Aligning LMs to human preferences can be resource-intensive and technically challenging.

Large language models are rapidly transforming social science research by enabling the automation of labor-intensive tasks like data annotation and text analysis. However, LLM outputs vary significantly depending on the implementation choices made by researchers (e.g., model selection or prompting strategy). Such variation can introduce systematic biases and random errors, which propagate to downstream analyses and cause Type I (false positive), Type II (false negative), Type S (wrong sign), or Type M (exaggerated effect) errors. We call this phenomenon where configuration choices lead to incorrect conclusions LLM hacking. We find that intentional LLM hacking is strikingly simple. By replicating 37 data annotation tasks from 21 published social science studies, we show that, with just a handful of prompt paraphrases, virtually anything can be presented as statistically significant. Beyond intentional manipulation, our analysis of 13 million labels from 18 different LLMs across 2361 realistic hypotheses shows that there is also a high risk of accidental LLM hacking, even when following standard research practices. We find incorrect conclusions in approximately 31% of hypotheses for state-of-the-art LLMs, and in half the hypotheses for smaller language models. While higher task performance and stronger general model capabilities reduce LLM hacking risk, even highly accurate models remain susceptible. The risk of LLM hacking decreases as effect sizes increase, indicating the need for more rigorous verification of LLM-based findings near significance thresholds. We analyze 21 mitigation techniques and find that human annotations provide crucial protection against false positives. Common regression estimator correction techniques can restore valid inference but trade off Type I vs. Type II errors. We publish a list of practical recommendations to prevent LLM hacking.

Artificial intelligence experts Tristan Harris and Aza Raskin analyze the technology on'One Nation.' A 14-year-old committed suicide after following the advice of an AI chatbot. Another family is suing the same one -- Character AI -- after it told an autistic 14-year-old to kill his parents. It also exposed an 11-year-old to sexual content. These stories are heavy reminders that young people are especially vulnerable on the internet, but AI isn't the only thing targeting them.

We present a new perspective on how readers integrate context during real-time language comprehension. Our proposals build on surprisal theory, which posits that the processing effort of a linguistic unit (e.g., a word) is an affine function of its in-context information content. We first observe that surprisal is only one out of many potential ways that a contextual predictor can be derived from a language model. Another one is the pointwise mutual information (PMI) between a unit and its context, which turns out to yield the same predictive power as surprisal when controlling for unigram frequency. Moreover, both PMI and surprisal are correlated with frequency. This means that neither PMI nor surprisal contains information about context alone. In response to this, we propose a technique where we project surprisal onto the orthogonal complement of frequency, yielding a new contextual predictor that is uncorrelated with frequency. Our experiments show that the proportion of variance in reading times explained by context is a lot smaller when context is represented by the orthogonalized predictor. From an interpretability standpoint, this indicates that previous studies may have overstated the role that context has in predicting reading times.

When prompting a language model (LM), users often expect the model to adhere to a set of behavioral principles across diverse tasks, such as producing insightful content while avoiding harmful or biased language. Instilling such principles (i.e., a constitution) into a model is resource-intensive, technically challenging, and generally requires human preference labels or examples. We introduce SAMI, an iterative algorithm that finetunes a pretrained language model (without requiring preference labels or demonstrations) to increase the conditional mutual information between constitutions and self-generated responses given queries from a dataset. On single-turn dialogue and summarization, a SAMI-trained mistral-7b outperforms the initial pretrained model, with win rates between 66% and 77%. Strikingly, it also surpasses an instruction-finetuned baseline (mistral-7b-instruct) with win rates between 55% and 57% on single-turn dialogue. SAMI requires a model that writes the principles. To avoid dependence on strong models for writing principles, we align a strong pretrained model (mixtral-8x7b) using constitutions written by a weak instruction-finetuned model (mistral-7b-instruct), achieving a 65% win rate on summarization. Finally, we investigate whether SAMI generalizes to diverse summarization principles (e.g., "summaries should be scientific") and scales to stronger models (llama3-70b), finding that it achieves win rates of up to 68% for learned and 67% for held-out principles compared to the base model. Our results show that a pretrained LM can learn to follow constitutions without using preference labels, demonstrations, or human oversight.

This paper explores the application of Shapley Value Regression in dissecting marketing performance at channel-partner level, complementing channel-level Marketing Mix Modeling (MMM). Utilizing real-world data from the financial services industry, we demonstrate the practicality of Shapley Value Regression in evaluating individual partner contributions. Although structured in-field testing along with cooperative game theory is most accurate, it can often be highly complex and expensive to conduct. Shapley Value Regression is thus a more feasible approach to disentangle the influence of each marketing partner within a marketing channel. We also propose a simple method to derive adjusted coefficients of Shapley Value Regression and compares it with alternative approaches.

Most datasets suffer from partial or complete missing values, which has downstream limitations on the available models on which to test the data and on any statistical inferences that can be made from the data. Several imputation techniques have been designed to replace missing data with stand in values. The various approaches have implications for calculating clinical scores, model building and model testing. The work showcased here offers a novel means for categorical imputation based on item response theory (IRT) and compares it against several methodologies currently used in the machine learning field including k-nearest neighbors (kNN), multiple imputed chained equations (MICE) and Amazon Web Services (AWS) deep learning method, Datawig. Analyses comparing these techniques were performed on three different datasets that represented ordinal, nominal and binary categories. The data were modified so that they also varied on both the proportion of data missing and the systematization of the missing data. Two different assessments of performance were conducted: accuracy in reproducing the missing values, and predictive performance using the imputed data. Results demonstrated that the new method, Item Response Theory for Categorical Imputation (IRTCI), fared quite well compared to currently used methods, outperforming several of them in many conditions. Given the theoretical basis for the new approach, and the unique generation of probabilistic terms for determining category belonging for missing cells, IRTCI offers a viable alternative to current approaches.

The role of verbal and non-verbal cues towards great public speaking has been a topic of exploration for many decades. We identify a commonality across present theories, the element of "variety or heterogeneity" in channels or modes of communication (e.g. resorting to stories, scientific facts, emotional connections, facial expressions etc.) which is essential for effectively communicating information. We use this observation to formalize a novel HEterogeneity Metric, HEM, that quantifies the quality of a talk both in the verbal and non-verbal domain (transcript and facial gestures). We use TED talks as an input repository of public speeches because it consists of speakers from a diverse community besides having a wide outreach. We show that there is an interesting relationship between HEM and the ratings of TED talks given to speakers by viewers. It emphasizes that HEM inherently and successfully represents the quality of a talk based on "variety or heterogeneity". Further, we also discover that HEM successfully captures the prevalent bias in ratings with respect to race and gender, that we call sensitive attributes (because prediction based on these might result in unfair outcome). We incorporate the HEM metric into the loss function of a neural network with the goal to reduce unfairness in rating predictions with respect to race and gender. Our results show that the modified loss function improves fairness in prediction without considerably affecting prediction accuracy of the neural network. Our work ties together a novel metric for public speeches in both verbal and non-verbal domain with the computational power of a neural network to design a fair prediction system for speakers.

Organizations are rapidly deploying artificial intelligence (AI) systems to manage their workers. However, AI has been found at times to be unfair to workers. Unfairness toward workers has been associated with decreased worker effort and increased worker turnover. To avoid such problems, AI systems must be designed to support fairness and redress instances of unfairness. Despite the attention related to AI unfairness, there has not been a theoretical and systematic approach to developing a design agenda. This paper addresses the issue in three ways. First, we introduce the organizational justice theory, three different fairness types (distributive, procedural, interactional), and the frameworks for redressing instances of unfairness (retributive justice, restorative justice). Second, we review the design literature that specifically focuses on issues of AI fairness in organizations. Third, we propose a design agenda for AI fairness in organizations that applies each of the fairness types to organizational scenarios. Then, the paper concludes with implications for future research.