machine unlearning
Making an AI forget specific info or behavior without full retraining. Like removing a memory surgically, keeping the rest intact. Knowledge is tangled, so removal can damage unrelated abilities. Most methods are approximate, not exact. Models leak traces via parameters or tokenization (splitting text), letting data be recovered.
Papers show it's fragile: side channels like saved snapshots or alternative tokenizations reconstruct deleted content. Learning style matters: generalizing models lose more unrelated knowledge than memorizing ones. Efficient methods gate neurons or suppress features, avoiding full rewrites. Removing deceptive behavior needs unlearning the behavior, not the knowledge. So, effective unlearning requires precise targeting and leak awareness.
- Papers
- 17
- Released code
- 0
- First seen
- Oct 2026
- Latest
- Nov 2026
17 papers in the last two months, against 0 in the two before.
Start here
Three papers to read first, in the order they came out- 1AI models can stop drawing something while still recognizing it
UniForget helps unified multimodal models forget target concepts in both image generation and visual recognition better than single-pathway methods.
Nov 2026 - 2AI models can pass deletion checks yet diverge during later training
Unlearning algorithms often fail to recover the same features as retraining, causing divergence under further training.
Nov 2026 - 3Language models' own hallucinated text can reveal which memories to erase
Hallucination traces reveal overfitted memories, enabling targeted unlearning that balances forgetting efficacy and general utility without retraining.
Nov 2026
How fast it is moving
Often studied with
mechanistic interpretabilityThe study of how neural networks compute through internal circuits, features, and causal interventions, rather than just input-output behavior.5 togetherdiffusion modelA type of generative model that learns to reverse a gradual noising process to create data.3 togethersafety alignmentProcess of training models to recognize and refuse harmful requests, ensuring outputs remain safe.2 togetherWho is working on it
The papers
Most central to this idea first, not most recent.
- Research Labcs.LG
AI models can pass deletion checks yet diverge during later training
Unlearning algorithms often fail to recover the same features as retraining, causing divergence under further training.
CISPA Helmholtz Center for Information Security, Saarland University · Nov 2026
- Top Universitycs.AI
Language models' own hallucinated text can reveal which memories to erase
Hallucination traces reveal overfitted memories, enabling targeted unlearning that balances forgetting efficacy and general utility without retraining.
Peking University, University of Technology Sydney · Nov 2026
- Chinese Techcs.CV
AI models can stop drawing something while still recognizing it
UniForget helps unified multimodal models forget target concepts in both image generation and visual recognition better than single-pathway methods.
Institute of Automation, Chinese Academy of Sciences, University of Chinese Academy of Sciences · Nov 2026
- Chinese Techcs.LG
AI models can hide individual forgetting failures behind good averages
SALA provides provable per-sample unlearning guarantees by aligning each sample's output to its member or non-member target distribution.
The Hong Kong Polytechnic University, Yunnan University · Nov 2026
- Research Labcs.CR
Language models’ saved versions can reveal supposedly deleted training text
Approximate unlearning in LLMs leaves recoverable traces that enable reconstruction of unlearned text without data-specific knowledge.
RMIT University, CSIRO · Oct 2026
- Research Labcs.CV
Image generators forget unwanted content more reliably while preserving image quality
ConceptACP uses sparse autoencoders to precisely remove target concepts from diffusion models without inference-time overhead.
Institute of Information Engineering, Chinese Academy of Sciences, Boston University · Oct 2026
- Independentcs.LG
Orthogonal Boundary Unlearning preserves benign skills while removing dangerous ones
Restricting unlearning updates to the orthogonal complement of a benign capability subspace removes hazardous dual-use capabilities without degrading benign performance.
Oct 2026
- Chinese Techcs.AI
Open-MMUnlearning framework unifies MLLM unlearning evaluation across five benchmarks and eight models.
Open-MMUnlearning unifies benchmarks, models, and methods for standardized MLLM unlearning evaluation.
Institute of Automation, Chinese Academy of Sciences, University of the Chinese Academy of Sciences · Oct 2026
- Top Universitycs.CR
Alternative tokenizations can bypass targeted LLM editing and unlearning
Alternative tokenizations of input strings bypass localized modifications in LLMs and recover suppressed knowledge.
National University of Singapore, A*STAR Institute of Advanced Intelligence and Computing, Singapore · Oct 2026
- Academiccs.LG
Machines can unlearn deception itself, not just the knowledge it distorts
PACT unlearns deceptive behaviors in LLMs by training with pressure-aware counterfactual targets, reducing deception to under 3%.
Purdue University · Oct 2026
- Academiccs.LG
Query-time neuron gates can selectively suppress unwanted language-model knowledge
Causal routing for unlearning freezes the base model and gates only neurons expressing the forget-set concept, cutting memory to 14 GiB and beating baselines on recall.
Sapienza University of Rome, Aarhus University · Oct 2026
- Top Universitycs.LG
Generalizing models suffer more collateral damage during machine unlearning
A model's reliance on generalization vs memorization predicts retain damage from unlearning: more generalization causes more damage.
Seoul National University, Hanyang University · Oct 2026
- Big Techcs.CV
Consistency error enables direct unlearning in few-step distilled diffusion models
Using consistency error instead of noise prediction error achieves effective unlearning in few step distilled text to image models.
Amazon AGI, Purdue University · Oct 2026
- Big Techcs.LG
Gradient sealing makes language models resist forbidden fine-tuning
Preemptive unlearning via gradient sealing blocks forbidden capability acquisition during fine-tuning by pushing pre-activations into ReLU's zero-derivative region.
State Key Laboratory of Internet of Things for Smart City, University of Macau, RIKEN Center for Advanced Intelligence Project · Oct 2026
- Academiccs.CV
Image generators’ copying is linked to a few attention heads
Memorization in diffusion models is driven by a few shallow-layer attention heads with abnormally strong semantic control.
Southeast University, Engineering Research Center of Blockchain ApplicationSupervision and Management (Southeast University), Ministry of Education · Nov 2026
- Big Techcs.CV
AI models can reuse layers while retaining image and text abilities
CLIP models tolerate substantial reuse of transformer blocks across depth while preserving transferable capabilities.
Samsung AI Center Warsaw, AGH University of Krakow · Oct 2026
- Academiccs.CR
Continually updated AI models reveal when data was used in training
A black-box attack infers whether and when individual training samples entered or left a continually updated model, beating prior methods by reading confidence patterns across snapshots.
Stevens Institute of Technology, University of Connecticut · Nov 2026
Concepts are extracted from each paper and reused across the corpus, so this page grows on its own as the desk reads.