Machine Unlearning
Benchmarking the Adversarial Robustness and Reversibility of LLM Unlearning
Generated automatically from the limitations stated in 6 papers (ICLR, ICML, NAACL), listed under Evidence. It is not a paper, and it does not come from papers submitted to CSPaper.
The problem
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.
Why it matters
Enables researchers and auditors to quantitatively verify whether an unlearning method provides persistent knowledge erasure or fragile behavioral suppression. It establishes the first standardized evaluation protocol for unlearning irreversibility under adversarial and relearning conditions.
Ways to approach it
Prior-work checks are free with an account. Results someone already ran are shown to everyone.
- 1
Unified Extraction Benchmark: Build an evaluation harness that subjects top unlearning methods (e.g., gradient ascent, preference optimization, weight localization) to standardized red-teaming: diverse jailbreak templates, white-box activation probing, and low-budget relearning (varying sample count $M \in [1, 50]$ and epochs). Measure knowledge extraction rate and relearning speedup relative to a blank-slate model.
- 2
Relearning-Resistant Objectives: Formulate unlearning objectives that optimize against relearning dynamics—such as regularizing the Fisher information along forget directions or projecting weight updates orthogonal to forget-related parameter submanifolds. Measure post-relearning retention of target knowledge versus standard task utility.
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 concurrent benchmark efforts rapidly standardize unlearning robustness evaluations, diminishing the novelty of purely empirical auditing, or if fundamental information-theoretic bounds prove that robust unlearning in overparameterized networks strictly requires full dataset retraining.
Evidence
Each paper's own statement of the limitation, verbatim.
- DRAGON: Guard LLM Unlearning in Context via Negative Detection and ReasoningICLR 2026
Unlearning is purely behavioral and in-context, meaning underlying memorized weights are not erased and may remain vulnerable to jailbreaks or detection evasion.
- Unlearning as multi-task optimization: A normalized gradient difference approach with an adaptive learning rateNAACL 2025
Complete knowledge removal cannot be guaranteed; unlearned models may still generate harmful or unwanted outputs
- Towards LLM Unlearning Resilient to Relearning Attacks: A Sharpness-Aware Minimization Perspective and BeyondICML 2025
Robustness still degrades as relearning scale grows: with ≥40 forget samples or ≥2 relearning epochs, UE of all smooth variants drops substantially (e.g., NPO+SAM falls from 0.74 to 0.59 at M=3), so unlearning is made harder to reverse, not irreversible
- Targeted Unlearning with Single Layer Unlearning GradientICML 2025
No theoretical guarantee of removal: authors state unlearned identities and concepts may potentially be retrieved via adversarial prompts or relearning attacks, and SLUG lacks robustness against white-box attacks
- Mitigating Memorization in Language ModelsICLR 2025
Even unlearned models can still be prompted to produce impermissible content (Shumailov et al., 2024), so unlearning does not guarantee safety against extraction
- Revisiting Who’s Harry Potter: Towards Targeted Unlearning from a Causal Intervention PerspectiveEMNLP 2024
No theoretical guarantee of unlearning; robustness is only empirically evaluated against two specific jailbreak attacks (Anil et al., 2024; Schwinn et al., 2024), so other attacks may still expose the unlearned knowledge
Nearest existing work
- Unlearning Isn't Deletion: Investigating Reversibility of Machine Unlearning in LLMsICML 2026
- Probing Hidden Knowledge Holes in Unlearned LLMsNeurIPS 2025
- Keeping an Eye on LLM Unlearning: The Hidden Risk and RemedyNeurIPS 2025
- Dual-Space Smoothness for Robust and Balanced LLM UnlearningICLR 2026
- Unlearning or Obfuscating? Jogging the Memory of Unlearned LLMs via Benign RelearningICLR 2025
- Rethinking Adversarial Robustness in the Context of the Right to be ForgottenICML 2024
- Textual Unlearning Gives a False Sense of UnlearningICML 2025
- SEPS: A Separability Measure for Robust Unlearning in LLMsEMNLP 2025
- A Closer Look at Machine Unlearning for Large Language ModelsICLR 2025
- Erase or Hide? Suppressing Spurious Unlearning Neurons for Robust UnlearningICLR 2026
- Machine Unlearning of Pre-trained Large Language ModelsACL 2024
- The Unseen Threat: Residual Knowledge in Machine Unlearning under Perturbed SamplesNeurIPS 2025
- ReLearn: Unlearning via Learning for Large Language ModelsACL 2025
- Verification of Machine Unlearning is FragileICML 2024
- Knowledge Unlearning for Mitigating Privacy Risks in Language ModelsACL 2023
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
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.
Machine Unlearning
Scalability and Robustness of Machine Unlearning Under Varying Forget-Set Regimes
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.