Mitigating explicit and implicit biases in Large Language Models (LLMs) has become a critical focus in the field of natural language processing. However, many current methodologies evaluate scenarios in isolation, without considering the broader context or the spectrum of potential biases within each situation. To address this, we introduce the Sensitivity Testing on Offensive Progressions (STOP) dataset, which includes 450 offensive progressions containing 2,700 unique sentences of varying severity that progressively escalate from less to more explicitly offensive. Covering a broad spectrum of 9 demographics and 46 sub-demographics, STOP ensures inclusivity and comprehensive coverage. We evaluate several leading closed- and open-source models, including GPT-4, Mixtral, and Llama 3. Our findings reveal that even the best-performing models detect bias inconsistently, with success rates ranging from 19.3% to 69.8%. Furthermore, we demonstrate how aligning models with human judgments on STOP can improve model answer rates on sensitive tasks such as BBQ, StereoSet, and CrowS-Pairs by up to 191%, while maintaining or even improving performance. STOP presents a novel framework for assessing the complex nature of biases in LLMs, which will enable more effective bias mitigation strategies and facilitates the creation of fairer language models.

This study explores the effectiveness of Large Language Models (LLMs) in creating personalized “mirror stories” that reflect and resonate with individual readers’ identities, addressing the significant lack of diversity in literature. We present MirrorStories, a corpus of 1,500 personalized short stories generated by integrating elements such as name, gender, age, ethnicity, reader interest, and story moral. We demonstrate that LLMs can effectively incorporate diverse identity elements into narratives, with human evaluators identifying personalized elements in the stories with high accuracy. Through a comprehensive evaluation involving 26 diverse human judges, we compare the effectiveness of MirrorStories against generic narratives. We find that personalized LLM-generated stories not only outscore generic human-written and LLM-generated ones across all metrics of engagement (with average ratings of 4.22 versus 3.37 on a 5-point scale), but also achieve higher textual diversity while preserving the intended moral. We also provide analyses that include bias assessments and a study on the potential for integrating images into personalized stories.

The Winograd Schema Challenge (WSC) serves as a prominent benchmark for evaluating machine understanding. While Large Language Models (LLMs) excel at answering WSC questions, their ability to generate such questions remains less explored. In this work, we propose Tree-of-Experts (ToE), a novel prompting method which enhances the generation of WSC instances (50% valid cases vs. 10% in recent methods). Using this approach, we introduce WSC+, a novel dataset comprising 3,026 LLM-generated sentences. Notably, we extend the WSC framework by incorporating new ‘ambiguous’ and ‘offensive’ categories, providing a deeper insight into model overconfidence and bias. Our analysis reveals nuances in generation-evaluation consistency, suggesting that LLMs may not always outperform in evaluating their own generated questions when compared to those crafted by other models. On WSC+, GPT-4, the top-performing LLM, achieves an accuracy of 68.7%, significantly below the human benchmark of 95.1%.

As the use of Large Language Models (LLMs) becomes more widespread, understanding their self-evaluation of confidence in generated responses becomes increasingly important as it is integral to the reliability of the output of these models. We introduce the concept of Confidence-Probability Alignment, that connects an LLM’s internal confidence, quantified by token probabilities, to the confidence conveyed in the model’s response when explicitly asked about its certainty. Using various datasets and prompting techniques that encourage model introspection, we probe the alignment between models’ internal and expressed confidence. These techniques encompass using structured evaluation scales to rate confidence, including answer options when prompting, and eliciting the model’s confidence level for outputs it does not recognize as its own. Notably, among the models analyzed, OpenAI’s GPT-4 showed the strongest confidence-probability alignment, with an average Spearman’s  ̂𝜌 of 0.42, across a wide range of tasks. Our work contributes to the ongoing efforts to facilitate risk assessment in the application of LLMs and to further our understanding of model trustworthiness.

Large Language Models (LLMs) have demonstrated remarkable success in tasks like the Winograd Schema Challenge (WSC), showcasing advanced textual common-sense reasoning. However, applying this reasoning to multimodal domains, where understanding text and images together is essential, remains a substantial challenge. To address this, we introduce WinoVis, a novel dataset specifically designed to probe text-to-image models on pronoun disambiguation within multimodal contexts. Utilizing GPT-4 for prompt generation and Diffusion Attentive Attribution Maps (DAAM) for heatmap analysis, we propose a novel evaluation framework that isolates the models’ ability in pronoun disambiguation from other visual processing challenges. Evaluation of successive model versions reveals that, despite incremental advancements, Stable Diffusion 2.0 achieves a precision of 56.7% on WinoVis, only marginally surpassing random guessing. Further error analysis identifies important areas for future research aimed at advancing text-to-image models in their ability to interpret and interact with the complex visual world.

Research on Large Language Models (LLMs) has often neglected subtle biases that, although less apparent, can significantly influence the models’ outputs toward particular social narratives. This study addresses two such biases within LLMs: representative bias, which denotes a tendency of LLMs to generate outputs that mirror the experiences of certain identity groups, and affinity bias, reflecting the models’ evaluative preferences for specific narratives or viewpoints. We introduce two novel metrics to measure these biases: the Representative Bias Score (RBS) and the Affinity Bias Score (ABS), and present the Creativity-Oriented Generation Suite (CoGS), a collection of open-ended tasks such as short story writing and poetry composition, designed with customized rubrics to detect these subtle biases. Our analysis uncovers marked representative biases in prominent LLMs, with a preference for identities associated with being white, straight, and men. Furthermore, our investigation of affinity bias reveals distinctive evaluative patterns within each model, akin to ‘bias fingerprints’. This trend is also seen in human evaluators, highlighting a complex interplay between human and machine bias perceptions.

Knowledge distillation allows smaller neural networks to emulate the performance of larger, teacher models with reduced computational demands. Traditional methods for Large Language Models (LLMs) often necessitate extensive fine-tuning, which limits their accessibility. To address this, we introduce Trace-of-Thought Prompting, a novel framework designed to distill critical reasoning capabilities from large-scale teacher models (over 8 billion parameters) to small-scale student models (up to 8 billion parameters). This approach leverages problem decomposition to enhance interpretability and facilitate human-in-the-loop interventions. Empirical evaluations on the GSM8K and MATH datasets show that student models achieve accuracy gains of up to 113% on GSM8K and 20% on MATH, with significant improvements particularly notable in smaller models like Llama 2 and Zephyr. Our results suggest a promising pathway for open-source, small-scale models to eventually serve as both students and teachers, potentially reducing our reliance on large-scale, proprietary models. Our code, featuring data analytics and testing scripts, is provided here: https://github.com/traceofthought/trace-of-thought-prompting/tree/main.

While Large Language Models (LLMs) excel at the Winograd Schema Challenge (WSC), a coreference resolution task testing common-sense reasoning through pronoun disambiguation, they struggle with instances that feature minor alterations or rewording. To address this, we introduce EvoGrad, an open-source platform that harnesses a human-in-the-loop approach to create a dynamic dataset tailored to such altered WSC instances. Leveraging ChatGPT’s capabilities, we expand our task instances from 182 to 3691, setting a new benchmark for diverse common-sense reasoning datasets. Additionally, we introduce the error depth metric, assessing model stability in dynamic tasks. Our results emphasize the challenge posed by EvoGrad: Even the best performing LLM, GPT-3.5, achieves an accuracy of 65.0% with an average error depth of 7.2, a stark contrast to human performance of 92.8% accuracy without perturbation errors. This highlights ongoing model limitations and the value of dynamic datasets in uncovering them.

In this paper, we study the viability of the deployment of language models towards non-playable character (NPC) scripts, by introducing a novel pipeline for the automatic construction of NPC scripts using Transformer-based believable scripts for a variety of game genres and specifications. In addition, we propose a self-diagnosis method inspired by previous work to develop language models, tailored specifically to desirable NPC qualities such as coherency, believability, and degree of repetition. Finally, we propose a new benchmark, called The Turing Quest, which we use to show that the pipeline, when applied to GPT-3, can generate for a variety of game genres and contexts, NPC scripts that can fool judges in thinking they have been written by humans. We believe that these findings can greatly benefit both the gaming industry and its global community of users, since many current games continue to base their NPCs on manually-curated scripts that are resource-demanding and may curb the immersiveness and enjoyment of the user.

Debiasing methods that seek to mitigate the tendency of Language Models (LMs) to occasionally output toxic or inappropriate text have recently gained traction. In this paper, we propose a standardized protocol which distinguishes methods that yield not only desirable results, but are also consistent with their mechanisms and specifications. For example, we ask, given a debiasing method that is developed to reduce toxicity in LMs, if the definition of toxicity used by the debiasing method is reversed, would the debiasing results also be reversed? We used such considerations to devise three criteria for our new protocol: Specification Polarity, Specification Importance, and Domain Transferability. As a case study, we apply our protocol to a popular debiasing method, Self-Debiasing, and compare it to one we propose, called Instructive Debiasing, and demonstrate that consistency is as important an aspect to debiasing viability as is simply a desirable result. We show that our protocol provides essential insights into the generalizability and interpretability of debiasing methods that may otherwise go overlooked.

A false contract is more likely to be rejected than a contract is, yet a false key is less likely than a key to open doors. While correctly interpreting and assessing the effects of such adjective-noun pairs (e.g., false key) on the plausibility of given events (e.g., opening doors) underpins many natural language understanding tasks, doing so often requires a significant degree of world knowledge and common-sense reasoning. We introduce ADEPT – a large-scale semantic plausibility task consisting of over 16 thousand sentences that are paired with slightly modified versions obtained by adding an adjective to a noun. Overall, we find that while the task appears easier for human judges (85% accuracy), it proves more difficult for transformer-based models like RoBERTa (71% accuracy). Our experiments also show that neither the adjective itself nor its taxonomic class suffice in determining the correct plausibility judgement, emphasizing the importance of endowing automatic natural language understanding systems with more context sensitivity and common-sense reasoning.

The Winograd Schema Challenge (WSC) and variants inspired by it have become important benchmarks for common-sense reasoning (CSR). Model performance on the WSC has quickly progressed from chance-level to near-human using neural language models trained on massive corpora. In this paper, we analyze the effects of varying degrees of overlaps that occur between these corpora and the test instances in WSC-style tasks. We find that a large number of test instances overlap considerably with the pretraining corpora on which state-of-the-art models are trained, and that a significant drop in classification accuracy occurs when models are evaluated on instances with minimal overlap. Based on these results, we provide the WSC-Web dataset, consisting of over 60k pronoun disambiguation problems scraped from web data, being both the largest corpus to date, and having a significantly lower proportion of overlaps with current pretraining corpora.

We introduce a new benchmark for coreference resolution and NLI, KnowRef, that targets common-sense understanding and world knowledge. Previous coreference resolution tasks can largely be solved by exploiting the number and gender of the antecedents, or have been handcrafted and do not reflect the diversity of naturally occurring text. We present a corpus of over 8,000 annotated text passages with ambiguous pronominal anaphora. These instances are both challenging and realistic. We show that various coreference systems, whether rule-based, feature-rich, or neural, perform significantly worse on the task than humans, who display high inter-annotator agreement. To explain this performance gap, we show empirically that state-of-the art models often fail to capture context, instead relying on the gender or number of candidate antecedents to make a decision. We then use problem-specific insights to propose a data-augmentation trick called antecedent switching to alleviate this tendency in models. Finally, we show that antecedent switching yields promising results on other tasks as well: we use it to achieve state-of-the-art results on the GAP coreference task.

Recent studies have significantly improved the state-of-the-art on common-sense reasoning (CSR) benchmarks like the Winograd Schema Challenge (WSC) and SWAG. The question we ask in this paper is whether improved performance on these benchmarks represents genuine progress towards common-sense-enabled systems. We make case studies of both benchmarks and design protocols that clarify and qualify the results of previous work by analyzing threats to the validity of previous experimental designs. Our protocols account for several properties prevalent in common-sense benchmarks including size limitations, structural regularities, and variable instance difficulty.

We introduce an automatic system that performs well on two common-sense reasoning tasks, the Winograd Schema Challenge (WSC) and the Choice of Plausible Alternatives (COPA). Problem instances from these tasks require diverse, complex forms of inference and knowledge to solve. Our method uses a knowledge-hunting module to gather text from the web, which serves as evidence for candidate problem resolutions. Given an input problem, our system generates relevant queries to send to a search engine. It extracts and classifies knowledge from the returned results and weighs it to make a resolution. Our approach improves F1 performance on the WSC by 0.16 over the previous best and is competitive with the state-of-the-art on COPA, demonstrating its general applicability.

We introduce an automatic system that achieves state-of-the-art results on the Winograd Schema Challenge (WSC), a common sense reasoning task that requires diverse, complex forms of inference and knowledge. Our method uses a knowledge hunting module to gather text from the web, which serves as evidence for candidate problem resolutions. Given an input problem, our system generates relevant queries to send to a search engine, then extracts and classifies knowledge from the returned results and weighs them to make a resolution. Our approach improves F1 performance on the full WSC by 0.21 over the previous best and represents the first system to exceed 0.5 F1. We further demonstrate that the approach is competitive on the Choice of Plausible Alternatives (COPA) task, which suggests that it is generally applicable.