would consider raising the score. Please let us know if any further clarification would be helpful.

Thanks for the author's additional experiments. I think the second question I was concerned about has been addressed.

AuthorsOpenReview ↗

Thank you!

Dear Reviewer bfY9,

Thank you for engaging with us in the discussion. We hope our response addressed your first question as well. Please let us know if you have any remaining concerns or if further clarification would be helpful.

If our response has addressed your concerns, we would be grateful if you would consider increasing your rating.

With kind regards,

The Authors

AuthorsOpenReview ↗

Follow-up Regarding Q1

Dear Reviewer bfY9,

Thank you again for your response. As the discussion period closes soon, we wanted to briefly follow up on question 1.

Following your suggestion, we evaluated GRAM with all modules active at the same time, without oracle routing. This configuration achieved zero utility before and after unlearning across all four TOFU backbones (Appendix E.8, Table 22), whereas MUSR preserved utility (e.g., (0.589) before and (0.579) after on Llama-3.2-1B-Instruct). Figure 2b also shows a utility drop off for GRAM in the pretraining setting if all modules are evaluated jointly.

Could you please let us know whether these experiments address question 1, or whether a specific comparison or clarification is still missing? We would be happy to respond promptly.

With kind regards, The Authors

Reviewer_ar8xOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes Modular Unlearning via Self-Routing (MUSR), an unlearning by construction method which trains a small module per added source or domain whose outputs are added to the output of chosen transformer MLP layers. While training still requires a labeled set of pre-defined sources, one of the novel contributions is to train each source-assigned module to effectively abstain on other sources/domains (Eq. 2), thus not relying on a source label as input at inference time. To unlearn a source, the module is simply disabled, returning performance to at/near the backbone model.

Critical strengths
  1. The experiments are quite extensive and generally support the authors claims well, covering numerous baselines (post-hoc, unlearning by construction, including router-free settings of other methods), and supporting both the pre-training and fine-tuning settings. The ablations are similarly very extensive.
  2. The method’s motivation is clear and well-defended by experiments. The writing is also clear and the paper is thorough.
Critical weaknesses
  1. The runtime and storage for this method increase linearly with the number of source domains kk. While this is common for unlearning by construction methods and the authors demonstrate it is reasonably efficient relative to peer methods (Table 16), this limits its applicability to some of the domains mentioned in the introduction, such as using individual users as domains, of which there could be thousands. Further, the analysis which includes latency is limited to batch size 1.
Questions for authors

None

2 comments

We thank the reviewer for the support and positive assessment of our work. We appreciate the recognition of the extensive evaluation, the method’s motivation and the clarity of the paper.

W1: Scaling and batch size

We added end-to-end generation measurements across batch sizes 11 to 256256, with up to 20002000 modules on Llama-1B and 200200 on Llama-8B. Selected latencies in seconds are shown below.

BackboneBatch sizeModulesFrozen base (s)MUSR (s)Routed GRAM (s)
Llama-1B3232200020002.8652.8653.3183.3183.3653.365
Llama-1B2562562002006.1886.1886.4256.4256.6776.677
Llama-1B256256100010006.1886.1887.2087.2086.7156.715
Llama-1B256256200020006.1886.1888.0018.0016.7166.716
Llama-8B25625620020022.81422.81423.64723.64724.61624.616

Most generation time is spent on the base model, while MUSR processes the small source modules together. At K=200K=200, MUSR has lower latency than routed GRAM at every tested batch size on both backbones. On Llama-1B, this also holds up to 20002000 modules at batch sizes up to 3232. At batches 128128 and 256256, routed GRAM becomes faster from 10001000 modules. Storage and dense computation grow linearly with the number of modules. We have extended Appendix F with these measurements and the benchmark details.

We believe the additional experiments and clarifications address the points raised, and we would appreciate it if the reviewer would champion our paper. We are happy to provide further details if needed.

AuthorsOpenReview ↗

The discussion phase is coming to an end

As the rebuttal period ends, thank you again for your positive assessment. Based on your feedback, we added generation and memory measurements across batch sizes up to 256256 and module counts up to 20002000. At K=200K=200, MUSR generates faster than oracle-routed GRAM at every tested batch size on both Llama-1B and Llama-8B, despite evaluating all modules (Appendix F).

We hope these results address your concerns and would appreciate it if the reviewer would champion our paper. We are happy to clarify any remaining questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank all reviewers for their constructive feedback. We appreciate their recognition of the practical benefit of removing inference-time source labels and external routing (bfY9), the clear motivation and extensive evaluation across pretraining and fine-tuning settings (ar8x), and the preservation of utility after unlearning (bfY9, xa59). The reviewers also highlighted the clarity of the method and the ablation showing the role of the output penalty (ar8x, cwwd).

We have expanded the revised manuscript with a significant amount of experiments and analyses addressing the reviewers’ feedback. These changes are marked in blue. We present the highlights of our rebuttal below:

  1. Reduced pretraining overshoot on LISP. (xa59) We reduced the excess cross-entropy on LISP by training the backbone alone on a fraction of shared core batches. This brings MUSR within 0.0530.053 of retraining on average across the four sources, while changing retained cross-entropy by only 0.0020.002. MUSR achieves better source learning than the compared baselines without an inference router. The updated results are in Table 1 and Appendix C, Tables 7 and 8.

  2. Extended GRAM comparison. (bfY9) We evaluated GRAM with all modules summed on all four original TOFU backbones. Without oracle routing, GRAM’s utility drops to zero before and after unlearning on every model. MUSR preserves source learning and retained utility without a router, demonstrating why our proposed output penalty is needed to make dense module evaluation work (Appendix E.8, Table 23).

  3. Diverse model families. (cwwd) We extended TOFU evaluation to OLMo and Gemma, bringing fine-tuning evaluation to six backbones across four model families. Both additional backbones preserve utility after unlearning (Table 14).

  4. Deletion under different source assignments. (cwwd) We tested deleting a source whose data is spread across several shared modules. Reinitializing and retraining the affected modules removes the requested information while preserving the utility on the remaining sources. Under random sample assignment, MUSR retains high utility while the baselines do not retain much information during training, showing that our advantage extends beyond perfect one-source-per-module assignments (Table 2).

  5. Strengthened privacy and relearning analysis. (bfY9, xa59) We added retained-module relearning attacks and included matched GRAM privacy comparisons. No tested relearning attack recovers deleted answers faster than a model trained without the forget data. Together with MUSR’s TOFU Privacy score of −0.83-0.83, close to the retraining reference of zero, these results support its privacy robustness under the tested attacks (Appendix E.6).

  6. Post-hoc unlearning in the pretraining setting. (xa59) We added GradDiff, NPO, SimNPO and RMU to the realistic pretraining evaluation. NPO is the only tested post-hoc method that approaches retraining with preserved retained performance, while requiring 10001000 steps per request. MUSR achieves both through module deletion alone. (Appendix C.3, Table 9).

  7. End-to-end inference measurements. (ar8x, cwwd, xa59) We extended generation latency and memory measurements across batch sizes and module counts. At K=200K=200, MUSR generates faster than oracle-routed GRAM at every tested batch size from 11 to 256256 on both Llama-1B and Llama-8B, despite evaluating all modules (Appendix F, Table 26).

  8. Module activity on non-source inputs. (xa59) We measured module activity on MMLU and extended the output penalty with general-text negatives from Alpaca [2]. This preserves MMLU performance close to the frozen base while maintaining LUME learning and unlearning performance, without additional inference computation (Appendix D.3, Table 17).

  9. Overlapping facts across sources. (xa59) Natural-overlap and shared-attribute experiments show that MUSR preserves shared information about retained authors when unlearning other authors’ facts (Appendix E.9, Table 24).

We have also clarified the source assumptions, partial-source requests, and the distinction to Gram, and deletion bookkeeping versus inference-time selection. Detailed results and responses to individual concerns are provided in the responses.

Additional Highlights

Beyond the requested experiments, we added two further evaluations:

  1. Comparison with a learned router. We added FlexOlmo [1] to realistic and synthetic pretraining, extending the comparison to learned routing. MUSR achieves better source learning in the realistic setting without an inference router (Table 1, Figure 2).

  2. Additional TOFU evaluation. We added TOFU's Forget Quality for all methods across four backbones, comparing forget-set truth ratios to complement ROUGE and membership attacks (Appendix D.1, Table 15).

[1] FlexOLMo: Open Language Models for Flexible Data Use, Shi et al., NeurIPS 2025.

[2] Stanford Alpaca: An Instruction-following LLaMA model, Taori et al., 2023

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.