Censored LLMs as a Natural Testbed for Secret Knowledge Elicitation
Researchers present censored Chinese open-weight language models as a testbed for studying whether models can be prompted to reveal knowledge they otherwise suppress and whether false answers can be detected. Across 90 questions on sensitive subjects, next-token completion without a chat template, few-shot prompting, and fine-tuning on generic honesty data most reliably increased factual responses. The two strongest inference-time methods also improved truthfulness in DeepSeek-R1-0528, Qwen3.5-397B, and MiniMax-M2.5. Asking a model to classify its own answers was an effective lie-detection method, while activation probes trained on unrelated data offered a cheaper alternative. The authors argue that censorship-related falsehoods in existing models provide a more natural research setting than deception deliberately implanted through training.
Haber tarzıBilgilendirici
Primarily reports experimental methods and results, with some evaluative claims about the promise and effectiveness of techniques.






