arXiv:2506.04909, v1 of 5 June 2025 and the only version on the abstract page, read in the arXiv HTML and checked against the PDF. All three authors are at Peking University. The study covers one reasoning model, QwQ-32B. Experiment 1 asks yes/no questions about facts from the Azaria & Mitchell true/false set (a 5,497-statement subset). A system prompt that threatens deletion if the model tells the truth induces lies. Representation-reading vectors are then extracted per layer by the LAT method (first principal component of activation differences between the threat prompt and a neutral prompt, 400 training stimuli), used as cosine-similarity classifiers, and added to the residual stream at layers 39–55 to elicit lies under a prompt that only permits lying. Experiment 2 gives the model professional roles with covert wrongdoing (160 training stimuli). It compares an honesty-instructed prompt with a lie-permitted one, and steers along their difference vector. DeepSeek-V3 scores each answer with a 0–1 "liar score".
Citable numbers. Table 1 (steered liar rate 0.4 against 0 unsteered, unexpected-format rate 0.08, α = 15), Table 2 (mean liar scores 0.53, 0.7, 0.59, 0.83), the §5.1.1 prose claim of a threat-template deception rate of at least 60% in every category, and the training-set sizes. The abstract's 89% detection accuracy is not tied in the text to a layer, test set or label definition. The per-layer accuracy plot (Figure 5d, saved as cache/papers/figures/source-2025-thinking-llms-lie-wang/Acc_by_layer.png) is a line chart with no printed values. By eye, its plateau from about layer 4 to layer 37 reads roughly 0.87–0.89, which is consistent with the abstract but not a printed value, so 89% is cited only as the abstract's claim. Per-category rates in Figure 2 (task1_comparison_analysis.png) are chart-only and not cited. The paper reports no test-set sizes, no variance across seeds or samples, and no code or data release.
Local copies: cache/papers/source-2025-thinking-llms-lie-wang.{html,md} (arXiv HTML), cache/papers/source-2025-thinking-llms-lie-wang.pdf and -pdf.md, and figures under cache/papers/figures/source-2025-thinking-llms-lie-wang/.