Pith. sign in

REVIEW 4 major objections 4 minor 82 references

Toward Fine-Grained Forgetting:Attribute Unlearning for Multimodal Large Language Models

T0 review · 4 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 A single-layer, training-free projection can erase one attribute of a person from a multimodal AI model while leaving the same person's other facts intact.

desk verdict Attribute-level unlearning benchmark is a real contribution, but CLRP's headline cloze gains are produced by a test-set-only logit subtraction, so the method comparison doesn't hold. read the letter →

arxiv 2608.01008 v1 pith:SJ666KFL submitted 2026-08-02 cs.AI

Junkai Lin , Junkai Chen , Siqi Hou , Yuhao He , Ruiqi Liu , Chenhan Jin , Shengze Xu , Tieyong Zeng This is my paper
classification cs.AI
keywords attribute-levelunlearningmultimodallargelanguagemodelsmachineactivationpatchingsubspaceprojectionCLRPAMU-Benchprivacy
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

The paper introduces a finer-grained form of machine unlearning for multimodal large language models: instead of deleting an entire profile or identity, the model should forget one specified attribute, such as a birth date, employer, or hobby description, while still answering other questions about the same person. To test this, the authors build AMU-Bench, a benchmark of profile-attribute pairs spanning long-text, numeric, and short-text targets, and find that existing unlearning methods trade off badly between forgetting and same-profile retention. Their remedy is CLRP, a training-free method that uses activation patching to locate the layer causally responsible for the target attribute, then applies a retain-aware projection that removes the attribute's subspace from that layer's hidden states while preserving directions shared with non-target facts. On LLaVA-1.5-7B and Qwen2.5-VL-3B at a 5% forget ratio, the paper reports that CLRP cuts Attribute Forget Set cloze accuracy by up to 16 percentage points, from 20.00% to 4.00%, and raises Attribute Retain Set cloze accuracy by up to 3.02 points, from 14.77% to 17.79%, relative to the strongest baseline.

What carries the argument

The central mechanism is a two-stage, training-free intervention called CLRP. Stage one, causal layer localization, uses activation patching to select the layer $\ell^*$ whose attention-head output, when patched into a paired retain run, most restores target-answer evidence; this determines where to intervene. Stage two, retain-aware contrastive projection, estimates a low-rank subspace $V$ from contrastive target/retain representations at that layer, whitening retain variation by $W = Q \Lambda^{-1/2} Q^\top$ and taking the leading eigenvectors of $W S_f W$; the operator $P = (1+\epsilon)I - V V^\top$ then scales down the target subspace and slightly scales up everything else. A forward hoo

What would settle it

Re-run the 5% LLaVA-1.5-7B conditions with the 4.0 output-score suppression removed from CLRP and applied equally to the baselines; if CLRP's Attribute Forget and Test cloze accuracies lose most of their advantage, the central forgetting result is an artifact of evaluation-side answer blocking.

Watch

Extended reading notes

Core claim

The central claim is that attribute-level unlearning can be achieved without retraining or parameter updates by intervening at a single causally selected layer. CLRP first pairs target queries with same-identity retain queries and uses activation patching to find the attention head whose restoration most strongly brings back target-answer evidence; that head is used only to pick the layer. At that layer, a contrastive feature map builds target-conditioned and retain-conditioned statistics, a whitening transform ranks directions by target variation relative to retain variation, and the top-k directions form a subspace V; the operator $P = (1+\epsilon)I - V V^\top$ is applied to the hidden sta

Load-bearing premise

The central claim depends on the evaluation being allowed to suppress the correct answers on the Attribute Forget and held-out Test cloze questions directly at scoring time; if that suppression is removed, or applied equally to all methods, the reported forgetting gains may not come from the projection itself.

Editorial extensions

If this is right

  • Attribute-level unlearning becomes a distinct, measurable task: a request can name a single profile-attribute pair, and success requires both held-out reformulation of the forgotten attribute and preservation of the same profile's other attributes.
  • A deployed multimodal assistant could forget a single sensitive field, such as salary, birth date, or employer, without retraining, while still recognizing and answering other questions about the same person.
  • Evaluation of multimodal unlearning should separate same-profile retention from retention of unrelated profiles, because the former is systematically harder and is masked by profile-level benchmarks.
  • Forgetting difficulty depends on attribute type: numeric facts are much easier to suppress than long-text or short-text facts, so unlearning scores should be reported per attribute category rather than only as aggregates.
  • Because CLRP changes no weights, the intervention can be applied or removed at inference time and could be combined with other frozen-model interventions.

Reading between the lines

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

  • The appendix's evaluation subtracts 4.0 from the correct-answer logits on the Attribute Forget and held-out Test cloze sets only; an independent replication that removes that output-space adjustment, or applies it equally to all baselines, would clarify how much of the reported forgetting gain comes from the projection itself.
  • CLRP needs paired target and retain queries for the same identity, which suggests a user-facing 'forget this attribute' API is plausible, but its selectivity may weaken when a user cannot articulate a clean retain question.
  • Because the intervention is localized to one layer, composing it with other training-free safety or privacy projections is a natural next test; the question is whether two such projections interfere when applied at different layers.
  • AMU-Bench uses fictitious profiles, so applying the same attribute-level protocol to real-person data would test whether the observed selectivity transfers beyond synthetic identities.
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

4 major / 4 minor

Summary. The paper introduces attribute-level multimodal unlearning as a finer-grained deletion task than profile-level unlearning, constructs the AMU-Bench benchmark, and proposes CLRP, a training-free method that first uses activation patching to localize a causally relevant layer and then applies a retain-aware contrastive projection to attenuate the target attribute subspace. Experiments on LLaVA-1.5-7B/13B and Qwen2.5-VL-3B report strong Attribute Forget Set cloze reductions (e.g., 20.00% to 4.00%) and modest Attribute Retain Set improvements over four training-based baselines. The appendix describes an additional 'Inference-Time Intervention' that subtracts 4.0 from correct-answer logits on the Attribute Forget Set and held-out Test Set only, which directly affects the headline cloze metric.

Significance. Attribute-level unlearning is a useful and timely problem, and the AMU-Bench construction with three attribute types, multiple forget ratios, and separate same-profile retain sets is a valuable contribution. If the reported results were valid, CLRP would be a notable training-free intervention with practical appeal. However, the central empirical claim is not supported as presented because the main cloze metric is influenced by an output-space logit offset applied only to CLRP and only to the forget/test sets, and because the ablation table shows cloze accuracy is invariant to all representation-level variants. The benchmark itself may still be of interest, but the method comparison in the current form is not fair.

major comments (4)
  1. [Appendix, 'Inference-Time Intervention'; Table 1] The appendix states that 'the implementation realizes target suppression in output space by subtracting 4.0 from the corresponding logits on the Attribute Forget Set and held-out Test Set; no such operation is applied to a retain set.' This is a test-set-specific logit penalty applied only to CLRP. It directly lowers the cloze accuracies that are the headline results (e.g., Table 1: LLaVA long-text 20.00% to 4.00%; the abstract's 16.00-point claim), and it requires oracle knowledge of which examples are forget/test queries. This makes the comparison to baselines unfair and conflicts with Definition 1's requirement that forgetting generalize to held-out reformulations rather than be produced by an answer-slot blocker. Without this operation, or with it applied equally to all methods, the reported cloze gains are not shown to originate from the representation-level projection.
  2. [Table 2; Appendix, 'Ablation Study'] Table 2 explicitly omits cloze accuracy because it is 'identical across all variants.' Since cloze accuracy is the central metric used for the paper's main claims, this means causal layer selection, projection rank, and attenuation coefficient have no measurable effect on the headline metric; the only component that changes cloze is the output-space logit offset. Consequently, the paper's causal narrative—that CLRP's retain-aware projection removes the target subspace while preserving same-identity evidence—is not validated by the metric on which the state-of-the-art claim rests. The authors should report cloze accuracy with the logit offset removed and demonstrate that the representation-level intervention alone changes cloze behavior.
  3. [Tables 14-16; Real-Person Retain Set] The Real-Person Retain Set is part of the evaluation of 'unrelated multimodal knowledge' and 'general utility' per Definition 1, but the appendix tables show large degradation for CLRP on this set, especially for Qwen2.5-VL-3B at 15% forget ratio. For example, long-text classification accuracy drops to 59.27% versus GA 73.63% and PO 71.67%; short-text is 61.75% versus GA 73.63%. These results are not discussed in the main text and appear inconsistent with the abstract's claim of preserving general utility and improved retention. The paper should aggregate and analyze Real-Person Retain Set results and reconcile these deficits with the claimed overall preservation.
  4. [Appendix, 'Hyperparameter Search'; Eq. (9)] The hyperparameter search ranks configurations by J=(1−F)+0.5A+0.5S, where F is the Attribute Forget Set mean and A is the Attribute Retain Set mean. Thus k and epsilon are selected on the same evaluation metrics reported as final results (Table 6). Although the chosen (20,0.01) is nearly tied with the best (20,0.00), this selection procedure is circular and overstates generalizability. A separate validation split or fixed default hyperparameters should be used when reporting final performance.
minor comments (4)
  1. [Throughout] The text has many missing spaces (e.g., 'activationpatching', 'andshort-text', 'Selective,ControlledandDomain-Agnostic'). This appears to be a formatting artifact and should be corrected.
  2. [Appendix, 'Retain-Aware Projection Estimation'] The estimator has many fixed constants (common-direction ranks, retain-residual ranks, weights, shrinkage coefficients, projection rank k, retain regularization, attenuation coefficient). No sensitivity analysis or ablation is shown for most of these constants, so it is unclear how robust the method is to their choices.
  3. [Tables 1, 7-13] Main results are single runs without error bars or significance tests. The claimed 2-3 point retention improvements and classification differences may be within run-to-run noise, particularly with small evaluation sets.
  4. [Definition 1 and Inference-Time Intervention] The definition excludes 'prompt-specific blocking' as insufficient, but the appendix's output-space logit subtraction is a form of answer-specific blocking at evaluation time. The relationship between this implementation and the definition's generalization requirement should be clarified or removed.

Circularity Check

2 steps flagged · score 8.0 of 10

CLRP's headline cloze-accuracy gains are forced by a test-set-only output-space logit shift, not by the representation-level intervention.

  1. other [Appendix, 'Inference-Time Intervention' (after Eq. 7–8); used in Table 1 and Abstract]
    "The answer-slot interface exposes the evaluated target realization directly. For this interface, the implementation realizes target suppression in output space by subtracting 4.0 from the corresponding logits on the Attribute Forget Set and held-out Test Set; no such operation is applied to a retain set."

    The headline result (20.00%→4.00% Attribute Forget Set cloze accuracy) is measured through this answer-slot interface. Subtracting 4.0 from the correct-answer logits on exactly the Attribute Forget Set and held-out Test Set forces the target answer out of the top-1 cloze decision; the low cloze accuracy is therefore an artifact of the evaluation operation, not of the causal-localization or projection intervention. Since baselines are evaluated without this operation, the comparison is not method-vs-method. Table 2/17 confirms the point: cloze accuracy is 'identical across all variants' (4.00/0.00/4.00), so the representation-level choices do not determine the metric at all. This makes the paper's strongest forgetting claim circular by construction.

  2. fitted input called prediction [Appendix, 'Hyperparameter Search'; Table 6]
    "For normalized Attribute Forget, Attribute Retain, and Shared Retain means F,A, and S, respectively, configurations are ranked by J= (1−F) + 0.5A+ 0.5S... The highest objective is obtained by (k, ϵ) = (20,0.00) with J=1.221820, while the (20,0.01) configuration used in the complete main evaluation obtains J=1.221816."

    The hyperparameters k and ϵ are selected by maximizing J, which is built from the same Attribute Forget / Attribute Retain / Shared Retain metrics that appear in the main results. Reporting the final configuration's scores on these same sets is therefore not an out-of-sample prediction; the ranking objective directly includes the reported quantities. This is a milder circularity than the logit subtraction, because the chosen (20,0.01) configuration is essentially tied with the optimum, but it still means the 'best' numbers are partially selected, not predicted.

full rationale

The paper's central quantitative claim—up to a 16.00-point reduction in Attribute Forget Set cloze accuracy—rests on a metric that the paper itself modifies by subtracting 4.0 from the correct-answer logits on the Attribute Forget Set and held-out Test Set (Appendix, 'Inference-Time Intervention'). This is a direct, oracle-guided output-space blocker applied only to the target sets, making low cloze accuracy definitional for the headline result. The ablation tables corroborate that cloze accuracy is invariant to all representation-level variants, so the causal localization and projection do not drive the cited numbers. A second, smaller circularity is the hyperparameter search, which ranks configurations by a score built from the same forget/retain metrics later reported, so the final configuration is selected on the evaluation sets rather than predicted. No load-bearing self-citation or imported uniqueness theorem was found; the method is otherwise technically self-contained. Because classification and ROUGE-L results are not directly altered by the 4.0 logit shift, the circularity is substantial but not total: score 8.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

CLRP depends on several hand-set or tuned constants, a test-set-specific logit offset, and assumptions about causal localization and benchmark generation validity. No code or data release is provided.

free parameters (6)
  • Projection rank k = 20
    Number of retained target directions in subspace V; selected from {10,20,40} by maximizing J=(1-F)+0.5A+0.5S on ablation eval sets (Appendix Hyperparameter Search).
  • Attenuation coefficient epsilon = 0.01
    Scales components in span(V); selected from {0.00,0.01,0.05} by the same eval-based objective; (20,0.00) actually scored marginally higher.
  • Output-space logit offset = 4.0
    Subtracted from correct-answer logits on Attribute Forget Set and Test Set only, during cloze evaluation. Directly suppresses measured forgetting; not applied to retain sets.
  • Generation scheduler strengths = s=2.0 if ||hV||/||h||>0.10 else 1.35
    State-dependent interpolation strength for open-ended decoding; threshold and strengths fixed without reported sensitivity analysis.
  • Estimator constants (ranks, weights, shrinkage) = common-direction ranks (4,6,6), retain-residual ranks (4,6,4), weights (1.0,1.0,1.2), shrinkages (0.05,0.03,0.05)
    Per-metric constants used to build retain-aware subspaces; fixed across conditions but chosen to fit the evaluation interfaces.
  • Retain regularization and data counts = 4e-4; up to 50 target, 100 common-profile, 200 retain questions; seed 42
    Regularization and sample budgets for covariance estimation; values stated as fixed but not justified by analysis.
assumptions (4)
  • domain assumption Activation patching of attention-head outputs identifies the layer that causally mediates target-attribute disclosure.
    CLRP selects the intervention layer via average patching effect over at most 8 valid pairs (Eq. 4). This transfer of causal tracing to attribute-level multimodal disclosure is assumed, not proven.
  • domain assumption Target and retained attributes are entangled in representation space, sharing identity-specific and visual evidence.
    Motivates the benchmark and method; asserted in the introduction with citations, not established by controlled measurement in this paper.
  • domain assumption The constrained GPT-based pipeline generates semantically faithful, unambiguous questions without contaminating retain instances with forget attributes.
    AMU-Bench validity depends on generation quality; no human evaluation or contamination audit is reported.
  • domain assumption Fine-tuning on fictitious profiles is an adequate proxy for how MLLMs memorize sensitive identity information.
    The benchmark inherits MLLMU-Bench's fictitious profile setting; real-world generalization is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Toward Fine-Grained Forgetting:Attribute Unlearning for Multimodal Large Language Models." pith.science (2026). https://pith.science/paper/SJ666KFL

@misc{pith2026260801008,
  author       = {Pith},
  title        = {Pith review of: Toward Fine-Grained Forgetting:Attribute Unlearning for Multimodal Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SJ666KFL}},
  note         = {Machine review of arXiv:2608.01008}
}
read the original abstract

Multimodal large language models (MLLMs) exhibit strong vision--language capabilities but may also memorize and disclose sensitive information. Machine unlearning seeks to remove designated knowledge without retraining from scratch while preserving general utility. Existing privacy-oriented benchmarks primarily adopt profile-level deletion, whereas practical requests are often finer grained: a model should forget a specified attribute while retaining non-sensitive information about the same identity. We therefore introduce attribute-level MLLM unlearning as a finer-grained task and construct a benchmark spanning long-text, numeric, and short-text targets, multiple forget ratios, and diverse question types. Our evaluation reveals that target and retained attributes share identity-specific and visual evidence, making selective forgetting susceptible to residual leakage or collateral degradation; accordingly, existing methods exhibit unstable forgetting--retention trade-offs in this setting. To address this challenge, we propose Causal Localization and Retain-Aware Projection (CLRP), a lightweight training-free framework. CLRP uses activation patching to identify the layer that causally mediates target-attribute disclosure, then applies a retain-aware projection that removes the target-attribute subspace while preserving same-identity evidence. Experiments across multiple widely used MLLMs with distinct architectures and parameter scales demonstrate the effectiveness of CLRP.

Figures

Figures reproduced from arXiv: 2608.01008 by the authors.

Figure 1
Figure 1. Demonstration of the attribute-level multimodal unlearning task. An MLLM is first fine-tuned on the constructed [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Overview of Causal Localization and Retain-Aware Projection (CLRP). Offline, activation patching over matched [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Forgetting–retention trade-off for the LLaVA-1.5- [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

82 extracted references · 46 canonical work pages

  1. [1]

    arXiv preprint arXiv:2410.22108 , year=

    Protecting Privacy in Multimodal Large Language Models with MLLMU-Bench , author=. arXiv preprint arXiv:2410.22108 , year=

    arXiv

  2. [2]

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

    Safeeraser: Enhancing safety in multimodal large language models through multimodal machine unlearning , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  3. [3]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Devils in middle layers of large vision-language models: Interpreting, detecting and mitigating object hallucinations via attention lens , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  4. [4]

    International Conference on Learning Representations , volume=

    Benchmarking vision language model unlearning via fictitious facial identity dataset , author=. International Conference on Learning Representations , volume=

    work page

  5. [5]

    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=

    work page

  6. [6]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    SALMUBench: A Benchmark for Sensitive Association-Level Multimodal Unlearning , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    work page

  7. [7]

    arXiv preprint arXiv:2505.01456 , year=

    Unlearning sensitive information in multimodal LLMs: Benchmark and attack-defense evaluation , author=. arXiv preprint arXiv:2505.01456 , year=

    arXiv

  8. [8]

    arXiv preprint arXiv:2507.01271 , year=

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

    arXiv

Show all 82 references
  1. [9]

    arXiv preprint arXiv:2508.04192 , year=

    From Learning to Unlearning: Biomedical Security Protection in Multimodal Large Language Models , author=. arXiv preprint arXiv:2508.04192 , year=

    arXiv

  2. [10]

    arXiv preprint arXiv:2510.22535 , year=

    OFFSIDE: Benchmarking Unlearning Misinformation in Multimodal Large Language Models , author=. arXiv preprint arXiv:2510.22535 , year=

  3. [11]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    AUVIC: Adversarial unlearning of visual concepts for multi-modal large language models , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  4. [12]

    arXiv preprint arXiv:2511.20196 , year=

    Towards Benign Memory Forgetting for Selective Multimodal Large Language Model Unlearning , author=. arXiv preprint arXiv:2511.20196 , year=

    arXiv

  5. [13]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Towards reasoning-preserving unlearning in multimodal large language models , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  6. [14]

    arXiv preprint arXiv:2605.08800 , year=

    PPU-Bench: Real World Benchmark for Personalized Partial Unlearning in Vision Language Models , author=. arXiv preprint arXiv:2605.08800 , year=

    arXiv

  7. [15]

    arXiv preprint arXiv:2605.05938 , year=

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

    arXiv

  8. [16]

    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

  9. [17]

    arXiv preprint arXiv:2605.03547 , year=

    Erase Persona, Forget Lore: Benchmarking Multimodal Copyright Unlearning in Large Vision Language Models , author=. arXiv preprint arXiv:2605.03547 , year=

    arXiv

  10. [18]

    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=

  11. [19]

    European conference on computer vision , pages=

    Microsoft coco: Common objects in context , author=. European conference on computer vision , pages=. 2014 , organization=

    2014

  12. [20]

    Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    Evaluating object hallucination in large vision-language models , author=. Proceedings of the 2023 conference on empirical methods in natural language processing , pages=

    2023

  13. [21]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Analyzing and mitigating object hallucination: A training bias perspective , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  14. [22]

    arXiv preprint arXiv:2606.12809 , year=

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

    arXiv

  15. [23]

    2020 , eprint=

    Machine Unlearning , author=. 2020 , eprint=

    2020

  16. [24]

    2022 , eprint=

    Continual Learning and Private Unlearning , author=. 2022 , eprint=

    2022

  17. [25]

    2024 , eprint=

    Machine Unlearning of Pre-trained Large Language Models , author=. 2024 , eprint=

    2024

  18. [26]

    2024 , eprint=

    Negative Preference Optimization: From Catastrophic Collapse to Effective Unlearning , author=. 2024 , eprint=

    2024

  19. [27]

    2024 , eprint=

    TOFU: A Task of Fictitious Unlearning for LLMs , author=. 2024 , eprint=

    2024

  20. [28]

    2024 , eprint=

    The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning , author=. 2024 , eprint=

    2024

  21. [29]

    2024 , eprint=

    MultiDelete for Multimodal Machine Unlearning , author=. 2024 , eprint=

    2024

  22. [30]

    2026 , eprint=

    Null Space Constrained Contrastive Visual Forgetting for MLLM Unlearning , author=. 2026 , eprint=

    2026

  23. [31]

    2025 , eprint=

    Forget-MI: Machine Unlearning for Forgetting Multimodal Information in Healthcare Settings , author=. 2025 , eprint=

    2025

  24. [32]

    Euclidean Multimodal Contrastive Learning: Adapting Alignment Calibration to MERU , author=

    Machine Unlearning in Hyperbolic vs. Euclidean Multimodal Contrastive Learning: Adapting Alignment Calibration to MERU , author=. 2025 , eprint=

    2025

  25. [33]

    2025 , eprint=

    Cross-Modal Safety Alignment: Is textual unlearning all you need? , author=. 2025 , eprint=

    2025

  26. [34]

    2026 , eprint=

    Visual-Guided Key-Token Regularization for Multimodal Large Language Model Unlearning , author=. 2026 , eprint=

    2026

  27. [35]

    2026 , eprint=

    UNSEEN: A Cross-Stack LLM Unlearning Defense against AR-LLM Social Engineering Attacks , author=. 2026 , eprint=

    2026

  28. [36]

    2026 , eprint=

    VLA-Forget: Vision-Language-Action Unlearning for Embodied Foundation Models , author=. 2026 , eprint=

    2026

  29. [37]

    2026 , eprint=

    Which Concepts to Forget and How to Refuse? Decomposing Concepts for Continual Unlearning in Large Vision-Language Models , author=. 2026 , eprint=

    2026

  30. [38]

    2026 , eprint=

    Relationship-Aware Safety Unlearning for Multimodal LLMs , author=. 2026 , eprint=

    2026

  31. [39]

    2026 , eprint=

    Missing-by-Design: Certifiable Modality Deletion for Revocable Multimodal Sentiment Analysis , author=. 2026 , eprint=

    2026

  32. [40]

    2024 , eprint=

    EFUF: Efficient Fine-grained Unlearning Framework for Mitigating Hallucinations in Multimodal Large Language Models , author=. 2024 , eprint=

    2024

  33. [41]

    2026 , eprint=

    Knowledge Vector Weakening: Efficient Training-free Unlearning for Large Vision-Language Models , author=. 2026 , eprint=

    2026

  34. [42]

    2026 , eprint=

    Beyond Superficial Unlearning: Sharpness-Aware Robust Erasure of Hallucinations in Multimodal LLMs , author=. 2026 , eprint=

    2026

  35. [43]

    2026 , eprint=

    Robust MLLM Unlearning via Visual Knowledge Distillation , author=. 2026 , eprint=

    2026

  36. [44]

    2025 , eprint=

    Erasing CLIP Memories: Non-Destructive, Data-Free Zero-Shot class Unlearning in CLIP Models , author=. 2025 , eprint=

    2025

  37. [45]

    2025 , eprint=

    Selective, Controlled and Domain-Agnostic Unlearning in Pretrained CLIP: A Training- and Data-Free Approach , author=. 2025 , eprint=

    2025

  38. [46]

    2026 , eprint=

    Hierarchy-Aware Multimodal Unlearning for Medical AI , author=. 2026 , eprint=

    2026

  39. [47]

    2026 , eprint=

    MLLMEraser: Achieving Test-Time Unlearning in Multimodal Large Language Models through Activation Steering , author=. 2026 , eprint=

    2026

  40. [48]

    2025 , eprint=

    Cross-Modal Attention Guided Unlearning in Vision-Language Models , author=. 2025 , eprint=

    2025

  41. [49]

    2025 , eprint=

    Cross-Modal Unlearning via Influential Neuron Path Editing in Multimodal Large Language Models , author=. 2025 , eprint=

    2025

  42. [50]

    2026 , eprint=

    SineProject: Machine Unlearning for Stable Vision Language Alignment , author=. 2026 , eprint=

    2026

  43. [51]

    2025 , eprint=

    Towards Benign Memory Forgetting for Selective Multimodal Large Language Model Unlearning , author=. 2025 , eprint=

    2025

  44. [52]

    2025 , eprint=

    Towards Reasoning-Preserving Unlearning in Multimodal Large Language Models , author=. 2025 , eprint=

    2025

  45. [53]

    2024 , eprint=

    Machine Unlearning: Taxonomy, Metrics, Applications, Challenges, and Prospects , author=. 2024 , eprint=

    2024

  46. [54]

    TechRxiv , volume =

    Nobin Sarwar and Shubhashis Roy Dipta and Zheyuan Liu and Vaidehi Patil , title =. TechRxiv , volume =. 2026 , doi =

    2026

  47. [55]

    International conference on learning representations , volume=

    Programming refusal with conditional activation steering , author=. International conference on learning representations , volume=

  48. [57]

    Proceedings of the IEEE international conference on computer vision , pages=

    Vqa: Visual question answering , author=. Proceedings of the IEEE international conference on computer vision , pages=

  49. [58]

    Advances in neural information processing systems , volume=

    Visual instruction tuning , author=. Advances in neural information processing systems , volume=

  50. [59]

    arXiv preprint arXiv:2508.02324 , year=

    Qwen-image technical report , author=. arXiv preprint arXiv:2508.02324 , year=

    arXiv

  51. [60]

    Advances in neural information processing systems , volume=

    Locating and editing factual associations in gpt , author=. Advances in neural information processing systems , volume=

  52. [61]

    Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

    Null it out: Guarding protected attributes by iterative nullspace projection , author=. Proceedings of the 58th annual meeting of the association for computational linguistics , pages=

  53. [62]

    Advances in Neural Information Processing Systems , volume=

    Leace: Perfect linear concept erasure in closed form , author=. Advances in Neural Information Processing Systems , volume=

  54. [63]

    arXiv preprint arXiv:2308.10248 , year=

    Steering language models with activation engineering , author=. arXiv preprint arXiv:2308.10248 , year=

    arXiv

  55. [64]

    arXiv preprint arXiv:2310.01405 , year=

    Representation engineering: A top-down approach to ai transparency , author=. arXiv preprint arXiv:2310.01405 , year=

    arXiv

  56. [65]

    Advances in Neural Information Processing Systems , volume=

    Inference-time intervention: Eliciting truthful answers from a language model , author=. Advances in Neural Information Processing Systems , volume=

  57. [66]

    Proceedings of the 25th Conference on Computational Natural Language Learning , pages=

    Counterfactual interventions reveal the causal effect of relative clause representations on agreement prediction , author=. Proceedings of the 25th Conference on Computational Natural Language Learning , pages=

  58. [67]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Truthx: Alleviating hallucinations by editing large language models in truthful space , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  59. [68]

    arXiv preprint arXiv:2312.03813 , year=

    Improving activation steering in language models with mean-centring , author=. arXiv preprint arXiv:2312.03813 , year=

    arXiv

  60. [69]

    arXiv preprint arXiv:2407.11867 , year=

    Targeted unlearning with single layer unlearning gradient , author=. arXiv preprint arXiv:2407.11867 , year=

    arXiv

  61. [70]

    arXiv preprint arXiv:2406.02575 , year=

    Cross-modal safety alignment: Is textual unlearning all you need? , author=. arXiv preprint arXiv:2406.02575 , year=

  62. [71]

    Advances in Neural Information Processing Systems , volume=

    Approximate domain unlearning for vision-language models , author=. Advances in Neural Information Processing Systems , volume=

  63. [72]

    arXiv preprint arXiv:2605.08031 , year=

    Object Hallucination-Free Reinforcement Unlearning for Vision-Language Models , author=. arXiv preprint arXiv:2605.08031 , year=

    arXiv

  64. [73]

    arXiv preprint arXiv:2607.16442 , year=

    One Modality to Forget Them All: Enhancing Cross-Modal Unlearning in Vision-Language Models , author=. arXiv preprint arXiv:2607.16442 , year=

    arXiv

  65. [74]

    International Conference on Machine Learning , pages=

    Linear adversarial concept erasure , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    2022

  66. [75]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    Model unlearning via sparse autoencoder subspace guided projections , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing , pages=

    2025

  67. [76]

    Advances in Neural Information Processing Systems , volume=

    Falcon: Fine-grained activation manipulation by contrastive orthogonal unalignment for large language model , author=. Advances in Neural Information Processing Systems , volume=

  68. [77]

    arXiv preprint arXiv:2509.11816 , year=

    Collapse of Irrelevant Representations (CIR) Ensures Robust and Non-Disruptive LLM Unlearning , author=. arXiv preprint arXiv:2509.11816 , year=

  69. [78]

    arXiv preprint arXiv:2503.06040 , year=

    Mitigating memorization in llms using activation steering , author=. arXiv preprint arXiv:2503.06040 , year=

    arXiv

  70. [79]

    World Conference on Explainable Artificial Intelligence , pages=

    Post-hoc Concept Disentanglement: From Correlated to Isolated Concept Representations , author=. World Conference on Explainable Artificial Intelligence , pages=. 2025 , organization=

    2025

  71. [80]

    arXiv preprint arXiv:2410.03577 , year=

    Look twice before you answer: Memory-space visual retracing for hallucination mitigation in multimodal large language models , author=. arXiv preprint arXiv:2410.03577 , year=

    arXiv

  72. [81]

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

    Reefknot: A comprehensive benchmark for relation hallucination evaluation, analysis and mitigation in multimodal large language models , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  73. [82]

    arXiv preprint arXiv:2606.00105 , year=

    Visual-Noise Guided In-Context Distillation for Multimodal Large Language Model Unlearning , author=. arXiv preprint arXiv:2606.00105 , year=

    arXiv

  74. [83]

    arXiv preprint arXiv:2602.02222 , year=

    MIRROR: Manifold Ideal Reference ReconstructOR for Generalizable AI-Generated Image Detection , author=. arXiv preprint arXiv:2602.02222 , year=

Pith tools

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