Unverified50% confidenceFactExact time
The researchers demonstrated the Synthesis Smuggling threat by training two open-weight models: one infers gender labels from conversational cues, the other infers age-group labels, and they encode the predictions using invisible characters or word choice.
1
Sources
50%
Confidence
Long-term
Relevance
11/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe hidden state steering experiments provide causal evidence that psycholinguistic factors such as commonness and distinctiveness are encoded inside the models, rather than being merely correlational observations.69% similarUnverifiedThe research constructed two kinds of "model organisms" to simulate misaligned behavior: a narrow-trigger model exhibits misbehavior only when the prompt contains a specific cue (for example, choosing the wrong answer in multiple-choice questions by following an implanted hint), while a broad-trigger model has no obvious cue and, after fine-tuning on unsafe code, gives harmful responses across a wide range of open-ended prompts.68% similarUnverifiedWhen the executing agent and the supervising agent originate from the same base model, they share biases and systematic misjudgments acquired during the pre-training phase, which is known as the model homogeneity risk.66% similarUnverifiedIf training data is built upon ambiguous, overlapping symptom labels, machine learning models based on those labels will inherit and even amplify diagnostic uncertainty.65% similarUnverifiedThe research team's core finding is: when training documents describe two behaviors together, the model learns the association between them; thereafter, fine-tuning on just one behavior makes the other appear as well.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1165995API
curl https://kongchang.com/api/v1/knowledge/claims/1165995MCP
get_claim(id=1165995)