mlsanityData to 14 Oct 2026

Inoculation Adapters: Improved Selective Generalization of Capabilities with Fewer Surprising Backdoors

arXiv 2606.30252 ↗First posted 29 Jun 2026cs.AIcited by 3

Not on the map. Why?

Abstract (arXiv)

Inoculation prompting is a selective-generalization technique used against Emergent Misalignment. We introduce inoculation adapters (IA), a family of methods that similarly reduce the optimization pressure to learn undesired traits by strengthening those traits during training. Inoculation adapters are LoRAs that are trained and used in three steps: (1) trained on undesired traits; (2) attached frozen while a separate task adapter is trained on data exhibiting both desired and undesired traits; (3) the IA is discarded at deployment, while only the task adapter is kept. We compare inoculation adapters with four selective-generalization baselines: inoculation prompting, preventative steering, Concept Ablation Fine-Tuning (CAFT), and KL regularization. Across nine setups and five model families, the inoculation adapter family spans a new Pareto frontier of desired trait retention vs. undesired trait suppression, although given wide confidence intervals the magnitude of improvement remains uncertain. Inoculation adapters also avoid two drawbacks of inoculation prompting: they can suppress capabilities and traits that cannot be reliably elicited by a prompt, and they introduce fewer surprising backdoors. However, no IA variant optimizes all objectives perfectly; gains in desired-trait generalization are generally accompanied by weaker suppression of the undesired trait and increased backdoor occurrence.

Key facts. Cited by 3 papers in the map (data through 14 Oct 2026). Builds on 33 of its 48 references that are in the map. First posted on arXiv 29 Jun 2026.

Builds on

33 of 48 references.

References · 33

  1. 2022Locating and Editing Factual Associations in GPTcited by 2,691
  2. 2019Invariant Risk Minimizationcited by 1,854
  3. 2019Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalizationcited by 1,247
  4. 2022Red Teaming Language Models with Language Modelscited by 1,026
  5. 2023Enhancing Chat Language Models by Scaling High-quality Instructional Conversationscited by 839
  6. 2022Discovering Latent Knowledge in Language Models Without Supervisioncited by 767
  7. 2024WildChat: 1M ChatGPT Interaction Logs in the Wildcited by 629
  8. 2021Just Train Twice: Improving Group Robustness without Training Group Informationcited by 523
  9. 2020Underspecification Presents Challenges for Credibility in Modern Machine Learningcited by 496
  10. 2023Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervisioncited by 467
  11. 2025Persona Vectors: Monitoring and Controlling Character Traits in Language Modelscited by 345
  12. 2020Pretrained Transformers Improve Out-of-Distribution Robustnesscited by 294
21 more
  1. 2021GFlowNet Foundationscited by 290
  2. 2023A Rank Stabilization Scaling Factor for Fine-Tuning with LoRAcited by 243
  3. 2023Finding Alignments Between Interpretable Causal Variables and Distributed Neural Representationscited by 209
  4. 2025Natural Emergent Misalignment from Reward Hacking in Production RLcited by 136
  5. 2022Goal Misgeneralization: Why Correct Specifications Aren't Enough For Correct Goalscited by 126
  6. 2021Goal Misgeneralization in Deep Reinforcement Learningcited by 123
  7. 2025Model Organisms for Emergent Misalignmentcited by 91
  8. 2021Evading the Simplicity Bias: Training a Diverse Set of Models Discovers Solutions with Superior OOD Generalizationcited by 83
  9. 2025Auditing language models for hidden objectivescited by 83
  10. 2025School of Reward Hacks: Hacking harmless tasks generalizes to misaligned behavior in LLMscited by 66
  11. 2022Generative Augmented Flow Networkscited by 50
  12. 2023Measuring Inductive Biases of In-Context Learning with Underspecified Demonstrationscited by 49
  13. 2025Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-timecited by 43
  14. 2025Inoculation Prompting: Instructing LLMs to misbehave at train-time improves test-time alignmentcited by 43
  15. 2025Steering Out-of-Distribution Generalization with Concept Ablation Fine-Tuningcited by 39
  16. 2026Conditional misalignment: common interventions can hide emergent misalignment behind contextual triggerscited by 17
  17. 2025Recontextualization Mitigates Specification Gaming without Modifying the Specificationcited by 14
  18. 2025Fundamental Limits of Perfect Concept Erasurecited by 6
  19. 2023Reinforcement Learning Fine-tuning of Language Models is Biased Towards More Extractable Featurescited by 3
  20. 2023Can Active Sampling Reduce Causal Confusion in Offline Reinforcement Learning?cited by 2
  21. 2023CoinRun: Solving Goal Misgeneralisationcited by 1

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 3.

3 papers.

  1. 2026Inducing Emergent Misalignment from Reward Hacks with Iterative DPOcited by 0
  2. 2026Don't Inoculate Everything: Stratified Inoculation Prompting Narrows Backdoor Triggers and Preserves Desired Traitscited by 0
  3. 2026Gradient Routing Localizes and Removes Unintended Behaviors in RLcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2026Subliminal Learning as Trait-Direction Drift: A Mechanism and Targeted Control under SFT Distillation
  2. 2026Emergent and Subliminal Misalignment Through the Lens of Data-Mediated Transfer
  3. 2025Emergent Misalignment via In-Context Learning: Narrow in-context examples can produce broadly misaligned LLMs
  4. 2026Overtrained, Not Misaligned
  5. 2026TAME: Token Attribution and Masking for Emergent misalignment
  6. 2026Inoculation Midtraining with Learned Neologisms
  7. 2026Scaling Model-Generated Distillation Data Can Make Latent Teacher Traits More Recoverable
  8. 2026Defence against the dark arts of distillation: Unsupervised discovery of subliminal learning