Do language models have beliefs about the world? Dennett (1995) famously argues that even thermostats have beliefs, on the view that a belief is simply an informational state decoupled from any motivational state. In this paper, we discuss approaches to detecting when models have beliefs about the world, and we improve on methods for updating model beliefs to be more truthful, with a focus on methods based on learned optimizers or hypernetworks. Our main contributions include: (1) new metrics for evaluating belief-updating methods that focus on the logical consistency of beliefs, (2) a training objective for Sequential, Local, and Generalizing model updates (SLAG) that improves the performance of learned optimizers, and (3) the introduction of the belief graph, which is a new form of interface with language models that shows the interdependencies between model beliefs. Our experiments suggest that models possess belief-like qualities to only a limited extent, but update methods can both fix incorrect model beliefs and greatly improve their consistency. Although off-the-shelf optimizers are surprisingly strong belief-updating baselines, our learned optimizers can outperform them in more difficult settings than have been considered in past work. Code is available at https://github.com/peterbhase/SLAG-Belief-Updating

We propose a unified multi-tasking framework to represent the complex and uncertain causal process of financial market dynamics, and then to predict the movement of any type of index with an application on the monthly direction of the S&P500 index. our solution is based on three main pillars: (i) the use of transfer learning to share knowledge and feature (representation, learning) between all financial markets, increase the size of the training sample and preserve the stability between training, validation and test sample. (ii) The combination of multidisciplinary knowledge (Financial economics, behavioral finance, market microstructure and portfolio construction theories) to represent a global top-down dynamics of any financial market, through a graph. (iii) The integration of forward looking unstructured data, different types of contexts (long, medium and short term) through latent variables/nodes and then, use a unique VAE network (parameter sharing) to learn simultaneously their distributional representation. We obtain Accuracy, F1-score, and Matthew Correlation of 74.3 %, 67 % and 0.42 above the industry and other benchmark on 12 years test period which include three unstable and difficult sub-period to predict.