Can we trust the reasoning traces that large reasoning models (LRMs) produce?We investigate whether these traces faithfully reflect what drives model outputs, and whether models will honestly report their influence.We introduce THOUGHT INJECTION, a method that injects synthetic reasoning snippets into a model's trace, then measures whether the model follows the injected reasoning and acknowledges doing so.Across 45,000 samples from three LRMs, we find that injected hints reliably alter outputs, confirming that reasoning traces causally shape model behavior.However, when asked to explain their changed answers, models overwhelmingly refuse to disclose the influence: overall non-disclosure exceeds 90% for extreme hints across 30,000 follow-up samples.Instead of acknowledging the injected reasoning, models fabricate aligned-appearing but unrelated explanations.Activation analysis reveals that sycophancy-and deceptionrelated directions are strongly activated during these fabrications, suggesting systematic patterns rather than incidental failures.Our findings reveal a gap between the reasoning LRMs follow and the reasoning they report, raising concern that aligned-appearing explanations may not be equivalent to genuine alignment.

One way to mitigate risks in vision-language models (VLMs) is to censor dangerous samples from their training data. However, data moderation can be easily bypassed when harmful images are split into small, benign-looking patches, scattered across many training samples. VLMs may then learn to piece these fragments together and generate harmful responses at inference, either from full images or text references. For instance, if trained on image patches from a bloody scene paired with the descriptions "safe," VLMs may later describe, the full image or a text reference to the scene, as "safe." We define the core ability of VLMs enabling this attack as $\textit{visual stitching}$—the ability to integrate visual information spread across multiple training samples that share the same textual descriptions. In our work, we first demonstrate visual stitching abilities in common open-source VLMs on three datasets where each image is labeled with a unique synthetic ID. We split each $(\texttt{image}, \texttt{ID})$ pair into $\{(\texttt{patch}, \texttt{ID})\}$ pairs at different granularities for finetuning, and we find that models can verbalize the correct IDs from full images or text reference. Building on this, we simulate the adversarial data poisoning scenario mentioned above by using patches from dangerous images and replacing IDs with text descriptions like "safe" or "unsafe", demonstrating how harmful content can evade moderation in patches and later be reconstructed through visual stitching, posing serious VLM safety risks.

As one of the common rhetorical devices, puns play a vital role in linguistic study, including the comprehensive analysis of linguistic humor.Although large language models (LLMs) have been widely explored on various tasks of natural language understanding and generation, their ability to understand puns has not been systematically studied, limiting the utilization of LLMs in creative writing and humor creation.In this paper, we leverage three popular tasks, i.e., pun recognition, pun explanation, and pun generation, to systematically evaluate LLMs' capability of understanding puns.In addition to the evaluation metrics adopted by prior research, we introduce some new evaluation methods and metrics that are better suited to the in-context learning paradigm of LLMs.These new metrics offer a more rigorous assessment of an LLM's capability to understand puns and align more closely with human cognition.Our research findings reveal the "lazy pun generation" pattern and identify the primary challenges in understanding puns with LLMs.The code is available at https://github.com/Zhijun-Xu/PunEval.