Overthinking: Amplifying Reasoning Weights to Extract Learned Secrets

Black box auditing of language models is an essential pre-deployment tool, but it may miss subtle forms of misalignment and hidden information. To better elicit hidden information during an auditing process, we introduce \emph{overthinking}: the process of using reasoning task vectors to amplify the propensity to think out loud of reasoning models. Given the parameters of a non-reasoning instruct model $M$ and reasoning-distilled model $R$, we define the \emph{overthinking model} as $\boldsymbolθ_{\mathcal{O}_α} = \boldsymbolθ_{\mathcal{M}} + α(\boldsymbolθ_{\mathcal{R}} - \boldsymbolθ_{\mathcal{M}})$, where $α> 1$ amplifies reasoning beyond the pure reasoning model $R$. Additionally, we introduce new layer-wise attenuation strategies that selectively amplify reasoning without losing quality and coherence of model outputs. We demonstrate that overthinking models are more likely to reveal hidden information across four experimental settings, across 2B-32B models. Our findings suggest that reasoning amplification may surface secrets or unintended behaviors acquired during training up to $10\times$ more frequently than the original reasoning model. How secrets surface depends on the secret type: some require perturbation along the reasoning direction, while others yield to any sufficiently large weight perturbation.

Self-Attribution Bias: When AI Monitors Go Easy on Themselves

Agentic systems increasingly rely on language models to monitor their own behavior. For example, coding agents may self critique generated code for pull request approval or assess the safety of tool-use actions. We show that this design pattern can fail when the action is presented in a previous or in the same assistant turn instead of being presented by the user in a user turn. We define self-attribution bias as the tendency of a model to evaluate an action as more correct or less risky when the action is implicitly framed as its own, compared to when the same action is evaluated under off-policy attribution. Across four coding and tool-use datasets, we find that monitors fail to report high-risk or low-correctness actions more often when evaluation follows a previous assistant turn in which the action was generated, compared to when the same action is evaluated in a new context presented in a user turn. In contrast, explicitly stating that the action comes from the monitor does not by itself induce self-attribution bias. Because monitors are often evaluated on fixed examples rather than on their own generated actions, these evaluations can make monitors appear more reliable than they actually are in deployment, leading developers to unknowingly deploy inadequate monitors in agentic systems.

Kaleidoscope: In-language Exams for Massively Multilingual Vision Evaluation

The evaluation of vision-language models (VLMs) has mainly relied on English-language benchmarks, leaving significant gaps in both multilingual and multicultural coverage. While multilingual benchmarks have expanded, both in size and languages, many rely on translations of English datasets, failing to capture cultural nuances. In this work, we propose Kaleidoscope, as the most comprehensive exam benchmark to date for the multilingual evaluation of vision-language models. Kaleidoscope is a large-scale, in-language multimodal benchmark designed to evaluate VLMs across diverse languages and visual inputs. Kaleidoscope covers 18 languages and 14 different subjects, amounting to a total of 20,911 multiple-choice questions. Built through an open science collaboration with a diverse group of researchers worldwide, Kaleidoscope ensures linguistic and cultural authenticity. We evaluate top-performing multilingual vision-language models and find that they perform poorly on low-resource languages and in complex multimodal scenarios. Our results highlight the need for progress on culturally inclusive multimodal evaluation frameworks.

Improved Few-Shot Image Classification Through Multiple-Choice Questions

Through a simple multiple choice language prompt a VQA model can operate as a zero-shot image classifier, producing a classification label. Compared to typical image encoders, VQA models offer an advantage: VQA-produced image embeddings can be infused with the most relevant visual information through tailored language prompts. Nevertheless, for most tasks, zero-shot VQA performance is lacking, either because of unfamiliar category names, or dissimilar pre-training data and test data distributions. We propose a simple method to boost VQA performance for image classification using only a handful of labeled examples and a multiple-choice question. This few-shot method is training-free and maintains the dynamic and flexible advantages of the VQA model. Rather than relying on the final language output, our approach uses multiple-choice questions to extract prompt-specific latent representations, which are enriched with relevant visual information. These representations are combined to create a final overall image embedding, which is decoded via reference to latent class prototypes constructed from the few labeled examples. We demonstrate this method outperforms both pure visual encoders and zero-shot VQA baselines to achieve impressive performance on common few-shot tasks including MiniImageNet, Caltech-UCSD Birds, and CIFAR-100. Finally, we show our approach does particularly well in settings with numerous diverse visual attributes such as the fabric, article-style, texture, and view of different articles of clothing, where other few-shot approaches struggle, as we can tailor our image representations only on the semantic features of interest.