Machine Unlearning
Scalability and Robustness of Machine Unlearning Under Varying Forget-Set Regimes
Generated automatically from the limitations stated in 4 papers (ICML, CVPR, TheWebConference), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
Existing machine unlearning techniques are only validated in narrow operating windows, typically limited to small forget budgets ($\le 10\%$). As forget sets scale beyond a few hundred samples or exceed 10–20% of the training distribution, retention accuracy drops significantly (e.g., a 14–15% gap relative to retraining from scratch on CIFAR-10), while tiny splits (e.g., 1%) yield negligible unlearning. Furthermore, unlearned representations remain fragile and easily recoverable when exposed to small fractions of the forgotten data during relearning. Consequently, current unlearning algorithms cannot be reliably deployed for large-scale data deletion compliance without catastrophic utility loss.
Why it matters
Enables practitioners to deploy unlearning algorithms with predictable safety and utility bounds for arbitrary deletion request volumes, and establishes standard criteria for true irreversible forgetting beyond toy deletion splits.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Systematic Empirical Benchmark Across Forget Scales: Implement leading unlearning families (gradient ascent/reweighting, weight scrubbers, representation orthogonalization) across a unified sweep of forget fractions (0.1%, 1%, 5%, 10%, 20%, 50%) on standard benchmarks (CIFAR-100, ImageNet subsets), measuring the exact Pareto frontier between forget quality (via membership inference attacks and relearning speed) and remaining-set generalization gap against true retraining.
- 2
Relearning Resilience Auditing: Measure the gradient dynamics and parameter trajectories during fine-tuning on small retention/forget mixtures (1–10% relearning budget) across various unlearning methods to quantify how deeply forgotten features are erased versus superficially masked.
- 3
Subspace-Preserving Dynamic Regularization: Formulate and evaluate projection or gradient-constrained objectives that dynamically scale update penalties as the forget-set cardinality grows, measuring whether retention degradation can be bounded within 3% of retraining across $>20\%$ deletion budgets.
Have a different approach?
Describe how you would tackle this problem and we'll look for papers that already do it. Free; your text stays private.
Why it might fail
If theoretical lower bounds establish that approximate linear/gradient-based unlearning without full data access fundamentally degrades utility proportional to forget-set cardinality, rendering true parity with retraining impossible without storing historical training trajectories.
Evidence
Each paper's own statement of the limitation, verbatim.
- Unlearning Isn't Deletion: Investigating Reversibility of Machine Unlearning in LLMsICML 2026
Irreversible, non-catastrophic forgetting is exceptionally challenging to achieve and failed whenever the model had access to even a 10% fraction of the original forget set during relearning.
- Towards Source-Free Machine UnlearningCVPR 2025
Performance degrades sharply as forget-set size grows: on CIFAR-10, forgetting 20% of data causes a 14-15% accuracy gap versus retraining, so it is only reliable for small forget fractions (≤10%)
- Towards Safe Machine Unlearning: a Paradigm that Mitigates Performance DegradationTheWebConference 2025
Performance on the remaining data still degrades significantly as the size of the forgetting set increases beyond 512 samples.
- Benchmarking Vision Language Model Unlearning via Fictitious Facial Identity DatasetICLR 2025
All evaluated methods trade off utility and forget quality sharply; with 1% forget splits none achieve meaningful forgetting
Nearest existing work
- Towards Unbounded Machine UnlearningNeurIPS 2023
- Remember What You Want to Forget: Algorithms for Machine UnlearningNeurIPS 2021
- What makes unlearning hard and what to do about itNeurIPS 2024
- Representation Unlearning: Forgetting through Information CompressionICML 2026
- Learning to Unlearn for Robust Machine UnlearningECCV 2024
- Descent-to-Delete: Gradient-Based Methods for Machine UnlearningALT 2021
- Countering Relearning with Perception Revising UnlearningACML 2024
- The Unseen Threat: Residual Knowledge in Machine Unlearning under Perturbed SamplesNeurIPS 2025
- CoUn: Empowering Machine Unlearning via Contrastive LearningNeurIPS 2025
- Towards Safe Machine Unlearning: a Paradigm that Mitigates Performance DegradationTheWebConference 2025
- The Utility and Complexity of In- and Out-of-Distribution Machine UnlearningICLR 2025
- Algorithms that Approximate Data Removal: New Results and LimitationsNeurIPS 2022
- Unlearning Isn't Deletion: Investigating Reversibility of Machine Unlearning in LLMsICML 2026
- Unlearning Isn’t Forgetting: Revealing Hidden Leakage in Class Unlearning EvaluationsICML 2026
- ESC: Erasing Space Concept for Knowledge DeletionCVPR 2025
Related open problems
Machine Unlearning
A Standardized Multi-Architecture Testbed of Memorized Concepts, Generation Fingerprints, and Editable Structures in Text-to-Image Models Beyond Stable Diffusion v1.x
Nearly every quantitative claim about text-to-image models — memorization, unlearning success, watermark robustness, deepfake provenance, concept editability — is anchored to a handful of frozen Stable Diffusion v1.x/v2 checkpoints, because those are the only models with established benchmarks, known memorized prompts, and validated concept labels. Newer architectures (DiT-based diffusion, DeepFloyd, autoregressive and GAN generators, video models) are largely untested, so results that appear robust may be artifacts of one 2022 model family. The field cannot tell whether its findings are properties of generative modeling or properties of Stable Diffusion 1.4.
Machine Unlearning
Benchmarking the Adversarial Robustness and Reversibility of LLM Unlearning
Existing LLM unlearning methods are predominantly evaluated on standard, benign queries, giving a false sense of compliance with privacy and copyright demands. Empirical evidence shows that "unlearned" knowledge remains extractable via adversarial jailbreaks, latent-space elicitation, and few-shot relearning. Without systematic evaluation across these extraction vectors, practitioners have no way to verify whether a model has actually eliminated sensitive data or merely applied a superficial suppression mask.
Machine Unlearning
Empirical Robustness and Sensitivity of Machine Unlearning Under Realistic Data and Checkpoint Constraints
Existing machine unlearning algorithms for language models are predominantly evaluated under idealized conditions: complete access to forget sets, clean entity anchors, standard prose, and checkpoints saved immediately before target data exposure. In practice, unlearning requests frequently present partial forget data, domain variations like code, noisy or alias-heavy entity mentions, and checkpoints separated from target exposure by billions or trillions of tokens. Practitioners currently cannot predict whether an unlearning method that succeeds on curated benchmarks like TOFU will retain any efficacy when deployed under these real-world data and provenance constraints.
Machine Unlearning
Machine Unlearning Under Degraded Operational Preconditions
Existing machine unlearning methods rely on strict operational preconditions: access to clean retain datasets ($D_r$), full white-box parameter access, historical pre-training checkpoints, or original pre-unlearned calibration weights. In production environments, intermediate training checkpoints are routinely deleted to save storage, retain data is often inaccessible due to data governance policies, and deployment interfaces may restrict full weight access. Without methods that function in the absence of these preconditions, deployed models cannot legally or practically comply with data deletion requests.