mlsanityData to 14 Oct 2026

Defence against the dark arts of distillation: Unsupervised discovery of subliminal learning

arXiv 2610.13536 ↗First posted 9 Oct 2026stat.ML cs.AI cs.LGcited by 0

On the map: Knowledge distillation → On-policy and self-distillation for LLMs also fits: Knowledge distillation for LLMs

Abstract (arXiv)

When a language model is distilled from a teacher that shares its base parameters, the student can inherit behavioural traits of the teacher---including misalignment---even from data that is semantically unrelated to the trait (Cloud et al., 2026). All existing defences require knowing the trait. We remove this requirement and prove that a subliminally transferred trait can be recovered from the parameter updates alone, without supervision, by exploiting a single asymmetry: the trait is expressed consistently across diverse datasets generated by the teacher, whereas task-specific and shared capabilities vary. The resulting estimator returns a weight-space direction that can be amplified to discover and interpret the trait or projected out to suppress it. Applied to the subliminal-learning benchmarks of (Cloud et al., 2026), our method recovers the transmitted animal-preference traits without any trait labels and suppresses their expression with almost no impact to task performance.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Placed in On-policy and self-distillation for LLMs, one of 12 directions in Knowledge distillation. Builds on 17 of its 25 references that are in the map. First posted on arXiv 9 Oct 2026.

Builds on

17 of 25 references.

Closest ancestors · 1

Same direction, newest first

  1. 2025Towards Understanding Subliminal Learning: When and How Hidden Biases Transfer28 Sep

Other references · 16

  1. 2021LoRA: Low-Rank Adaptation of Large Language Modelscited by 14,380
  2. 2021Training Verifiers to Solve Math Word Problemscited by 10,109
  3. 2018CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledgecited by 1,822
  4. 2017Get To The Point: Summarization with Pointer-Generator Networkscited by 1,292
  5. 2024Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Trainingcited by 529
  6. 2025Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-timecited by 43
  7. 2021Estimating Categorical Counterfactuals via Deep Twin Networkscited by 19
  8. 2026Subliminal Effects in Your Data: A General Mechanism via Log-Linearitycited by 19
  9. 2026Subliminal Steering: Stronger Encoding of Hidden Signalscited by 14
  10. 2026Subliminal Learning Is Steering Vector Distillationcited by 13
  11. 2026From Data to Behavior: Predicting Unintended Model Behaviors Before Trainingcited by 8
  12. 2026Subliminal Learning is a LoRA Artifactcited by 6
4 more
  1. 2026Causal Representation Learning for Generalisable Recommendationcited by 2
  2. 2026Unsupervised Identification and Removal of Spurious Correlations During Fine-Tuningcited by 1
  3. 2026Inference-Time Consensus for Mitigating Hidden Behaviors from LLM Fine-Tuningcited by 1
  4. 2026Incremental Recommendation via Causal Modelscited by 1

Built on by

Not cited yet.

Similar, not cited

Close in wording, no citation either way

  1. 2026On Mitigation of Subliminal Learning in Large Language Models
  2. 2026Cross-Trait Transfer in Subliminal Learning
  3. 2026Subliminal Learning as Trait-Direction Drift: A Mechanism and Targeted Control under SFT Distillation
  4. 2026Beyond Owls: Subliminal Learning Can Transfer Learned Capabilities and Backdoors
  5. 2025Subliminal Learning: Language models transmit behavioral traits via hidden signals in data
  6. 2026You Didn't Have to Say It like That: Subliminal Learning from Faithful Paraphrases
  7. 2026Reproducing and Evaluating the Generalizability of Subliminal Learning in Open-Weight Models
  8. 2026Subliminal Learning is Non-Semantic Distillation