Pith. sign in

REVIEW 3 major objections 4 minor 60 references

A Model Merging Approach for Continual MLLM Unlearning

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Continual unlearning for multimodal LLMs can be done by merging one-shot adapters instead of retraining the model on each new deletion request.

desk verdict A genuinely new and honest method paper for continual multimodal unlearning via adapter merging; the empirical gains look real, but the causal story about Gram reconfiguration outruns the evidence. read the letter →

arxiv 2608.04548 v1 pith:7QE55K4W submitted 2026-08-05 cs.LG cs.AI

Yuhang Wang , Linlin Zhang , Haoxuan Ji , Xianmin Ye , Zhenxing Niu , Haichang Gao This is my paper · ORCID
classification cs.LGcs.AI
keywords continualunlearningmultimodallargelanguagemodelsmodelmergingLoRAadapterscross-taskinterferencetaskvectorsmachineparametergeometry
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Continual unlearning for multimodal large language models usually means applying one unlearning step after another, and each step can overwrite earlier forgetfulness, revive supposedly deleted knowledge, and erode what the model should keep. The paper argues these failures are avoidable: each deletion request can be handled independently by a one-shot LoRA adapter, and the adapters can then be merged into a single update. To make the merge work, MCU must deal with the fact that adapters trained for different requests share overlapping directions, which helps transfer forgetfulness across tasks but also causes interference. The proposed solution projects all adapters into a shared low-rank core space, keeps their dominant directions, dampens over-concentrated coordinates, and reconfigures cross-task directional dependencies before summing the updates. On ICU-Bench and MLLMU-Bench, this merge-only procedure outperforms sequential retraining baselines and ordinary merging methods on both forget and retain metrics.

What carries the argument

The central object is the unlearning task vector: the LoRA update $\Delta W_t = B_t A_t$ produced independently for each request from the same base model. MCU's machinery has four stages: (1) a shared core-space projection $\Delta W_t \mapsto M_t = P^\top \Delta W_t Q$ built from the joint column and row spaces of all accumulated adapters; (2) dominant-direction selection, truncating each $M_t$ to its top-$k$ singular components; (3) row-capacity control, softly rescaling rows of the retained update whose norms exceed a quantile threshold; and (4) dependency-aware reconfiguration, which enforces non-negative cross-request similarities on the left and right Gram matrices $G_U = S_U^\top S_U$, $G_V = S_V^\top S_V$, then recovers directions by orthogonal Procrustes alignment and merges the reconstructed updates. The load-bearing identity is the sign-invariant interaction $\chi_{pq} = (u_p^\top u_q)(v_p^\top v_q)$ between rank-one update components, whose non-negativity the Gram reconfiguration approximates through a separable sufficient surrogate.

What would settle it

Run MCU on a pair of requests whose sign-invariant interaction $\chi_{pq}$ is negative after merging; if the merged model still shows cancellation of forgetting or degraded retain performance on those two targets, then the surrogate did not control interference. More directly, a leave-one-task-out experiment in which dropping one adapter produces almost no diagonal recovery on its own target would falsify the claimed cross-task transfer mechanism, since MCU predicts that reconfiguration should substantially raise diagonal recovery and reduce off-diagonal deviations.

Watch

Extended reading notes

Core claim

The central claim is that continually unlearning many sensitive items from an MLLM does not require continual modification of the model. Instead, one-shot unlearning adapters, each trained from the same base checkpoint for one request, can be consolidated into a single merged adapter that preserves each request's forgetting effect while keeping retained knowledge intact. The discovery that makes this possible is cross-task unlearning transfer: a leave-one-out merging study shows that omitting an adapter for a given target still leaves part of that target forgotten, because adapters for other requests share overlapping update directions. MCU treats these shared directions as a mixture of synergy and interference, and reshapes them through a Gram-matrix optimization that forces non-negative cross-request similarities on both left and right singular-direction spaces, thereby suppressing antagonistic interactions without destroying beneficial transfer. The empirical claim, demonstrated for up to 100 sequential requests on two backbones, is that this merged model matches or beats one-shot baselines on forgetting while sharply reducing rebound, retention drift, and utility collapse.

Load-bearing premise

MCU assumes that making pairwise update directions non-negative in the shared parameter space is a sufficient stand-in for avoiding actual behavioral interference between unlearning requests; the paper's own limitations section notes that parameter geometry does not fully determine functional behavior, so if this geometric-to-behavioral link fails the method's advantage is unsupported.

Editorial extensions

If this is right

  • On ICU-Bench, MCU's forget and retain scores at 100 tasks stay close to its own task-10 scores, while sequential baselines degrade sharply over the sequence.
  • At task 100 on Qwen2-VL-7B, MCU lowers forgetting rebound to 1.12 and retain stability rate to 1.10, and generation quality scores stay near 1.99 out of 2.
  • The merge is order-independent: original, reversed, and random permutations of the same adapter bank give identical forget and retain scores, because the procedure operates on the collection of adapters, not on a running model state.
  • Consolidating 50 adapters takes about 98 seconds on CPU with no GPU memory allocation, and the final deployed model has a size independent of the number of accumulated requests.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the Gram reconfiguration acts on pairwise geometry rather than on forward behavior, a natural next step is to optimize the reconfiguration against measured output divergence on retain and forget sets; the paper explicitly lists this functional characterization as open.
  • The order-independence result implies the method could serve as an append-only unlearning service: requests arrive asynchronously, adapters are trained independently, and consolidation happens in a single batch, which fits auditing and rollback workflows.
  • The signed-graph obstruction, where three mutually conflicting directions cannot all be made non-negative by sign flips, sets a limit on the separable surrogate for more than two strongly conflicting requests, pointing toward block-wise or hierarchical merging as a remedy.
  • If cross-task transfer is a general property of LoRA-based unlearning rather than of these two document benchmarks, the same merging recipe should work on other MLLM unlearning settings, such as entity-level or scene-level forgetting.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces Merging for Continual Unlearning (MCU), a model-merging framework for continual multimodal LLM unlearning. Each new deletion request is handled by independently training a one-shot LoRA unlearning adapter from the same base model, and MCU merges the accumulated adapters into a single unified update. The method projects adapters into a shared core space, keeps their top-k singular directions, applies a soft row-capacity constraint, and then reconfigures cross-request singular-direction geometry by solving a Gram-matrix optimization problem with non-negative cross-request entries. The authors report that MCU outperforms sequential unlearning baselines and standard merging baselines on ICU-Bench and MLLMU-Bench with two 7B backbones, reducing unlearning rebound and retention drift while preserving retained knowledge and external utility.

Significance. If the central claim holds, MCU offers a genuinely different paradigm for continual unlearning: instead of repeatedly updating a model, one merges independently trained one-shot adapters, which is training-free at merge time and gives a compact request history. The paper is empirically broad (two benchmarks, two backbones, sequence lengths up to 100 requests), and it includes useful sanity checks such as leave-one-task-out attribution, order-robustness, external VQAv2 utility, generation-quality checks, and explicit cost measurements. The authors also state several limitations honestly, including the conservative surrogate and the open link between parameter geometry and functional behavior. The main weaknesses are the absence of uncertainty quantification, the selection of core hyperparameters on the evaluation benchmarks, and the lack of a direct causal test connecting the optimized Gram geometry to the observed unlearning/retention gains.

major comments (3)
  1. [Method, Dependency-aware direction reconfiguration; Supplementary, Limitations and Scope] The central mechanistic claim is that suppressing antagonistic cross-task interactions while retaining synergistic ones explains MCU's gains, but the paper does not provide a direct test of this causal link. Equation (8) defines χ_ij,ab as a parameter-space product of left and right direction similarities, and the optimization in Eq. (16) only enforces non-negative cross-request Gram entries. As the authors acknowledge in the supplementary Limitations, this condition is sufficient but not necessary for a non-antagonistic rank-one interaction. More importantly, the sign of χ_ij,ab does not determine the functional consequence: cancellation can either weaken a target unlearning update or reduce collateral damage to retained knowledge. The only behavioral evidence for the geometric-to-behavioral link is the 20-task leave-one-out VQA analysis (Fig. 3 and Table 7), which is not replicated on MLLMU-Bench, at longer horizons, or for QA and Retain axes, and which does not correlate χ values with functional deviations. Without such a test, the ablation improvements in Table 4 could be driven by low-rank truncation and capacity control rather than by the Gram reconfiguration presented as the core contribution.
  2. [Tables 1-4 and Supplementary, Hyperparameter Settings] The main empirical claims rest on single-point accuracy numbers without error bars, standard deviations, or significance tests. MCU's hyperparameters k=6, q=0.95, and the per-backbone α are selected on the same ICU-Bench and MLLMU-Bench evaluation sets used to report the headline results, and no code or trained checkpoints are provided. This combination makes it difficult to determine whether MCU's large reported gains are robust or partly reflect tuning to the evaluation benchmarks. I would like to see repeated runs or at least bootstrap intervals, and a validation protocol in which hyperparameters are chosen on a held-out subset or on one benchmark and evaluated on the other.
  3. [Experiments, Continual unlearning on MLLMU-Bench] The MLLMU-Bench continual protocol uses only five stages and a single fixed retain subset, which provides a relatively weak long-horizon test; the paper's strongest claims about 'long-horizon' continual unlearning depend primarily on ICU-Bench. The MLLMU-Bench stage-wise results in Fig. 5 are presented only as heatmaps without numeric values, making the claimed 'stable Retain performance' hard to quantify. Reporting the actual stage-wise Forget and Retain numbers, including historical targets, in a table would strengthen the central claim.
minor comments (4)
  1. [Abstract and full text] The rendered text contains many missing spaces (e.g., 'Multimodallargelanguagemodel', 'MergingforContinualUnlearning', 'Extensiveexperiments'), which appear to be a PDF-extraction artifact; the camera-ready version should be checked for this.
  2. [Figure 3] The caption and the in-text discussion refer to 'Before Dependency-Aware Reconfiguration' and 'After Dependency-Aware Reconfiguration', but the figure itself does not label the two matrices clearly; adding explicit panel labels would improve readability.
  3. [Table 3] The abbreviations RSR, FR, and GQ-F/GQ-R are used in the main text but defined only in the supplementary material; a brief parenthetical definition at first use in the main paper would help.
  4. [Supplementary, Robustness to Request Order] Table 8 shows identical numbers under original, reversed, and random orders. Since the algorithm is permutation-invariant by construction (it operates on the set of adapters rather than their order), this result is a sanity check rather than an empirical robustness finding; the text should present it as such.

Circularity Check

0 steps flagged · score 3.0 of 10

No derivation-level circularity; mild self-evaluation loop because ICU-Bench is authored by the same group and the MCU hyperparameters are selected from ICU-Bench statistics on which the headline results are reported.

full rationale

Walking the derivation chain: core-space projection (Eq. 4), dominant-direction truncation (Eq. 9), row-capacity control (Eq. 11), Gram reconfiguration (Eq. 16), Procrustes recovery (Eq. 18), and summation (Eq. 20) are each well-defined optimizations with stated objectives; the final merged update is not equivalent to any fitted input by construction, and the paper explicitly disclaims optimality for downstream behavior. The 'cross-task unlearning transfer' phenomenon is an empirical leave-one-out claim (Fig. 3, Table 7), not a definitional consequence of the merging formula. The paper is also transparent about the surrogate's insufficiency (supplementary 'Limitations and Scope': non-negativity on both Gram sides is 'sufficient, but not necessary'; 'a complete functional characterization remains an open direction'), which is a validity gap, not a circularity. The two mild circularity-adjacent issues are: (1) the retained rank, capacity quantile, and merging coefficient are selected from ICU-Bench statistics and headlong results are reported on the same ICU-Bench sequence, so the headline numbers are partly fit-validated-on-fit, mitigated by the fixed-configuration evaluation on external MLLMU-Bench and selection-free VQAv2; (2) ICU-Bench itself is a same-group benchmark, so the strongest long-horizon claim leans on a self-referential evaluation base, though external corroboration prevents the claim from reducing to self-citation. Overall score 3: below the threshold of construction-level circularity, but the self-evaluation loop is real and should be disclosed as a benchmarking limitation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central effectiveness claim rests on three tuned hyperparameters, an unreported LoRA rank, and the assumption that Gram-geometry reconfiguration is behaviorally beneficial. The mathematical propositions are standard, but the behavioral link between the reconfiguration and improved unlearning is empirical rather than derived.

free parameters (4)
  • Top-k singular directions = 6
    Retains six leading singular components per structured update; chosen based on cumulative spectral energy on ICU-Bench and validated by ablations, not derived from first principles.
  • Row-capacity quantile q = 0.95
    Rows above the 95th percentile are rescaled; chosen from row-norm concentration observed on ICU-Bench.
  • Global merging coefficient alpha = 0.4 (Qwen2-VL-7B), 1.0 (LLaVA-1.5-7B)
    Applied to the merged update in Eq. (3); tuned per backbone on the evaluation benchmarks.
  • LoRA rank r = not reported
    The rank of one-shot LoRA adapters is never stated; it controls the shared core-space dimension and is likely inherited from ICU-Bench.
assumptions (5)
  • standard math Eckart-Young-Mirsky theorem
    Used to justify top-k SVD truncation in Proposition 1 of the supplementary material.
  • standard math Euclidean projection onto row-norm ball
    Used to justify the row-capacity control operator in Proposition 2.
  • domain assumption Independently trained adapters from the same base model can be merged in a shared core space
    Eqs. (1)-(4) and (37)-(43) assume all updates share the same base initialization, enabling reversible core-space projection.
  • domain assumption Gram-matrix geometry reflects functional unlearning behavior
    The method reshapes parameter-space Gram matrices to reduce interference, but the paper's limitations state that parameter geometry does not fully determine functional behavior.
  • ad hoc to paper Separable non-negativity surrogate is an adequate tractable proxy
    Eq. (16) and the supplementary state the constraint is sufficient but not necessary for non-antagonistic rank-one interactions; adopted for tractability rather than derived from first principles.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Model Merging Approach for Continual MLLM Unlearning." pith.science (2026). https://pith.science/paper/7QE55K4W

@misc{pith2026260804548,
  author       = {Pith},
  title        = {Pith review of: A Model Merging Approach for Continual MLLM Unlearning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7QE55K4W}},
  note         = {Machine review of arXiv:2608.04548}
}
read the original abstract

Multimodal large language model (MLLM) unlearning methods have been proposed to remove private, sensitive, or proprietary information from well-trained models. However, most existing MLLM unlearning methods are designed for one-shot requests and fail to adequately address continual scenarios, as repeatedly applying one-shot operations leads to cumulative utility degradation, unlearning rebound, and retention drift. We introduce Merging for Continual Unlearning (MCU), an approach that dynamically merges multiple one-shot unlearning adapters into a unified adapter upon receiving each new unlearning request.Through a leave-one-out merging analysis, we reveal that these unlearning adapters exhibit strong cross-task dependencies. Such dependencies have two contrasting effects: they can facilitate cross-task unlearning transferability, but they can also introduce severe interference that degrades unlearning effectiveness and compromises retained knowledge. To address this challenge, MCU projects the adapters into a shared representation space, preserves their dominant directions, suppresses over-concentrated coordinates, and reconfigures cross-task dependencies to mitigate interference while enhancing transferability. Experiments on ICU-Bench and MLLMU-Bench demonstrate that MCU achieves superior unlearning effectiveness while preserving both retained knowledge and general multimodal utility.

Figures

Figures reproduced from arXiv: 2608.04548 by the authors.

Figure 1
Figure 1. Comparison between sequential continual unlearn [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall framework of MCU. (1) One-shot unlearning LoRA adapters derived from the same base model are projected [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Leave-one-task-out VQA analysis before and af [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: (a) Off-diagonal task-direction overlap of the singular subspaces. unlearning adapters exhibit substantially stronger [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Stage-wise VQA performance under the five-stage continual unlearning protocol on MLLMU-Bench. Columns denote [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 14 canonical work pages

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

  5. [5]

    Clancey and Glenn Rennels , abstract =

    Diane Warner Hasling and William J. Clancey and Glenn Rennels , abstract =. Strategic explanations for a diagnostic consultation system , journal =. 1984 , issn =. doi:https://doi.org/10.1016/S0020-7373(84)80003-6 , url =

    doi:10.1016/s0020-7373(84)80003-6 1984

  6. [6]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

  7. [7]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

  8. [8]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

Show all 60 references
  1. [9]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

  2. [10]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    2023

  3. [11]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

  4. [12]

    Intersoft Consulting, Accessed in October , volume=

    General data protection regulation (GDPR) , author=. Intersoft Consulting, Accessed in October , volume=

  5. [13]

    Santa Clara Univ

    An introduction to the california consumer privacy act (ccpa) , author=. Santa Clara Univ. Legal Studies Research Paper , year=

  6. [14]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Mllm-protector: Ensuring mllm’s safety without hurting performance , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  7. [15]

    arXiv preprint arXiv:2401.00676 , year=

    Digger: Detecting copyright content mis-usage in large language model training , author=. arXiv preprint arXiv:2401.00676 , year=

    arXiv

  8. [16]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Performance gap in entity knowledge extraction across modalities in vision language models , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  9. [17]

    Annual International Conference on the Theory and Applications of Cryptographic Techniques , pages=

    Formalizing data deletion in the context of the right to be forgotten , author=. Annual International Conference on the Theory and Applications of Cryptographic Techniques , pages=. 2020 , organization=

    2020

  10. [18]

    Advances in neural information processing systems , volume=

    Making ai forget you: Data deletion in machine learning , author=. Advances in neural information processing systems , volume=

  11. [19]

    2015 IEEE symposium on security and privacy , pages=

    Towards making systems forget with machine unlearning , author=. 2015 IEEE symposium on security and privacy , pages=. 2015 , organization=

    2015

  12. [20]

    Advances in Neural Information Processing Systems , volume=

    Adaptive machine unlearning , author=. Advances in Neural Information Processing Systems , volume=

  13. [21]

    Advances in Neural Information Processing Systems , volume=

    Remember what you want to forget: Algorithms for machine unlearning , author=. Advances in Neural Information Processing Systems , volume=

  14. [22]

    International conference on machine learning , pages=

    Machine unlearning for random forests , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  15. [23]

    arXiv preprint arXiv:2606.12809 , year=

    MLUBench: A Benchmark for Lifelong Unlearning Evaluation in MLLMs , author=. arXiv preprint arXiv:2606.12809 , year=

    arXiv

  16. [24]

    arXiv preprint arXiv:2605.05938 , year=

    ICU-Bench: Benchmarking Continual Unlearning in Multimodal Large Language Models , author=. arXiv preprint arXiv:2605.05938 , year=

    arXiv

  17. [25]

    International Conference on Learning Representations , volume=

    On large language model continual unlearning , author=. International Conference on Learning Representations , volume=

  18. [26]

    International Conference on Learning Representations , volume=

    Muse: Machine unlearning six-way evaluation for language models , author=. International Conference on Learning Representations , volume=

  19. [27]

    arXiv preprint arXiv:2507.01271 , year=

    Pulse: Practical evaluation scenarios for large multimodal model unlearning , author=. arXiv preprint arXiv:2507.01271 , year=

  20. [28]

    Protecting privacy in multimodal large language models with mllmu-bench , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2025

  21. [29]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Cross-modal unlearning via influential neuron path editing in multimodal large language models , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  22. [30]

    2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P) , pages=

    Unrolling sgd: Understanding factors influencing machine unlearning , author=. 2022 IEEE 7th European Symposium on Security and Privacy (EuroS&P) , pages=. 2022 , organization=

    2022

  23. [31]

    Conference on Lifelong Learning Agents , pages=

    Continual learning and private unlearning , author=. Conference on Lifelong Learning Agents , pages=. 2022 , organization=

    2022

  24. [32]

    arXiv preprint arXiv:2401.06121 , year=

    Tofu: A task of fictitious unlearning for llms , author=. arXiv preprint arXiv:2401.06121 , year=

    arXiv

  25. [33]

    Advances in neural information processing systems , volume=

    Direct preference optimization: Your language model is secretly a reward model , author=. Advances in neural information processing systems , volume=

  26. [34]

    arXiv preprint arXiv:2404.05868 , year=

    Negative preference optimization: From catastrophic collapse to effective unlearning , author=. arXiv preprint arXiv:2404.05868 , year=

    arXiv

  27. [35]

    arXiv preprint arXiv:2507.19894 , year=

    A survey on generative model unlearning: Fundamentals, taxonomy, evaluation, and future direction , author=. arXiv preprint arXiv:2507.19894 , year=

    arXiv

  28. [36]

    arXiv preprint arXiv:2605.18879 , year=

    ZeroUnlearn: Few-Shot Knowledge Unlearning in Large Language Models , author=. arXiv preprint arXiv:2605.18879 , year=

    arXiv

  29. [37]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Mmunlearner: Reformulating multimodal machine unlearning in the era of multimodal large language models , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  30. [38]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Modality-aware neuron pruning for unlearning in multimodal large language models , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  31. [39]

    Advances in Neural Information Processing Systems , volume=

    Single image unlearning: Efficient machine unlearning in multimodal large language models , author=. Advances in Neural Information Processing Systems , volume=

  32. [40]

    arXiv preprint arXiv:2512.11325 , year=

    MLLM Machine Unlearning via Visual Knowledge Distillation , author=. arXiv preprint arXiv:2512.11325 , year=

  33. [41]

    arXiv preprint arXiv:2605.05909 , year=

    Null Space Constrained Contrastive Visual Forgetting for MLLM Unlearning , author=. arXiv preprint arXiv:2605.05909 , year=

    arXiv

  34. [42]

    arXiv preprint arXiv:2406.14796 , year=

    Mu-bench: A multitask multimodal benchmark for machine unlearning , author=. arXiv preprint arXiv:2406.14796 , year=

    arXiv

  35. [43]

    arXiv preprint arXiv:2503.12545 , year=

    Pebench: A fictitious dataset to benchmark machine unlearning for multimodal large language models , author=. arXiv preprint arXiv:2503.12545 , year=

    arXiv

  36. [44]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Clear: Character unlearning in textual and visual modalities , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  37. [45]

    Advances in Neural Information Processing Systems , volume=

    Umu-bench: Closing the modality gap in multimodal unlearning evaluation , author=. Advances in Neural Information Processing Systems , volume=

  38. [46]

    Engineering Applications of Artificial Intelligence , volume=

    Forgetme: Benchmarking the selective forgetting capabilities of generative models , author=. Engineering Applications of Artificial Intelligence , volume=. 2025 , publisher=

    2025

  39. [47]

    Advances in Neural Information Processing Systems , volume=

    Merging models with fisher-weighted averaging , author=. Advances in Neural Information Processing Systems , volume=

  40. [48]

    International conference on machine learning , pages=

    Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  41. [49]

    arXiv preprint arXiv:2212.04089 , year=

    Editing models with task arithmetic , author=. arXiv preprint arXiv:2212.04089 , year=

    arXiv

  42. [50]

    Advances in neural information processing systems , volume=

    Ties-merging: Resolving interference when merging models , author=. Advances in neural information processing systems , volume=

  43. [51]

    Forty-first International Conference on Machine Learning , year=

    Language models are super mario: Absorbing abilities from homologous models as a free lunch , author=. Forty-first International Conference on Machine Learning , year=

  44. [52]

    arXiv preprint arXiv:2410.03617 , year=

    What matters for model merging at scale? , author=. arXiv preprint arXiv:2410.03617 , year=

    arXiv

  45. [53]

    Dynamic fisher-weighted model merging via bayesian optimization , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2025

  46. [54]

    Advances in Neural Information Processing Systems , volume=

    Accurate and efficient low-rank model merging in core space , author=. Advances in Neural Information Processing Systems , volume=

  47. [55]

    Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

    Task singular vectors: Reducing task interference in model merging , author=. Proceedings of the Computer Vision and Pattern Recognition Conference , pages=

  48. [56]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. Iclr , volume=

  49. [57]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Improved baselines with visual instruction tuning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  50. [58]

    arXiv preprint arXiv:2409.12191 , year=

    Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution , author=. arXiv preprint arXiv:2409.12191 , year=

    arXiv

  51. [59]

    arXiv preprint arXiv:2605.18610 , year=

    CATA: Continual Machine Unlearning via Conflict-Averse Task Arithmetic , author=. arXiv preprint arXiv:2605.18610 , year=

    arXiv

  52. [60]

    Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    Lmms-eval: Reality check on the evaluation of large multimodal models , author=. Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.