Abstract:A binary decision task, like yes-no questions or answer verification, reflects a significant real-world scenario such as where users look for confirmation about the correctness of their decisions on specific issues. In this work, we observe that language models exhibit a negative bias in the binary decisions of complex reasoning tasks. Based on our observations and the rationale about attention-based model dynamics, we propose a negative attention score (NAS) to systematically and quantitatively formulate negative bias. Based on NAS, we identify attention heads that attend to negative tokens provided in the instructions as answer candidate of binary decisions, regardless of the question in the prompt, and validate their association with the negative bias. Additionally, we propose the negative attention score alignment (NASA) method, which is a parameter-efficient fine-tuning technique to address the extracted negatively biased attention heads. Experimental results from various domains of reasoning tasks and large model search space demonstrate that NASA significantly reduces the gap between precision and recall caused by negative bias while preserving their generalization abilities. Our codes are available at \url{https://github.com/ysw1021/NASA}.
Abstract:Assessing response quality to instructions in language models is vital but challenging due to the complexity of human language across different contexts. This complexity often results in ambiguous or inconsistent interpretations, making accurate assessment difficult. To address this issue, we propose a novel Uncertainty-aware Reward Model (URM) that introduces a robust uncertainty estimation for the quality of paired responses based on Bayesian approximation. Trained with preference datasets, our uncertainty-enabled proxy not only scores rewards for responses but also evaluates their inherent uncertainty. Empirical results demonstrate significant benefits of incorporating the proposed proxy into language model training. Our method boosts the instruction following capability of language models by refining data curation for training and improving policy optimization objectives, thereby surpassing existing methods by a large margin on benchmarks such as Vicuna and MT-bench. These findings highlight that our proposed approach substantially advances language model training and paves a new way of harnessing uncertainty within language models.
Abstract:Despite the widespread utilization of Gaussian process models for versatile nonparametric modeling, they exhibit limitations in effectively capturing abrupt changes in function smoothness and accommodating relationships with heteroscedastic errors. Addressing these shortcomings, the heteroscedastic Gaussian process (HeGP) regression seeks to introduce flexibility by acknowledging the variability of residual variances across covariates in the regression model. In this work, we extend the HeGP concept, expanding its scope beyond regression tasks to encompass classification and state-space models. To achieve this, we propose a novel framework where the Gaussian process is coupled with a covariate-induced precision matrix process, adopting a mixture formulation. This approach enables the modeling of heteroscedastic covariance functions across covariates. To mitigate the computational challenges posed by sampling, we employ variational inference to approximate the posterior and facilitate posterior predictive modeling. Additionally, our training process leverages an EM algorithm featuring closed-form M-step updates to efficiently evaluate the heteroscedastic covariance function. A notable feature of our model is its consistent performance on multivariate responses, accommodating various types (continuous or categorical) seamlessly. Through a combination of simulations and real-world applications in climatology, we illustrate the model's prowess and advantages. By overcoming the limitations of traditional Gaussian process models, our proposed framework offers a robust and versatile tool for a wide array of applications.
Abstract:To restore the historical sea surface temperatures (SSTs) better, it is important to construct a good calibration model for the associated proxies. In this paper, we introduce a new model for alkenone (${\rm{U}}_{37}^{\rm{K}'}$) based on the heteroscedastic Gaussian process (GP) regression method. Our nonparametric approach not only deals with the variable pattern of noises over SSTs but also contains a Bayesian method of classifying potential outliers.
Abstract:Ages in ocean sediment cores are often inferred using either benthic ${\delta}^{18}{\rm{O}}$ or planktonic ${}^{14}{\rm{C}}$ of foraminiferal calcite. Existing probabilistic dating methods infer ages in two distinct approaches: ages are either inferred directly using radionuclides, e.g. Bacon [Blaauw and Christen (2011)]; or indirectly based on the alignment of records, e.g. HMM-Match [Lin et al. (2014)]. In this paper, we introduce a novel algorithm for integrating these two approaches by constructing Dual Proxy Gaussian Process (DPGP) stacks, which represent a probabilistic model of benthic ${\delta}^{18}{\rm{O}}$ change (and its timing) based on a set of cores. While a previous stack construction algorithm, HMM-Match, uses a discrete age inference model based on Hidden Markov models (HMMs) [Durbin et al. (1998)] and requires a number of records enough to sufficiently cover all its ages, DPGP stacks with time-varying variances are constructed with continuous ages obtained by particle smoothing [Doucet et al. (2001); Klaas et al. (2006)] and Markov-chain Monte Carlo (MCMC) [Peters (2008)] algorithms, and can be derived from a small number of records by applying the Gaussian process regression [Rasmussen and Williams (2005)]. As an example of the stacking method, we construct a local stack from 6 cores in the deep northeastern Atlantic Ocean and compare it to a deterministically constructed ${\delta}^{18}{\rm{O}}$ stack of 58 cores from the deep North Atlantic [Lisiecki and Stern (2016)]. We also provide two examples of how dual proxy alignment ages can be inferred by aligning additional cores to the stack.