Fei Sha

Jackson Petty, Sjoerd van Steenkiste, Ishita Dasgupta et al. · deepmind

To process novel sentences, language models (LMs) must generalize compositionally -- combine familiar elements in new ways. What aspects of a model's structure promote compositional generalization? Focusing on transformers, we test the hypothesis, motivated by theoretical and empirical work, that deeper transformers generalize more compositionally. Simply adding layers increases the total number of parameters; to address this confound between depth and size, we construct three classes of models which trade off depth for width such that the total number of parameters is kept constant (41M, 134M and 374M parameters). We pretrain all models as LMs and fine-tune them on tasks that test for compositional generalization. We report three main conclusions: (1) after fine-tuning, deeper models generalize more compositionally than shallower models do, but the benefit of additional layers diminishes rapidly; (2) within each family, deeper models show better language modeling performance, but returns are similarly diminishing; (3) the benefits of depth for compositional generalization cannot be attributed solely to better performance on language modeling. Because model latency is approximately linear in the number of layers, these results lead us to the recommendation that, with a given total parameter budget, transformers can be made shallower than is typical without sacrificing performance.

Roberto Molinaro, Samuel Lanthaler, Bogdan Raonić et al.

We present a generative AI algorithm for addressing the pressing task of fast, accurate, and robust statistical computation of three-dimensional turbulent fluid flows. Our algorithm, termed as GenCFD, is based on an end-to-end conditional score-based diffusion model. Through extensive numerical experimentation with a set of challenging fluid flows, we demonstrate that GenCFD provides an accurate approximation of relevant statistical quantities of interest while also efficiently generating high-quality realistic samples of turbulent fluid flows and ensuring excellent spectral resolution. In contrast, ensembles of deterministic ML algorithms, trained to minimize mean square errors, regress to the mean flow. We present rigorous theoretical results uncovering the surprising mechanisms through which diffusion models accurately generate fluid flows. These mechanisms are illustrated with solvable toy models that exhibit the mathematically relevant features of turbulent fluid flows while being amenable to explicit analytical formulae. Our codes are publicly available at https://github.com/camlab-ethz/GenCFD.

Tiwalayo Eisape, MH Tessler, Ishita Dasgupta et al.

A central component of rational behavior is logical inference: the process of determining which conclusions follow from a set of premises. Psychologists have documented several ways in which humans' inferences deviate from the rules of logic. Do language models, which are trained on text generated by humans, replicate such human biases, or are they able to overcome them? Focusing on the case of syllogisms -- inferences from two simple premises -- we show that, within the PaLM2 family of transformer language models, larger models are more logical than smaller ones, and also more logical than humans. At the same time, even the largest models make systematic errors, some of which mirror human reasoning biases: they show sensitivity to the (irrelevant) ordering of the variables in the syllogism, and draw confident but incorrect inferences from particular syllogisms (syllogistic fallacies). Overall, we find that language models often mimic the human biases included in their training data, but are able to overcome them in some cases.

Linlu Qiu, Fei Sha, Kelsey Allen et al.

Artificial intelligence systems based on large language models (LLMs) are increasingly used as agents that interact with users and with the world. To do so successfully, LLMs need to construct internal representations of the world and form probabilistic beliefs about those representations. To provide a user with personalized recommendations, for example, the LLM needs to gradually infer the user's preferences, over the course of multiple interactions. To evaluate whether contemporary LLMs are able to do so, we use the Bayesian inference framework from probability theory, which lays out the optimal way to update an agent's beliefs as it receives new information. We first show that LLMs do not update their beliefs as expected from the Bayesian framework, and that consequently their predictions do not improve as expected as more information becomes available. To address this issue, we teach the LLMs to reason in a Bayesian manner by training them to mimic the predictions of the normative Bayesian model. We find that this approach not only significantly improves the LLM's performance on the particular recommendation task it is trained on, but also enables generalization to other tasks. This suggests that this method teaches the LLM to better approximate Bayesian reasoning. More generally, our results indicate that LLMs can effectively learn reasoning skills from examples and generalize those skills to new domains.

Zhong Yi Wan, Ignacio Lopez-Gomez, Robert Carver et al.

Accurate, actionable climate information at km scales is crucial for robust natural hazard risk assessment and infrastructure planning. Simulating climate at these resolutions remains intractable, forcing reliance on downscaling: either physics-based or statistical methods that transform climate simulations from coarse to impact-relevant resolutions. One major challenge for downscaling is to comprehensively capture the interdependency among climate processes of interest, a prerequisite for representing climate hazards. However, current approaches either lack the desired scalability or are bespoke to specific types of hazards. We introduce GenFocal, a computationally efficient, general-purpose, end-to-end generative framework that gives rise to full probabilistic characterizations of complex climate processes interacting at fine spatiotemporal scales. GenFocal more accurately assesses extreme risk in the current climate than leading approaches, including one used in the US 5th National Climate Assessment. It produces plausible tracks of tropical cyclones, providing accurate statistics of their genesis and evolution, even when they are absent from the corresponding climate simulations. GenFocal also shows compelling results that are consistent with the literature on projecting climate impact on decadal timescales. GenFocal revolutionizes how climate simulations can be efficiently augmented with observations and harnessed to enable future climate impact assessments at the spatiotemporal scales relevant to local and regional communities. We believe this work establishes genAI as an effective paradigm for modeling complex, high-dimensional multivariate statistical correlations that have deterred precise quantification of climate risks associated with hazards such as wildfires, extreme heat, tropical cyclones, and flooding; thereby enabling the evaluation of adaptation strategies.