Chain-of-Thought (CoT) reasoning has significantly advanced state-of-the-art AI capabilities. However, recent studies have shown that CoT reasoning is not always faithful, i.e. CoT reasoning does not always reflect how models arrive at conclusions. So far, most of these studies have focused on unfaithfulness in unnatural contexts where an explicit bias has been introduced. In contrast, we show that unfaithful CoT can occur on realistic prompts with no artificial bias. Our results reveal non-negligible rates of several forms of unfaithful reasoning in frontier models: Sonnet 3.7 (16.3%), DeepSeek R1 (5.3%) and ChatGPT-4o (7.0%) all answer a notable proportion of question pairs unfaithfully. Specifically, we find that models rationalize their implicit biases in answers to binary questions ("implicit post-hoc rationalization"). For example, when separately presented with the questions "Is X bigger than Y?" and "Is Y bigger than X?", models sometimes produce superficially coherent arguments to justify answering Yes to both questions or No to both questions, despite such responses being logically contradictory. We also investigate restoration errors (Dziri et al., 2023), where models make and then silently correct errors in their reasoning, and unfaithful shortcuts, where models use clearly illogical reasoning to simplify solving problems in Putnam questions (a hard benchmark). Our findings raise challenges for AI safety work that relies on monitoring CoT to detect undesired behavior.

In this work, we propose and evaluate the feasibility of a two-stage pipeline to evaluate literary machine translation, in a fine-grained manner, from English to Korean. The results show that our framework provides fine-grained, interpretable metrics suited for literary translation and obtains a higher correlation with human judgment than traditional machine translation metrics. Nonetheless, it still fails to match interhuman agreement, especially in metrics like Korean Honorifics. We also observe that LLMs tend to favor translations generated by other LLMs, and we highlight the necessity of developing more sophisticated evaluation methods to ensure accurate and culturally sensitive machine translation of literary works. Figure 1: The overview of our proposed framework: we evaluate translation of literary works in two stages.

Unimodality constitutes a key property indicating grouping behavior of the data around a single mode of its density. We propose a method that partitions univariate data into unimodal subsets through recursive splitting around valley points of the data density. For valley point detection, we introduce properties of critical points on the convex hull of the empirical cumulative density function (ecdf) plot that provide indications on the existence of density valleys. Next, we apply a unimodal data modeling approach that provides a statistical model for each obtained unimodal subset in the form of a Uniform Mixture Model (UMM). Consequently, a hierarchical statistical model of the initial dataset is obtained in the form of a mixture of UMMs, named as the Unimodal Mixture Model (UDMM). The proposed method is non-parametric, hyperparameter-free, automatically estimates the number of unimodal subsets and provides accurate statistical models as indicated by experimental results on clustering and density estimation tasks.

The Coronavirus Disease 2019 (COVID-19) has had a profound impact on global health and economy, making it crucial to build accurate and interpretable data-driven predictive models for COVID-19 cases to improve public policy making. The extremely large scale of the pandemic and the intrinsically changing transmission characteristics pose a great challenge for effectively predicting COVID-19 cases. To address this challenge, we propose a novel hybrid model in which the interpretability of the Autoregressive model (AR) and the predictive power of the long short-term memory neural networks (LSTM) join forces. The proposed hybrid model is formalized as a neural network with an architecture that connects two composing model blocks, of which the relative contribution is decided data-adaptively in the training procedure. We demonstrate the favorable performance of the hybrid model over its two single composing models as well as other popular predictive models through comprehensive numerical studies on two data sources under multiple evaluation metrics. Specifically, in county-level data of 8 California counties, our hybrid model achieves 4.173% MAPE, outperforming the composing AR (5.629%) and LSTM (4.934%) alone on average. In country-level datasets, our hybrid model outperforms the widely-used predictive models such as AR, LSTM, Support Vector Machines, Gradient Boosting, and Random Forest, in predicting the COVID-19 cases in Japan, Canada, Brazil, Argentina, Singapore, Italy, and the United Kingdom. In addition to the predictive performance, we illustrate the interpretability of our proposed hybrid model using the estimated AR component, which is a key feature that is not shared by most black-box predictive models for COVID-19 cases. Our study provides a new and promising direction for building effective and interpretable data-driven models for COVID-19 cases, which could have significant implications for public health policy making and control of the current COVID-19 and potential future pandemics.

Trucking paid for Scott Spindola to take a road trip down the coast of Spain, climb halfway up Machu Picchu, and sample a Costa Rican beach for two weeks. The 44-year-old from Covina now makes up to $70,000 per year, with overtime, hauling goods from the port of Long Beach. He has full medical coverage and plans to drive until he retires. But in a decade, his big rig may not have any need for him. Carmaking giants and ride-sharing upstarts racing to put autonomous vehicles on the road are dead set on replacing drivers, and that includes truckers.