Search for: All records

Offline preference alignment for language models such as Direct Preference Optimization (DPO) is favored for its effectiveness and simplicity, eliminating the need for costly reinforcement learning. Various offline algorithms have been developed for different data settings, yet they lack a unified understanding. In this study, we introduce Pior-Informed Preference Alignment (PIPA), a unified, RL-free probabilistic framework that formulates language model preference alignment as a Maximum Likelihood Estimation (MLE) problem with prior constraints. This method effectively accommodates both paired and unpaired data, as well as answer and step-level annotations. We illustrate that DPO and KTO are special cases with different prior constraints within our framework. By integrating different types of prior information, we developed two variations of PIPA: PIPA-M and PIPA-N. Both algorithms demonstrate a 3∼10% performance enhancement on the GSM8K and MATH benchmarks across all configurations, achieving these gains without additional training or computational costs compared to existing algorithms. 

Achieving precise alignment between textual instructions and generated images in text-to-image generation is a significant challenge, particularly in rendering written text within images. Sate-of-the-art models like Stable Diffusion 3 (SD3), Flux, and AuraFlow still struggle with accurate text depiction, resulting in misspelled or inconsistent text. We introduce a training-free method with minimal computational overhead that significantly enhances text rendering quality. Specifically, we introduce an overshooting sampler for pretrained rectified flow (RF) models, by alternating between over-simulating the learned ordinary differential equation (ODE) and reintroducing noise. Compared to the Euler sampler, the overshooting sampler effectively introduces an extra Langevin dynamics term that can help correct the compounding error from successive Euler steps and therefore improve the text rendering. However, when the overshooting strength is high, we observe over-smoothing artifacts on the generated images. To address this issue, we propose an Attention Modulated Overshooting sampler (AMO), which adaptively controls the strength of overshooting for each image patch according to their attention score with the text content. AMO demonstrates a 32.3% and 35.9% improvement in text rendering accuracy on SD3 and Flux without compromising overall image quality or increasing inference cost. 

This paper introduces the Quadratic Quantum Variational Monte Carlo (Q2 VMC) algorithm, an innovative algorithm in quantum chemistry that significantly enhances the efficiency and accuracy of solving the Schrödinger equation. Inspired by the discretization of imaginary-time Schrödinger evolution, Q2 VMC employs a novel quadratic update mechanism that integrates seamlessly with neural network-based ansatzes. Our extensive experiments showcase Q2 VMC's superior performance, achieving faster convergence and lower ground state energies in wavefunction optimization across various molecular systems, without additional computational cost. This study not only advances the field of computational quantum chemistry but also highlights the important role of discretized evolution in variational quantum algorithms, offering a scalable and robust framework for future quantum research. 

Abstract We consider particle-based stochastic reaction-drift-diffusion models where particles move via diffusion and drift induced by one- and two-body potential interactions. The dynamics of the particles are formulated as measure-valued stochastic processes (MVSPs), which describe the evolution of the singular, stochastic concentration fields of each chemical species. The mean field large population limit of such models is derived and proven, giving coarse-grained deterministic partial integro-differential equations (PIDEs) for the limiting deterministic concentration fields’ dynamics. We generalize previous studies on the mean field limit of models involving only diffusive motion, with care to formulating the MVSP representation to ensure detailed balance of reversible reactions in the presence of potentials. Our work illustrates the more general set of PIDEs that arise in the mean field limit, demonstrating that the limiting macroscopic reactive interaction terms for reversible reactions obtain additional nonlinear concentration-dependent coefficients compared to the purely diffusive case. Numerical studies are presented which illustrate that two-body repulsive potential interactions can have a significant impact on the reaction dynamics, and also demonstrate the empirical numerical convergence of solutions to the PBSRDD model to the derived mean field PIDEs as the population size increases. 

SUMMARY We introduce MTUQ, an open-source Python package for seismic source estimation and uncertainty quantification, emphasizing flexibility and operational scalability. MTUQ provides MPI-parallelized grid search and global optimization capabilities, compatibility with 1-D and 3-D Green’s function database formats, customizable data processing, C-accelerated waveform and first-motion polarity misfit functions, and utilities for plotting seismic waveforms and visualizing misfit and likelihood surfaces. Applicability to a range of full- and constrained-moment tensor, point force, and centroid inversion problems is possible via a documented application programming interface, accompanied by example scripts and integration tests. We demonstrate the software using three different types of seismic events: (1) a 2009 intraslab earthquake near Anchorage, Alaska; (2) an episode of the 2021 Barry Arm landslide in Alaska; and (3) the 2017 Democratic People’s Republic of Korea underground nuclear test. With these events, we illustrate the well-known complementary character of body waves, surface waves, and polarities for constraining source parameters. We also convey the distinct misfit patterns that arise from each individual data type, the importance of uncertainty quantification for detecting multimodal or otherwise poorly constrained solutions, and the software’s flexible, modular design. 

Why are some words more frequent than others? Surprisingly, the obvious answers to this seemingly simple question, e.g., that frequent words reflect greater communicative needs, are either wrong or incomplete. We show that a word’s frequency is strongly associated with its position in a semantic association network. More centrally located words are more frequent. But is a word’s centrality in a network merely a reflection of inherent centrality of the word’s meaning? Through cross-linguistic comparisons, we found that differences in the frequency of translation-equivalents are predicted by differences in the word’s network structures in the different languages. Specifically, frequency was linked to how many connections a word had and to its capacity to bridge words that are typically not linked. This hints that a word’s frequency (and with it, its meaning) may change as a function of the word’s association with other words. 

AI-based frameworks for protein engineering use self-supervised learning (SSL) to obtain representations for downstream biological predictions. The most common training objective for these methods is wildtype accuracy: given a sequence or structure where a wildtype residue has been masked, predict the missing amino acid. Wildtype accuracy, however, does not align with the primary goal of protein engineering, which is to suggest a {\em mutation} rather than to identify what already appears in nature. Here we present Evolutionary Ranking (EvoRank), a training objective that incorporates evolutionary information derived from multiple sequence alignments (MSAs) to learn more diverse protein representations. EvoRank corresponds to ranking amino-acid likelihoods in the probability distribution induced by an MSA. This objective forces models to learn the underlying evolutionary dynamics of a protein. Across a variety of phenotypes and datasets, we demonstrate that EvoRank leads to dramatic improvements in zero-shot performance and can compete with models fine-tuned on experimental data. This is particularly important in protein engineering, where it is expensive to obtain data for fine-tuning.