Atharvan Dogra, Krishna Pillutla, Ameet Deshpande, Ananya B. Sai, John J Nay, Tanmay Rajpurohit, Ashwin Kalyan, Balaraman Ravindran. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Recent works have showcased the ability of large-scale language models (LLMs) to embody diverse personas in their responses, exemplified by prompts like ‘_You are Yoda. Explain the Theory of Relativity._’ While this ability allows personalization of LLMs and enables human behavior simulation, its effect on LLMs’ capabilities remains unclear. To fill this gap, we present the first extensive study of the unintended side-effects of persona assignment on the ability of LLMs to perform _basic reasoning tasks_. Our study covers 24 reasoning datasets (spanning mathematics, law, medicine, morals, and more), 4 LLMs (2 versions of ChatGPT-3.5, GPT-4-Turbo, and Llama-2-70b-chat), and 19 diverse personas (e.g., ‘an Asian person’) spanning 5 socio-demographic groups: race, gender, religion, disability, and political affiliation. Our experiments unveil that LLMs harbor deep rooted bias against various socio-demographics underneath a veneer of fairness. While they overtly reject stereotypes when explicitly asked (‘_Are Black people less skilled at mathematics?_’), they manifest stereotypical and often erroneous presumptions when prompted to answer questions while adopting a persona. These can be observed as abstentions in the model’s response, e.g., ‘_As a Black person, I am unable to answer this question as it requires math knowledge_’, and generally result in a substantial drop in performance on reasoning tasks. Our experiments with ChatGPT-3.5 show that this bias is _ubiquitous_—80% of our personas demonstrate bias; it is _significant_—some datasets show performance drops of 70%+; and can be especially _harmful for certain groups_—some personas suffer statistically significant drops on 80%+ of the datasets. Overall, all four LLMs exhibit persona-induced bias to varying extents, with GPT-4-Turbo showing the least but still a problematic amount of bias (evident in 42% of the personas). Further analysis shows that these persona-induced errors can be hard-to-discern as they do not always manifest as explicit abstentions, and can also be hard-to-avoid—we find de-biasing prompts to have minimal to no effect. Our findings serve as a cautionary tale that the practice of assigning personas to LLMs—a trend on the rise—can surface their deep-rooted biases and have unforeseeable and detrimental side-effects.

The advent of large language models (LLMs) has ushered in a new paradigm of search engines that use generative models to gather and summarize information to answer user queries. This emerging technology, which we formalize under the unified framework of generative engines (GEs), can generate accurate and personalized responses, rapidly replacing traditional search engines like Google and Bing. Generative Engines typically satisfy queries by synthesizing information from multiple sources and summarizing them using LLMs. While this shift significantly improvesuser utility and generative search engine traffic, it poses a huge challenge for the third stakeholder -- website and content creators. Given the black-box and fast-moving nature of generative engines, content creators have little to no control over when and how their content is displayed. With generative engines here to stay, we must ensure the creator economy is not disadvantaged. To address this, we introduce Generative Engine Optimization (GEO), the first novel paradigm to aid content creators in improving their content visibility in generative engine responses through a flexible black-box optimization framework for optimizing and defining visibility metrics. We facilitate systematic evaluation by introducing GEO-bench, a large-scale benchmark of diverse user queries across multiple domains, along with relevant web sources to answer these queries. Through rigorous evaluation, we demonstrate that GEO can boost visibility by up to 40% in generative engine responses. Moreover, we show the efficacy of these strategies varies across domains, underscoring the need for domain-specific optimization methods. Our work opens a new frontier in information discovery systems, with profound implications for both developers of generative engines and content creators.

Vishvak Murahari, Ameet Deshpande, Peter Clark, Tanmay Rajpurohit, Ashish Sabharwal, Karthik Narasimhan, Ashwin Kalyan. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Afra Feyza Akyurek, Ekin Akyurek, Ashwin Kalyan, Peter Clark, Derry Tanti Wijaya, Niket Tandon. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Ameet Deshpande, Carlos Jimenez, Howard Chen, Vishvak Murahari, Victoria Graf, Tanmay Rajpurohit, Ashwin Kalyan, Danqi Chen, Karthik Narasimhan. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

In this paper, we present a novel approach for distilling math word problem solving capabilities from large language models (LLMs) into smaller, more efficient student models. Our approach is designed to consider the student model's weaknesses and foster a tailored learning experience by generating targeted exercises aligned with educational science principles, such as knowledge tracing and personalized learning. Concretely, we let GPT-3 be a math tutor and run two steps iteratively: 1) assessing the student model's current learning status on a GPT-generated exercise book, and 2) improving the student model by training it with tailored exercise samples generated by GPT-3. Experimental results reveal that our approach outperforms LLMs (e.g., GPT-3 and PaLM) in accuracy across three distinct benchmarks while employing significantly fewer parameters. Furthermore, we provide a comprehensive analysis of the various components within our methodology to substantiate their efficacy.

Swaroop Mishra, Arindam Mitra, Neeraj Varshney, Bhavdeep Sachdeva, Peter Clark, Chitta Baral, Ashwin Kalyan. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Swaroop Mishra, Matthew Finlayson, Pan Lu, Leonard Tang, Sean Welleck, Chitta Baral, Tanmay Rajpurohit, Oyvind Tafjord, Ashish Sabharwal, Peter Clark, Ashwin Kalyan. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022.

When answering a question, humans utilize the information available across different modalities to synthesize a consistent and complete chain of thought (CoT). This process is normally a black box in the case of deep learning models like large-scale language models. Recently, science question benchmarks have been used to diagnose the multi-hop reasoning ability and interpretability of an AI system. However, existing datasets fail to provide annotations for the answers, or are restricted to the textual-only modality, small scales, and limited domain diversity. To this end, we present Science Question Answering (ScienceQA), a new benchmark that consists of ~21k multimodal multiple choice questions with a diverse set of science topics and annotations of their answers with corresponding lectures and explanations. We further design language models to learn to generate lectures and explanations as the chain of thought (CoT) to mimic the multi-hop reasoning process when answering ScienceQA questions. ScienceQA demonstrates the utility of CoT in language models, as CoT improves the question answering performance by 1.20% in few-shot GPT-3 and 3.99% in fine-tuned UnifiedQA. We also explore the upper bound for models to leverage explanations by feeding those in the input; we observe that it improves the few-shot performance of GPT-3 by 18.96%. Our analysis further shows that language models, similar to humans, benefit from explanations to learn from fewer data and achieve the same performance with just 40% of the data. The data and code are available at https://scienceqa.github.io.

Many real-world problems require the combined application of multiple reasoning abilities-employing suitable abstractions, commonsense knowledge, and creative synthesis of problem-solving strategies.To help advance AI systems towards such capabilities, we propose a new reasoning challenge, namely Fermi Problems (FPs), which are questions whose answers can only be approximately estimated because their precise computation is either impractical or impossible.For example, "How much would the sea level rise if all ice in the world melted?"FPs are commonly used in quizzes and interviews to bring out and evaluate the creative reasoning abilities of humans.To do the same for AI systems, we present two datasets: 1) A collection of 1k real-world FPs sourced from quizzes and olympiads; and 2) a bank of 10k synthetic FPs of intermediate complexity to serve as a sandbox for the harder real-world challenge.In addition to question-answer pairs, the datasets contain detailed solutions in the form of an executable program and supporting facts, helping in supervision and evaluation of intermediate steps.We demonstrate that even extensively fine-tuned large-scale language models perform poorly on these datasets, on average making estimates that are off by two orders of magnitude.Our contribution is thus the crystallization of several unsolved AI problems into a single, new challenge that we hope will spur further advances in building systems that can reason. Solving a Fermi Problem How much would the sea level rise if all the ice melted?Ice on land causes rise in sea levels. Div(a, b)Vol. of ice in the world? million mi²Area of ice?Area of ocean?Thickness of ice? 3 mi Area of Antarctica?Vol. of ice on land?How many Antarcticas fit in the world map?Vol. of ice in Antarctica?Mul(a, b) Ice on land Ice in Antarctica ≈ Area of Antarctica Area of ice in Antarctica

Many sequence prediction tasks admit multiple correct outputs and so, it is often useful to decode a set of outputs that maximize some task-specific set-level metric. However, retooling standard sequence prediction procedures tailored towards predicting the single best output leads to the decoding of sets containing very similar sequences; failing to capture the variation in the output space. To address this, we propose $\nabla$BS, a trainable decoding procedure that outputs a set of sequences, highly valued according to the metric. Our method tightly integrates the training and decoding phases and further allows for the optimization of the task-specific metric addressing the shortcomings of standard sequence prediction. Further, we discuss the trade-offs of commonly used set-level metrics and motivate a new set-level metric that naturally evaluates the notion of “capturing the variation in the output space”. Finally, we show results on the image captioning task and find that our model outperforms standard techniques and natural ablations.

Many structured prediction problems (particularly in vision and language domains) are ambiguous, with multiple outputs being ‘correct’ for an input {–} e.g. there are many ways of describing an image, multiple ways of translating a sentence; however, exhaustively annotating the applicability of all possible outputs is intractable due to exponentially large output spaces (e.g. all English sentences). In practice, these problems are cast as multi-class prediction, with the likelihood of only a sparse set of annotations being maximized {–} unfortunately penalizing for placing beliefs on plausible but unannotated outputs. We make and test the following hypothesis {–} for a given input, the annotations of its neighbors may serve as an additional supervisory signal. Specifically, we propose an objective that transfers supervision from neighboring examples. We first study the properties of our developed method in a controlled toy setup before reporting results on multi-label classification and two image-grounded sequence modeling tasks {–} captioning and question generation. We evaluate using standard task-specific metrics and measures of output diversity, finding consistent improvements over standard maximum likelihood training and other baselines.