We theoretically point out the reward over-optimization stems from the inacuracy at the high-reward tail. Accordingly, we investigate whether rubric-based rewards can help mitigate the issue by: (1) leveraging high-quality off-policy responses, and (2) designing with a focus on differentiating among great and diverse responses.

We develop a temperature annealing decoding strategy: start with a higher-than-standard temperature to encourage exploration and gradually reduce to a lower temperature to ensure the quality of the final output. This simple strategy allows effective exploration for reinforcement fine tuning for LLMs.

Recent findings show that learned embeddings of large language models do not lie in a Euclidean space, questioning the default use of cosine similarity. We adopt a more geometry-aware similarity metric and investigate it with two tasks: (1) learning interpretable features via sparse autoencoders, and (2) efficiently retrieving top-𝑘 likely next tokens.

We prove that the underlying sampling distribution of the best-of-n sampling is essentially the optimal model (distribution) for post-training alignment of large language models. It strikes a balance between enhancing output quality, such as helpfulness and harmlessness, and preserving closeness to the base model. Based on this theoretical finding, we propose a fine-tuning method, BoNBoN, to explicitly approximate the best-of-𝑛 distribution.

We link the score representations of the text-controlled diffusion models to real-world concepts, enabling controlled concept shifts in generated images through direct manipulation of these score representations.

We formulated a formal causal estimand tailored to the causal inference of the text-attribute question, and verified its identifiability under minimal conditions. We provided a computationally efficient estimation of the uncertainty quantification of this causal estimand, supported by theoretical assurances.

We establish the asymptotic validity of the heavy-tailed combination test under a broad class of dependence structures, including both asymptotic independence and dependence. We further show that its power advantage over the Bonferroni's test is pronounced under asymptotic dependence, with the gain increasing as the level of dependence strengthens.

We study the cell type deconvolution problem, which seeks to estimate the proportions of different cell types in a mixed sample from gene expression data. We identify key conditions required for the identifiability of these proportions. To address this problem, we propose MEAD, a comprehensive statistical framework for both estimating and conducting inference on cell type proportions. MEAD further enables the comparison of deconvolved proportions across individuals, accounting for both gene–gene correlations and biological variability.

We introduce the heavy-tailed combination test encompassing the widely used Cauchy combination test and harmonic mean p-value as special cases. Under asymptotic independence, we establish its validity but show that it is equivalent to the Bonferroni test in this theoretical setting. In contrast, under stronger dependence, empirical results suggest that the heavy-tailed combination test can substantially outperform Bonferroni.

We introduced an innovative multiple testing procedure that enhances detection power by adaptively filtering out unlikely candidates of PC nulls, and theoretically established the control of both Family-Wise Error Rate (FWER) and False Discovery Rate (FDR) for this method.