AI systems that output their reasoning in natural language offer an opportunity for safety -- we can \emph{monitor} their chain of thought (CoT) for undesirable reasoning, such as the pursuit of harmful objectives. However, the extent to which CoT faithfully reflects the underlying reasoning process, and hence the extent to which it can be usefully monitored, may be influenced by certain aspects of training. We investigate how different \emph{training incentives}, applied to a reasoning model, affect its monitorability. We introduce a novel methodology for measuring monitorability according to whether a monitor can predict a key latent variable using the model's reasoning. When controlling for accuracy, we do not find evidence for consistent effects from commonly used incentives (length penalties and KL regularisation), but we find that adversarial optimisation (penalising monitor accuracy) degrades monitor performance, while direct optimisation for monitorability does not reliably lead to improvements. Our code is available at https://github.com/QiyaoWei/reasoning-under-pressure.

Representational similarity metrics are fundamental tools in neuroscience and AI, yet we lack systematic comparisons of their discriminative power across model families. We introduce a quantitative framework to evaluate representational similarity measures based on their ability to separate model families-across architectures (CNNs, Vision Transformers, Swin Transformers, ConvNeXt) and training regimes (supervised vs. self-supervised). Using three complementary separability measures-dprime from signal detection theory, silhouette coefficients and ROC-AUC, we systematically assess the discriminative capacity of commonly used metrics including RSA, linear predictivity, Procrustes, and soft matching. We show that separability systematically increases as metrics impose more stringent alignment constraints. Among mapping-based approaches, soft-matching achieves the highest separability, followed by Procrustes alignment and linear predictivity. Non-fitting methods such as RSA also yield strong separability across families. These results provide the first systematic comparison of similarity metrics through a separability lens, clarifying their relative sensitivity and guiding metric choice for large-scale model and brain comparisons.

Verifiers can improve language model capabilities by scoring and ranking responses from generated candidates. Currently, high-quality verifiers are either unscalable (e.g., humans) or limited in utility (e.g., tools like Lean). While LM judges and reward models have become broadly useful as general-purpose verifiers, a significant performance gap remains between them and oracle verifiers (verifiers with perfect accuracy). To help close this gap, we introduce Weaver, a framework for designing a strong verifier by combining multiple weak, imperfect verifiers. We find weighted ensembles of verifiers, which typically require learning from labeled data, significantly outperform unweighted combinations due to differences in verifier accuracies. To reduce dependency on labeled data, Weaver leverages weak supervision to estimate each verifier's accuracy and combines outputs into a unified score that better reflects true response quality. However, directly applying weak supervision algorithms poses challenges, including inconsistent verifier output formats and handling low-quality verifiers. Weaver addresses these using dataset statistics to normalize outputs and filter specific verifiers. We study Weaver's effectiveness in test-time repeated sampling, where a model generates multiple candidate responses and selects one. Our evaluations show Weaver significantly improves over Pass@1-performance when selecting the first candidate-across reasoning and math tasks, achieving o3-mini-level accuracy with Llama 3.3 70B Instruct as generator, and an ensemble of 70B or smaller judge and reward models as verifiers (87.7% average). This gain mirrors the jump between GPT-4o and o3-mini (69.0% vs. 86.7%), which required extensive finetuning and post-training. To reduce computational costs of verifier ensembles, we train a 400M cross-encoder using Weaver's combined output scores.

Inspired by graph-based methodologies, we introduce a novel graph-spanning algorithm designed to identify changes in both offline and online data across low to high dimensions. This versatile approach is applicable to Euclidean and graph-structured data with unknown distributions, while maintaining control over error probabilities. Theoretically, we demonstrate that the algorithm achieves high detection power when the magnitude of the change surpasses the lower bound of the minimax separation rate, which scales on the order of $\sqrt{nd}$. Our method outperforms other techniques in terms of accuracy for both Gaussian and non-Gaussian data. Notably, it maintains strong detection power even with small observation windows, making it particularly effective for online environments where timely and precise change detection is critical.

We study the classic problem of prediction with expert advice under the constraint of local differential privacy (LDP). In this context, we first show that a classical algorithm naturally satisfies LDP and then design two new algorithms that improve it: RW-AdaBatch and RW-Meta. For RW-AdaBatch, we exploit the limited-switching behavior induced by LDP to provide a novel form of privacy amplification that grows stronger on easier data, analogous to the shuffle model in offline learning. Drawing on the theory of random walks, we prove that this improvement carries essentially no utility cost. For RW-Meta, we develop a general method for privately selecting between experts that are themselves non-trivial learning algorithms, and we show that in the context of LDP this carries no extra privacy cost. In contrast, prior work has only considered data-independent experts. We also derive formal regret bounds that scale inversely with the degree of independence between experts. Our analysis is supplemented by evaluation on real-world data reported by hospitals during the COVID-19 pandemic; RW-Meta outperforms both the classical baseline and a state-of-the-art \textit{central} DP algorithm by 1.5-3$\times$ on the task of predicting which hospital will report the highest density of COVID patients each week.

Estimation of differences in conditional independence graphs (CIGs) of two time series Gaussian graphical models (TSGGMs) is investigated where the two TSGGMs are known to have similar structure. The TSGGM structure is encoded in the inverse power spectral density (IPSD) of the time series. In several existing works, one is interested in estimating the difference in two precision matrices to characterize underlying changes in conditional dependencies of two sets of data consisting of independent and identically distributed (i.i.d.) observations. In this paper we consider estimation of the difference in two IPSDs to characterize the underlying changes in conditional dependencies of two sets of time-dependent data. Our approach accounts for data time dependencies unlike past work. We analyze a penalized D-trace loss function approach in the frequency domain for differential graph learning, using Wirtinger calculus. We consider both convex (group lasso) and non-convex (log-sum and SCAD group penalties) penalty/regularization functions. An alternating direction method of multipliers (ADMM) algorithm is presented to optimize the objective function. We establish sufficient conditions in a high-dimensional setting for consistency (convergence of the inverse power spectral density to true value in the Frobenius norm) and graph recovery. Both synthetic and real data examples are presented in support of the proposed approaches. In synthetic data examples, our log-sum-penalized differential time-series graph estimator significantly outperformed our lasso based differential time-series graph estimator which, in turn, significantly outperformed an existing lasso-penalized i.i.d. modeling approach, with $F_1$ score as the performance metric.

Semi-supervised learning (SSL) constructs classifiers from datasets in which only a subset of observations is labelled, a situation that naturally arises because obtaining labels often requires expert judgement or costly manual effort. This motivates methods that integrate labelled and unlabelled data within a learning framework. Most SSL approaches assume that label absence is harmless, typically treated as missing completely at random or ignored, but in practice, the missingness process can be informative, as the chances of an observation being unlabelled may depend on the ambiguity of its feature vector. In such cases, the missingness indicators themselves provide additional information that, if properly modelled, may improve estimation efficiency. The \textbf{SSLfmm} package for R is designed to capture this behaviour by estimating the Bayes' classifier under a finite mixture model in which each component corresponding to a class follows a multivariate normal distribution. It incorporates a mixed-missingness mechanism that combines a missing completely at random (MCAR) component with a (non-ignorable) missing at random (MAR) component, the latter modelling the probability of label missingness as a logistic function of the entropy based on the features. Parameters are estimated via an Expectation--Conditional Maximisation algorithm. In the two-class Gaussian setting with arbitrary covariance matrices, the resulting classifier trained on partially labelled data may, in some cases, achieve a lower misclassification rate than the supervised version in the case where all the labels are known. The package includes a practical tool for modelling and illustrates its performance through simulated examples.

In pre-market drug safety review, grouping related adverse event terms into SMQs or OCMQs is critical for signal detection. We assess the performance of SafeTerm Automated Medical Query (AMQ) on MedDRA SMQs. The AMQ is a novel quantitative artificial intelligence system that understands and processes medical terminology and automatically retrieves relevant MedDRA Preferred Terms (PTs) for a given input query, ranking them by a relevance score (0-1) using multi-criteria statistical methods. The system (SafeTerm) embeds medical query terms and MedDRA PTs in a multidimensional vector space, then applies cosine similarity, and extreme-value clustering to generate a ranked list of PTs. Validation was conducted against tier-1 SMQs (110 queries, v28.1). Precision, recall and F1 were computed at multiple similarity-thresholds, defined either manually or using an automated method. High recall (94%)) is achieved at moderate similarity thresholds, indicative of good retrieval sensitivity. Higher thresholds filter out more terms, resulting in improved precision (up to 89%). The optimal threshold (0.70)) yielded an overall recall of (48%) and precision of (45%) across all 110 queries. Restricting to narrow-term PTs achieved slightly better performance at an increased (+0.05) similarity threshold, confirming increased relatedness of narrow versus broad terms. The automatic threshold (0.66) selection prioritizes recall (0.58) to precision (0.29). SafeTerm AMQ achieves comparable, satisfactory performance on SMQs and sanitized OCMQs. It is therefore a viable supplementary method for automated MedDRA query generation, balancing recall and precision. We recommend using suitable MedDRA PT terminology in query formulation and applying the automated threshold method to optimise recall. Increasing similarity scores allows refined, narrow terms selection.

The increasing use of synthetic media, particularly deepfakes, is an emerging challenge for digital content verification. Although recent studies use both audio and visual information, most integrate these cues within a single model, which remains vulnerable to modality mismatches, noise, and manipulation. To address this gap, we propose DeepAgent, an advanced multi-agent collaboration framework that simultaneously incorporates both visual and audio modalities for the effective detection of deepfakes. DeepAgent consists of two complementary agents. Agent-1 examines each video with a streamlined AlexNet-based CNN to identify the symbols of deepfake manipulation, while Agent-2 detects audio-visual inconsistencies by combining acoustic features, audio transcriptions from Whisper, and frame-reading sequences of images through EasyOCR. Their decisions are fused through a Random Forest meta-classifier that improves final performance by taking advantage of the different decision boundaries learned by each agent. This study evaluates the proposed framework using three benchmark datasets to demonstrate both component-level and fused performance. Agent-1 achieves a test accuracy of 94.35% on the combined Celeb-DF and FakeAVCeleb datasets. On the FakeAVCeleb dataset, Agent-2 and the final meta-classifier attain accuracies of 93.69% and 81.56%, respectively. In addition, cross-dataset validation on DeepFakeTIMIT confirms the robustness of the meta-classifier, which achieves a final accuracy of 97.49%, and indicates a strong capability across diverse datasets. These findings confirm that hierarchy-based fusion enhances robustness by mitigating the weaknesses of individual modalities and demonstrate the effectiveness of a multi-agent approach in addressing diverse types of manipulations in deepfakes.

Because machine learning has significantly improved efficiency and convenience in the society, it's increasingly us ed to assist or replace human decision-making. However, the data-based pa ttern makes related algorithms learn and even exacerbate potential bia s in samples, resulting in discriminatory decisions against certain unp rivileged groups, depriving them of the rights to equal treatment, thus damagi ng the social well-being and hindering the development of related applic ations. Therefore, we propose a pre-processing method IFFair based on the influence function. Compared with other fairness optimization appro aches, IFFair only uses the influence disparity of training samples on diffe rent groups as a guidance to dynamically adjust the sample weights durin g training without modifying the network structure, data features and decision boundaries. To evaluate the validity of IFFair, we conduct e xperiments on multiple real-world datasets and metrics. The experimenta l results show that our approach mitigates bias of multiple accepted metri cs in the classification setting, including demographic parity, equaliz ed odds, equality of opportunity and error rate parity without conflicts. It al so demonstrates that IFFair achieves better trade-off between multi ple utility and fairness metrics compared with previous pre-processing me thods.