Theory of mind (ToM) evaluations currently focus on testing models using passive narratives that inherently lack interactivity. We introduce FANToM, a new benchmark designed to stress-test ToM within information-asymmetric conversational contexts via question answering. Our benchmark draws upon important theoretical requisites from psychology and necessary empirical considerations when evaluating large language models (LLMs). In particular, we formulate multiple types of questions that demand the same underlying reasoning to identify illusory or false sense of ToM capabilities in LLMs. We show that FANToM is challenging for state-of-the-art LLMs, which perform significantly worse than humans even with chain-of-thought reasoning or fine-tuning.
Humans are social beings; we pursue social goals in our daily interactions, which is a crucial aspect of social intelligence. Yet, AI systems' abilities in this realm remain elusive. We present SOTOPIA, an open-ended environment to simulate complex social interactions between artificial agents and evaluate their social intelligence. In our environment, agents role-play and interact under a wide variety of scenarios; they coordinate, collaborate, exchange, and compete with each other to achieve complex social goals. We simulate the role-play interaction between LLM-based agents and humans within this task space and evaluate their performance with a holistic evaluation framework called SOTOPIA-Eval. With SOTOPIA, we find significant differences between these models in terms of their social intelligence, and we identify a subset of SOTOPIA scenarios, SOTOPIA-hard, that is generally challenging for all models. We find that on this subset, GPT-4 achieves a significantly lower goal completion rate than humans and struggles to exhibit social commonsense reasoning and strategic communication skills. These findings demonstrate SOTOPIA's promise as a general platform for research on evaluating and improving social intelligence in artificial agents.
Human values are crucial to human decision-making. Value pluralism is the view that multiple correct values may be held in tension with one another (e.g., when considering lying to a friend to protect their feelings, how does one balance honesty with friendship?). As statistical learners, AI systems fit to averages by default, washing out these potentially irreducible value conflicts. To improve AI systems to better reflect value pluralism, the first-order challenge is to explore the extent to which AI systems can model pluralistic human values, rights, and duties as well as their interaction. We introduce ValuePrism, a large-scale dataset of 218k values, rights, and duties connected to 31k human-written situations. ValuePrism's contextualized values are generated by GPT-4 and deemed high-quality by human annotators 91% of the time. We conduct a large-scale study with annotators across diverse social and demographic backgrounds to try to understand whose values are represented. With ValuePrism, we build Kaleido, an open, light-weight, and structured language-based multi-task model that generates, explains, and assesses the relevance and valence (i.e., support or oppose) of human values, rights, and duties within a specific context. Humans prefer the sets of values output by our system over the teacher GPT-4, finding them more accurate and with broader coverage. In addition, we demonstrate that Kaleido can help explain variability in human decision-making by outputting contrasting values. Finally, we show that Kaleido's representations transfer to other philosophical frameworks and datasets, confirming the benefit of an explicit, modular, and interpretable approach to value pluralism. We hope that our work will serve as a step to making more explicit the implicit values behind human decision-making and to steering AI systems to make decisions that are more in accordance with them.
Warning: This paper contains content that may be offensive or upsetting. Understanding the harms and offensiveness of statements requires reasoning about the social and situational context in which statements are made. For example, the utterance "your English is very good" may implicitly signal an insult when uttered by a white man to a non-white colleague, but uttered by an ESL teacher to their student would be interpreted as a genuine compliment. Such contextual factors have been largely ignored by previous approaches to toxic language detection. We introduce COBRA frames, the first context-aware formalism for explaining the intents, reactions, and harms of offensive or biased statements grounded in their social and situational context. We create COBRACORPUS, a dataset of 33k potentially offensive statements paired with machine-generated contexts and free-text explanations of offensiveness, implied biases, speaker intents, and listener reactions. To study the contextual dynamics of offensiveness, we train models to generate COBRA explanations, with and without access to the context. We find that explanations by context-agnostic models are significantly worse than by context-aware ones, especially in situations where the context inverts the statement's offensiveness (29% accuracy drop). Our work highlights the importance and feasibility of contextualized NLP by modeling social factors.
Design biases in NLP systems, such as performance differences for different populations, often stem from their creator's positionality, i.e., views and lived experiences shaped by identity and background. Despite the prevalence and risks of design biases, they are hard to quantify because researcher, system, and dataset positionality is often unobserved. We introduce NLPositionality, a framework for characterizing design biases and quantifying the positionality of NLP datasets and models. Our framework continuously collects annotations from a diverse pool of volunteer participants on LabintheWild, and statistically quantifies alignment with dataset labels and model predictions. We apply NLPositionality to existing datasets and models for two tasks -- social acceptability and hate speech detection. To date, we have collected 16,299 annotations in over a year for 600 instances from 1,096 annotators across 87 countries. We find that datasets and models align predominantly with Western, White, college-educated, and younger populations. Additionally, certain groups, such as non-binary people and non-native English speakers, are further marginalized by datasets and models as they rank least in alignment across all tasks. Finally, we draw from prior literature to discuss how researchers can examine their own positionality and that of their datasets and models, opening the door for more inclusive NLP systems.
Dogwhistles are coded expressions that simultaneously convey one meaning to a broad audience and a second one, often hateful or provocative, to a narrow in-group; they are deployed to evade both political repercussions and algorithmic content moderation. For example, in the sentence 'we need to end the cosmopolitan experiment,' the word 'cosmopolitan' likely means 'worldly' to many, but secretly means 'Jewish' to a select few. We present the first large-scale computational investigation of dogwhistles. We develop a typology of dogwhistles, curate the largest-to-date glossary of over 300 dogwhistles with rich contextual information and examples, and analyze their usage in historical U.S. politicians' speeches. We then assess whether a large language model (GPT-3) can identify dogwhistles and their meanings, and find that GPT-3's performance varies widely across types of dogwhistles and targeted groups. Finally, we show that harmful content containing dogwhistles avoids toxicity detection, highlighting online risks of such coded language. This work sheds light on the theoretical and applied importance of dogwhistles in both NLP and computational social science, and provides resources for future research in modeling dogwhistles and mitigating their online harms.
The escalating debate on AI's capabilities warrants developing reliable metrics to assess machine "intelligence". Recently, many anecdotal examples were used to suggest that newer large language models (LLMs) like ChatGPT and GPT-4 exhibit Neural Theory-of-Mind (N-ToM); however, prior work reached conflicting conclusions regarding those abilities. We investigate the extent of LLMs' N-ToM through an extensive evaluation on 6 tasks and find that while LLMs exhibit certain N-ToM abilities, this behavior is far from being robust. We further examine the factors impacting performance on N-ToM tasks and discover that LLMs struggle with adversarial examples, indicating reliance on shallow heuristics rather than robust ToM abilities. We caution against drawing conclusions from anecdotal examples, limited benchmark testing, and using human-designed psychological tests to evaluate models.
Most existing stylistic text rewriting methods operate on a sentence level, but ignoring the broader context of the text can lead to generic, ambiguous, and incoherent rewrites. In this paper, we propose the integration of preceding textual context into both the rewriting and evaluation stages of stylistic text rewriting, focusing on formality, toxicity, and sentiment transfer tasks. We conduct a comparative evaluation of rewriting through few-shot prompting of GPT-3.5 and GPT NeoX, comparing non-contextual rewrites to contextual rewrites. Our experiments show that humans often prefer contextual rewrites over non-contextual ones, but automatic metrics (e.g., BLEU, sBERT) do not. To bridge this gap, we propose context-infused versions of common automatic metrics, and show that these better reflect human preferences. Overall, our paper highlights the importance of integrating preceding textual context into both the rewriting and evaluation stages of stylistic text rewriting.
Improving language model generations according to some user-defined quality or style constraints is challenging. Typical approaches include learning on additional human-written data, filtering ``low-quality'' data using heuristics and/or using reinforcement learning with human feedback (RLHF). However, filtering can remove valuable training signals, whereas data collection and RLHF constantly require additional human-written or LM exploration data which can be costly to obtain. A natural question to ask is ``Can we leverage RL to optimize LM utility on existing crowd-sourced and internet data?'' To this end, we present Left-over Lunch RL (LoL-RL), a simple training algorithm that uses offline policy gradients for learning language generation tasks as a 1-step RL game. LoL-RL can finetune LMs to optimize arbitrary classifier-based or human-defined utility functions on any sequence-to-sequence data. Experiments with five different language generation tasks using models of varying sizes and multiple rewards show that models trained with LoL-RL can consistently outperform the best supervised learning models. We also release our experimental code. https://github.com/abaheti95/LoL-RL
The most meaningful connections between people are often fostered through expression of shared vulnerability and emotional experiences in personal narratives. We introduce a new task of identifying similarity in personal stories based on empathic resonance, i.e., the extent to which two people empathize with each others' experiences, as opposed to raw semantic or lexical similarity, as has predominantly been studied in NLP. Using insights from social psychology, we craft a framework that operationalizes empathic similarity in terms of three key features of stories: main events, emotional trajectories, and overall morals or takeaways. We create EmpathicStories, a dataset of 1,500 personal stories annotated with our empathic similarity features, and 2,000 pairs of stories annotated with empathic similarity scores. Using our dataset, we fine-tune a model to compute empathic similarity of story pairs, and show that this outperforms semantic similarity models on automated correlation and retrieval metrics. Through a user study with 150 participants, we also assess the effect our model has on retrieving stories that users empathize with, compared to naive semantic similarity-based retrieval, and find that participants empathized significantly more with stories retrieved by our model. Our work has strong implications for the use of empathy-aware models to foster human connection and empathy between people.