Large language models (LLMs) can pass explicit social bias tests but still harbor implicit biases, similar to humans who endorse egalitarian beliefs yet exhibit subtle biases. Measuring such implicit biases can be a challenge: as LLMs become increasingly proprietary, it may not be possible to access their embeddings and apply existing bias measures; furthermore, implicit biases are primarily a concern if they affect the actual decisions that these systems make. We address both challenges by introducing two new measures of bias: LLM Implicit Bias, a prompt-based method for revealing implicit bias; and LLM Decision Bias, a strategy to detect subtle discrimination in decision-making tasks. Both measures are based on psychological research: LLM Implicit Bias adapts the Implicit Association Test, widely used to study the automatic associations between concepts held in human minds; and LLM Decision Bias operationalizes psychological results indicating that relative evaluations between two candidates, not absolute evaluations assessing each independently, are more diagnostic of implicit biases. Using these measures, we found pervasive stereotype biases mirroring those in society in 8 value-aligned models across 4 social categories (race, gender, religion, health) in 21 stereotypes (such as race and criminality, race and weapons, gender and science, age and negativity). Our prompt-based LLM Implicit Bias measure correlates with existing language model embedding-based bias methods, but better predicts downstream behaviors measured by LLM Decision Bias. These new prompt-based measures draw from psychology's long history of research into measuring stereotype biases based on purely observable behavior; they expose nuanced biases in proprietary value-aligned LLMs that appear unbiased according to standard benchmarks.
相關內容
Despite the growing prevalence of large language model (LLM) architectures, a crucial concern persists regarding their energy and power consumption, which still lags far behind the remarkable energy efficiency of the human brain. Recent strides in spiking language models (LM) and transformer architectures aim to address this concern by harnessing the spiking activity of biological neurons to enhance energy/power efficiency. Doubling down on the principles of model quantization and energy efficiency, this paper proposes the development of a novel binary/ternary (1/1.58-bit) spiking LM architecture. Achieving scalability comparable to a deep spiking LM architecture is facilitated by an efficient knowledge distillation technique, wherein knowledge from a non-spiking full-precision "teacher" model is transferred to an extremely weight quantized spiking "student" LM. Our proposed model represents a significant advancement as the first-of-its-kind 1/1.58-bit spiking LM, and its performance is rigorously evaluated on multiple text classification tasks of the GLUE benchmark.
While numerous works have assessed the generative performance of language models (LMs) on tasks requiring Theory of Mind reasoning, research into the models' internal representation of mental states remains limited. Recent work has used probing to demonstrate that LMs can represent beliefs of themselves and others. However, these claims are accompanied by limited evaluation, making it difficult to assess how mental state representations are affected by model design and training choices. We report an extensive benchmark with various LM types with different model sizes, fine-tuning approaches, and prompt designs to study the robustness of mental state representations and memorisation issues within the probes. Our results show that the quality of models' internal representations of the beliefs of others increases with model size and, more crucially, with fine-tuning. We are the first to study how prompt variations impact probing performance on theory of mind tasks. We demonstrate that models' representations are sensitive to prompt variations, even when such variations should be beneficial. Finally, we complement previous activation editing experiments on Theory of Mind tasks and show that it is possible to improve models' reasoning performance by steering their activations without the need to train any probe.
Large language models show compelling performance on reasoning tasks but they tend to perform much worse in languages other than English. This is unsurprising given that their training data largely consists of English text and instructions. A typical solution is to translate instruction data into all languages of interest, and then train on the resulting multilingual data, which is called translate-training. This approach not only incurs high cost, but also results in poorly translated data due to the non-standard formatting of mathematical chain-of-thought. In this paper, we explore the benefits of question alignment, where we train the model to translate reasoning questions into English by finetuning on X-English parallel question data. In this way we perform targeted, in-domain language alignment which makes best use of English instruction data to unlock the LLMs' multilingual reasoning abilities. Experimental results on LLaMA2-13B show that question alignment leads to consistent improvements over the translate-training approach: an average improvement of 11.3% and 16.1% accuracy across ten languages on the MGSM and MSVAMP multilingual reasoning benchmarks. The project will be available at: //github.com/NJUNLP/QAlign.
Disparities in health or well-being experienced by minority groups can be difficult to study using the traditional exposure-outcome paradigm in causal inference, since potential outcomes in variables such as race or sexual minority status are challenging to interpret. Causal decomposition analysis addresses this gap by positing causal effects on disparities under interventions to other, intervenable exposures that may play a mediating role in the disparity. While invoking weaker assumptions than causal mediation approaches, decomposition analyses are often conducted in observational settings and require uncheckable assumptions that eliminate unmeasured confounders. Leveraging the marginal sensitivity model, we develop a sensitivity analysis for weighted causal decomposition estimators and use the percentile bootstrap to construct valid confidence intervals for causal effects on disparities. We also propose a two-parameter amplification that enhances interpretability and facilitates an intuitive understanding of the plausibility of unmeasured confounders and their effects. We illustrate our framework on a study examining the effect of parental acceptance on disparities in suicidal ideation among sexual minority youth. We find that the effect is small and sensitive to unmeasured confounding, suggesting that further screening studies are needed to identify mitigating interventions in this vulnerable population.
Probabilistic deterministic finite automata (PDFA) are discrete event systems modeling conditional probabilities over languages: Given an already seen sequence of tokens they return the probability of tokens of interest to appear next. These types of models have gained interest in the domain of explainable machine learning, where they are used as surrogate models for neural networks trained as language models. In this work we present an algorithm to distill PDFA from neural networks. Our algorithm is a derivative of the L# algorithm and capable of learning PDFA from a new type of query, in which the algorithm infers conditional probabilities from the probability of the queried string to occur. We show its effectiveness on a recent public dataset by distilling PDFA from a set of trained neural networks.
As large language models (LLMs) become more capable, fine-tuning techniques for aligning with human intent are increasingly important. A key consideration for aligning these models is how to most effectively use human resources, or model resources in the case where LLMs themselves are used as oracles. Reinforcement learning from Human or AI preferences (RLHF/RLAIF) is the most prominent example of such a technique, but is complex and often unstable. Direct Preference Optimization (DPO) has recently been proposed as a simpler and more stable alternative. In this work, we develop an active learning strategy for DPO to make better use of preference labels. We propose a practical acquisition function for prompt/completion pairs based on the predictive entropy of the language model and a measure of certainty of the implicit preference model optimized by DPO. We demonstrate how our approach improves both the rate of learning and final performance of fine-tuning on pairwise preference data.
Uncertainty quantification approaches have been more critical in large language models (LLMs), particularly high-risk applications requiring reliable outputs. However, traditional methods for uncertainty quantification, such as probabilistic models and ensemble techniques, face challenges when applied to the complex and high-dimensional nature of LLM-generated outputs. This study proposes a novel geometric approach to uncertainty quantification using convex hull analysis. The proposed method leverages the spatial properties of response embeddings to measure the dispersion and variability of model outputs. The prompts are categorized into three types, i.e., `easy', `moderate', and `confusing', to generate multiple responses using different LLMs at varying temperature settings. The responses are transformed into high-dimensional embeddings via a BERT model and subsequently projected into a two-dimensional space using Principal Component Analysis (PCA). The Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm is utilized to cluster the embeddings and compute the convex hull for each selected cluster. The experimental results indicate that the uncertainty of the model for LLMs depends on the prompt complexity, the model, and the temperature setting.
Pre-trained language models (PLMs) leverage chains-of-thought (CoT) to simulate human reasoning and inference processes, achieving proficient performance in multi-hop QA. However, a gap persists between PLMs' reasoning abilities and those of humans when tackling complex problems. Psychological studies suggest a vital connection between explicit information in passages and human prior knowledge during reading. Nevertheless, current research has given insufficient attention to linking input passages and PLMs' pre-training-based knowledge from the perspective of human cognition studies. In this study, we introduce a Prompting Explicit and Implicit knowledge (PEI) framework, which uses prompts to connect explicit and implicit knowledge, aligning with human reading process for multi-hop QA. We consider the input passages as explicit knowledge, employing them to elicit implicit knowledge through unified prompt reasoning. Furthermore, our model incorporates type-specific reasoning via prompts, a form of implicit knowledge. Experimental results show that PEI performs comparably to the state-of-the-art on HotpotQA. Ablation studies confirm the efficacy of our model in bridging and integrating explicit and implicit knowledge.
Recent contrastive representation learning methods rely on estimating mutual information (MI) between multiple views of an underlying context. E.g., we can derive multiple views of a given image by applying data augmentation, or we can split a sequence into views comprising the past and future of some step in the sequence. Contrastive lower bounds on MI are easy to optimize, but have a strong underestimation bias when estimating large amounts of MI. We propose decomposing the full MI estimation problem into a sum of smaller estimation problems by splitting one of the views into progressively more informed subviews and by applying the chain rule on MI between the decomposed views. This expression contains a sum of unconditional and conditional MI terms, each measuring modest chunks of the total MI, which facilitates approximation via contrastive bounds. To maximize the sum, we formulate a contrastive lower bound on the conditional MI which can be approximated efficiently. We refer to our general approach as Decomposed Estimation of Mutual Information (DEMI). We show that DEMI can capture a larger amount of MI than standard non-decomposed contrastive bounds in a synthetic setting, and learns better representations in a vision domain and for dialogue generation.
Graph neural networks (GNNs) are a popular class of machine learning models whose major advantage is their ability to incorporate a sparse and discrete dependency structure between data points. Unfortunately, GNNs can only be used when such a graph-structure is available. In practice, however, real-world graphs are often noisy and incomplete or might not be available at all. With this work, we propose to jointly learn the graph structure and the parameters of graph convolutional networks (GCNs) by approximately solving a bilevel program that learns a discrete probability distribution on the edges of the graph. This allows one to apply GCNs not only in scenarios where the given graph is incomplete or corrupted but also in those where a graph is not available. We conduct a series of experiments that analyze the behavior of the proposed method and demonstrate that it outperforms related methods by a significant margin.
小貼士
登錄享
相關主題
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191