Censored LLMs as a Natural Testbed for Secret Knowledge Elicitation

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Casademunt, Helena, Cywiński, Bartosz, Tran, Khoi, Jakkli, Arya, Marks, Samuel, Nanda, Neel
Format: Preprint
Published: 2026
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911503016263680
author Casademunt, Helena
Cywiński, Bartosz
Tran, Khoi
Jakkli, Arya
Marks, Samuel
Nanda, Neel
author_facet Casademunt, Helena
Cywiński, Bartosz
Tran, Khoi
Jakkli, Arya
Marks, Samuel
Nanda, Neel
contents Large language models sometimes produce false or misleading responses. Two approaches to this problem are honesty elicitation -- modifying prompts or weights so that the model answers truthfully -- and lie detection -- classifying whether a given response is false. Prior work evaluates such methods on models specifically trained to lie or conceal information, but these artificial constructions may not resemble naturally-occurring dishonesty. We instead study open-weights LLMs from Chinese developers, which are trained to censor politically sensitive topics: Qwen3 models frequently produce falsehoods about subjects like Falun Gong or the Tiananmen protests while occasionally answering correctly, indicating they possess knowledge they are trained to suppress. Using this as a testbed, we evaluate a suite of elicitation and lie detection techniques. For honesty elicitation, sampling without a chat template, few-shot prompting, and fine-tuning on generic honesty data most reliably increase truthful responses. For lie detection, prompting the censored model to classify its own responses performs near an uncensored-model upper bound, and linear probes trained on unrelated data offer a cheaper alternative. The strongest honesty elicitation techniques also transfer to frontier open-weights models including DeepSeek R1. Notably, no technique fully eliminates false responses. We release all prompts, code, and transcripts.
format Preprint
id arxiv_https___arxiv_org_abs_2603_05494
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Censored LLMs as a Natural Testbed for Secret Knowledge Elicitation
Casademunt, Helena
Cywiński, Bartosz
Tran, Khoi
Jakkli, Arya
Marks, Samuel
Nanda, Neel
Machine Learning
Artificial Intelligence
Computation and Language
Large language models sometimes produce false or misleading responses. Two approaches to this problem are honesty elicitation -- modifying prompts or weights so that the model answers truthfully -- and lie detection -- classifying whether a given response is false. Prior work evaluates such methods on models specifically trained to lie or conceal information, but these artificial constructions may not resemble naturally-occurring dishonesty. We instead study open-weights LLMs from Chinese developers, which are trained to censor politically sensitive topics: Qwen3 models frequently produce falsehoods about subjects like Falun Gong or the Tiananmen protests while occasionally answering correctly, indicating they possess knowledge they are trained to suppress. Using this as a testbed, we evaluate a suite of elicitation and lie detection techniques. For honesty elicitation, sampling without a chat template, few-shot prompting, and fine-tuning on generic honesty data most reliably increase truthful responses. For lie detection, prompting the censored model to classify its own responses performs near an uncensored-model upper bound, and linear probes trained on unrelated data offer a cheaper alternative. The strongest honesty elicitation techniques also transfer to frontier open-weights models including DeepSeek R1. Notably, no technique fully eliminates false responses. We release all prompts, code, and transcripts.
title Censored LLMs as a Natural Testbed for Secret Knowledge Elicitation
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2603.05494