Jiao Sun

Jiao Sun, Thibault Sellam, Elizabeth Clark et al. · deepmind, uw

Evaluation metrics that are not robust to dialect variation make it impossible to tell how well systems perform for many groups of users, and can even penalize systems for producing text in lower-resource dialects. However, currently, there exists no way to quantify how metrics respond to change in the dialect of a generated utterance. We thus formalize dialect robustness and dialect awareness as goals for NLG evaluation metrics. We introduce a suite of methods and corresponding statistical tests one can use to assess metrics in light of the two goals. Applying the suite to current state-of-the-art metrics, we demonstrate that they are not dialect-robust and that semantic perturbations frequently lead to smaller decreases in a metric than the introduction of dialect features. As a first step to overcome this limitation, we propose a training schema, NANO, which introduces regional and language information to the pretraining process of a metric. We demonstrate that NANO provides a size-efficient way for models to improve the dialect robustness while simultaneously improving their performance on the standard metric benchmark.

Zhixuan Zhou, Jiao Sun, Jiaxin Pei et al. · stanford

Fairy tales are a common resource for young children to learn a language or understand how a society works. However, gender bias, e.g., stereotypical gender roles, in this literature may cause harm and skew children's world view. Instead of decades of qualitative and manual analysis of gender bias in fairy tales, we computationally analyze gender bias in a fairy tale dataset containing 624 fairy tales from 7 different cultures. We specifically examine gender difference in terms of moral foundations, which are measures of human morality, and events, which reveal human activities associated with each character. We find that the number of male characters is two times that of female characters, showing a disproportionate gender representation. Our analysis further reveal stereotypical portrayals of both male and female characters in terms of moral foundations and events. Female characters turn out more associated with care-, loyalty- and sanctity- related moral words, while male characters are more associated with fairness- and authority- related moral words. Female characters' events are often about emotion (e.g., weep), appearance (e.g., comb), household (e.g., bake), etc.; while male characters' events are more about profession (e.g., hunt), violence (e.g., destroy), justice (e.g., judge), etc. Gender bias in terms of moral foundations shows an obvious difference across cultures. For example, female characters are more associated with care and sanctity in high uncertainty-avoidance cultures which are less open to changes and unpredictability. Based on the results, we propose implications for children's literature and early literacy research.

Jiao Sun, Swabha Swayamdipta, Jonathan May et al. · cmu

Free-form rationales aim to aid model interpretability by supplying the background knowledge that can help understand model decisions. Crowdsourced rationales are provided for commonsense QA instances in popular datasets such as CoS-E and ECQA, but their utility remains under-investigated. We present human studies which show that ECQA rationales indeed provide additional background information to understand a decision, while over 88% of CoS-E rationales do not. Inspired by this finding, we ask: can the additional context provided by free-form rationales benefit models, similar to human users? We investigate the utility of rationales as an additional source of supervision, by varying the quantity and quality of rationales during training. After controlling for instances where rationales leak the correct answer while not providing additional background knowledge, we find that incorporating only 5% of rationales during training can boost model performance by 47.22% for CoS-E and 57.14% for ECQA during inference. Moreover, we also show that rationale quality matters: compared to crowdsourced rationales, T5-generated rationales provide not only weaker supervision to models, but are also not helpful for humans in aiding model interpretability.

Jiao Sun, Anjali Narayan-Chen, Shereen Oraby et al. · amazon-science

The tasks of humor understanding and generation are challenging and subjective even for humans, requiring commonsense and real-world knowledge to master. Puns, in particular, add the challenge of fusing that knowledge with the ability to interpret lexical-semantic ambiguity. In this paper, we present the ExPUNations (ExPUN) dataset, in which we augment an existing dataset of puns with detailed crowdsourced annotations of keywords denoting the most distinctive words that make the text funny, pun explanations describing why the text is funny, and fine-grained funniness ratings. This is the first humor dataset with such extensive and fine-grained annotations specifically for puns. Based on these annotations, we propose two tasks: explanation generation to aid with pun classification and keyword-conditioned pun generation, to challenge the current state-of-the-art natural language understanding and generation models' ability to understand and generate humor. We showcase that the annotated keywords we collect are helpful for generating better novel humorous texts in human evaluation, and that our natural language explanations can be leveraged to improve both the accuracy and robustness of humor classifiers.

Jiao Sun, Anjali Narayan-Chen, Shereen Oraby et al. · amazon-science

Previous work on pun generation commonly begins with a given pun word (a pair of homophones for heterographic pun generation and a polyseme for homographic pun generation) and seeks to generate an appropriate pun. While this may enable efficient pun generation, we believe that a pun is most entertaining if it fits appropriately within a given context, e.g., a given situation or dialogue. In this work, we propose a new task, context-situated pun generation, where a specific context represented by a set of keywords is provided, and the task is to first identify suitable pun words that are appropriate for the context, then generate puns based on the context keywords and the identified pun words. We collect CUP (Context-sitUated Pun), containing 4.5k tuples of context words and pun pairs. Based on the new data and setup, we propose a pipeline system for context-situated pun generation, including a pun word retrieval module that identifies suitable pun words for a given context, and a generation module that generates puns from context keywords and pun words. Human evaluation shows that 69% of our top retrieved pun words can be used to generate context-situated puns, and our generation module yields successful puns 31% of the time given a plausible tuple of context words and pun pair, almost tripling the yield of a state-of-the-art pun generation model. With an end-to-end evaluation, our pipeline system with the top-1 retrieved pun pair for a given context can generate successful puns 40% of the time, better than all other modeling variations but 32% lower than the human success rate. This highlights the difficulty of the task, and encourages more research in this direction.

Yixin Wan, George Pu, Jiao Sun et al.

Large Language Models (LLMs) have recently emerged as an effective tool to assist individuals in writing various types of content, including professional documents such as recommendation letters. Though bringing convenience, this application also introduces unprecedented fairness concerns. Model-generated reference letters might be directly used by users in professional scenarios. If underlying biases exist in these model-constructed letters, using them without scrutinization could lead to direct societal harms, such as sabotaging application success rates for female applicants. In light of this pressing issue, it is imminent and necessary to comprehensively study fairness issues and associated harms in this real-world use case. In this paper, we critically examine gender biases in LLM-generated reference letters. Drawing inspiration from social science findings, we design evaluation methods to manifest biases through 2 dimensions: (1) biases in language style and (2) biases in lexical content. We further investigate the extent of bias propagation by analyzing the hallucination bias of models, a term that we define to be bias exacerbation in model-hallucinated contents. Through benchmarking evaluation on 2 popular LLMs- ChatGPT and Alpaca, we reveal significant gender biases in LLM-generated recommendation letters. Our findings not only warn against using LLMs for this application without scrutinization, but also illuminate the importance of thoroughly studying hidden biases and harms in LLM-generated professional documents.

Arshiya Aggarwal, Jiao Sun, Nanyun Peng

We present a robust methodology for evaluating biases in natural language generation(NLG) systems. Previous works use fixed hand-crafted prefix templates with mentions of various demographic groups to prompt models to generate continuations for bias analysis. These fixed prefix templates could themselves be specific in terms of styles or linguistic structures, which may lead to unreliable fairness conclusions that are not representative of the general trends from tone varying prompts. To study this problem, we paraphrase the prompts with different syntactic structures and use these to evaluate demographic bias in NLG systems. Our results suggest similar overall bias trends but some syntactic structures lead to contradictory conclusions compared to past works. We show that our methodology is more robust and that some syntactic structures prompt more toxic content while others could prompt less biased generation. This suggests the importance of not relying on a fixed syntactic structure and using tone-invariant prompts. Introducing syntactically-diverse prompts can achieve more robust NLG (bias) evaluation.

Jiao Sun, Yufei Tian, Wangchunshu Zhou et al.

While recent studies have looked into the abilities of large language models in various benchmark tasks, including question generation, reading comprehension, multilingual and etc, there have been few studies looking into the controllability of large language models on generation tasks. We present an extensive analysis of various benchmarks including a sentence planning benchmark with different granularities. After comparing large language models against state-of-the-start finetuned smaller models, we present a spectrum showing large language models falling behind, are comparable, or exceed the ability of smaller models. We conclude that **large language models struggle at meeting fine-grained hard constraints**.

Jiao Sun, Deqing Fu, Yushi Hu et al.

Despite their wide-spread success, Text-to-Image models (T2I) still struggle to produce images that are both aesthetically pleasing and faithful to the user's input text. We introduce DreamSync, a model-agnostic training algorithm by design that improves T2I models to be faithful to the text input. DreamSync builds off a recent insight from TIFA's evaluation framework -- that large vision-language models (VLMs) can effectively identify the fine-grained discrepancies between generated images and the text inputs. DreamSync uses this insight to train T2I models without any labeled data; it improves T2I models using its own generations. First, it prompts the model to generate several candidate images for a given input text. Then, it uses two VLMs to select the best generation: a Visual Question Answering model that measures the alignment of generated images to the text, and another that measures the generation's aesthetic quality. After selection, we use LoRA to iteratively finetune the T2I model to guide its generation towards the selected best generations. DreamSync does not need any additional human annotation. model architecture changes, or reinforcement learning. Despite its simplicity, DreamSync improves both the semantic alignment and aesthetic appeal of two diffusion-based T2I models, evidenced by multiple benchmarks (+1.7% on TIFA, +2.9% on DSG1K, +3.4% on VILA aesthetic) and human evaluation.

Jiao Sun, Yu Hou, Jiin Kim et al.

Information-seeking AI assistant systems aim to answer users' queries about knowledge in a timely manner. However, both the human-perceived helpfulness of information-seeking assistant systems and its fairness implication are under-explored. In this paper, we study computational measurements of helpfulness. We collect human annotations on the helpfulness of dialogue responses, develop models for automatic helpfulness evaluation, and then propose to use the helpfulness level of a dialogue system towards different user queries to gauge the fairness of a dialogue system. Experiments with state-of-the-art dialogue systems, including ChatGPT, under three information-seeking scenarios reveal that existing systems tend to be more helpful for questions regarding concepts from highly-developed countries than less-developed countries, uncovering potential fairness concerns underlying the current information-seeking assistant systems.

Youwei Liang, Junfeng He, Gang Li et al.

Recent Text-to-Image (T2I) generation models such as Stable Diffusion and Imagen have made significant progress in generating high-resolution images based on text descriptions. However, many generated images still suffer from issues such as artifacts/implausibility, misalignment with text descriptions, and low aesthetic quality. Inspired by the success of Reinforcement Learning with Human Feedback (RLHF) for large language models, prior works collected human-provided scores as feedback on generated images and trained a reward model to improve the T2I generation. In this paper, we enrich the feedback signal by (i) marking image regions that are implausible or misaligned with the text, and (ii) annotating which words in the text prompt are misrepresented or missing on the image. We collect such rich human feedback on 18K generated images (RichHF-18K) and train a multimodal transformer to predict the rich feedback automatically. We show that the predicted rich human feedback can be leveraged to improve image generation, for example, by selecting high-quality training data to finetune and improve the generative models, or by creating masks with predicted heatmaps to inpaint the problematic regions. Notably, the improvements generalize to models (Muse) beyond those used to generate the images on which human feedback data were collected (Stable Diffusion variants). The RichHF-18K data set will be released in our GitHub repository: https://github.com/google-research/google-research/tree/master/richhf_18k.

Gemini Team, Petko Georgiev, Ving Ian Lei et al. · deepmind, mila

In this report, we introduce the Gemini 1.5 family of models, representing the next generation of highly compute-efficient multimodal models capable of recalling and reasoning over fine-grained information from millions of tokens of context, including multiple long documents and hours of video and audio. The family includes two new models: (1) an updated Gemini 1.5 Pro, which exceeds the February version on the great majority of capabilities and benchmarks; (2) Gemini 1.5 Flash, a more lightweight variant designed for efficiency with minimal regression in quality. Gemini 1.5 models achieve near-perfect recall on long-context retrieval tasks across modalities, improve the state-of-the-art in long-document QA, long-video QA and long-context ASR, and match or surpass Gemini 1.0 Ultra's state-of-the-art performance across a broad set of benchmarks. Studying the limits of Gemini 1.5's long-context ability, we find continued improvement in next-token prediction and near-perfect retrieval (>99%) up to at least 10M tokens, a generational leap over existing models such as Claude 3.0 (200k) and GPT-4 Turbo (128k). Finally, we highlight real-world use cases, such as Gemini 1.5 collaborating with professionals on completing their tasks achieving 26 to 75% time savings across 10 different job categories, as well as surprising new capabilities of large language models at the frontier; when given a grammar manual for Kalamang, a language with fewer than 200 speakers worldwide, the model learns to translate English to Kalamang at a similar level to a person who learned from the same content.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Haoyi Qiu, Kung-Hsiang Huang, Ruichen Zheng et al.

Large vision-language models (LVLMs) are increasingly deployed in globally distributed applications, such as tourism assistants, yet their ability to produce culturally appropriate responses remains underexplored. Existing multimodal safety benchmarks primarily focus on physical safety and overlook violations rooted in cultural norms, which can result in symbolic harm. To address this gap, we introduce CROSS, a benchmark designed to assess the cultural safety reasoning capabilities of LVLMs. CROSS includes 1,284 multilingual visually grounded queries from 16 countries, three everyday domains, and 14 languages, where cultural norm violations emerge only when images are interpreted in context. We propose CROSS-Eval, an intercultural theory-based framework that measures four key dimensions: cultural awareness, norm education, compliance, and helpfulness. Using this framework, we evaluate 21 leading LVLMs, including mixture-of-experts models and reasoning models. Results reveal significant cultural safety gaps: the best-performing model achieves only 61.79% in awareness and 37.73% in compliance. While some open-source models reach GPT-4o-level performance, they still fall notably short of proprietary models. Our results further show that increasing reasoning capacity improves cultural alignment but does not fully resolve the issue. To improve model performance, we develop two enhancement strategies: supervised fine-tuning with culturally grounded, open-ended data and preference tuning with contrastive response pairs that highlight safe versus unsafe behaviors. These methods substantially improve GPT-4o's cultural awareness (+60.14%) and compliance (+55.2%), while preserving general multimodal capabilities with minimal performance reduction on general multimodal understanding benchmarks.

Xin Liu, Rongwu Xu, Xinyi Jia et al. · uw

The rise of large language models (LLMs) has enabled the generation of highly persuasive spam reviews that closely mimic human writing. These reviews pose significant challenges for existing detection systems and threaten the credibility of online platforms. In this work, we first create three realistic LLM-generated spam review datasets using three distinct LLMs, each guided by product metadata and genuine reference reviews. Evaluations by GPT-4.1 confirm the high persuasion and deceptive potential of these reviews. To address this threat, we propose FraudSquad, a hybrid detection model that integrates text embeddings from a pre-trained language model with a gated graph transformer for spam node classification. FraudSquad captures both semantic and behavioral signals without relying on manual feature engineering or massive training resources. Experiments show that FraudSquad outperforms state-of-the-art baselines by up to 44.22% in precision and 43.01% in recall on three LLM-generated datasets, while also achieving promising results on two human-written spam datasets. Furthermore, FraudSquad maintains a modest model size and requires minimal labeled training data, making it a practical solution for real-world applications. Our contributions include new synthetic datasets, a practical detection framework, and empirical evidence highlighting the urgency of adapting spam detection to the LLM era. Our code and datasets are available at: https://anonymous.4open.science/r/FraudSquad-5389/.

Longju Bai, Zhemin Huang, Xingyao Wang et al.

The wide adoption of AI agents in complex human workflows is driving rapid growth in LLM token consumption. When agents are deployed on tasks that require a significant amount of tokens, three questions naturally arise: (1) Where do AI agents spend the tokens? (2) Which models are more token-efficient? and (3) Can agents predict their token usage before task execution? In this paper, we present the first systematic study of token consumption patterns in agentic coding tasks. We analyze trajectories from eight frontier LLMs on SWE-bench Verified and evaluate models' ability to predict their own token costs before task execution. We find that: (1) agentic tasks are uniquely expensive, consuming 1000x more tokens than code reasoning and code chat, with input tokens rather than output tokens driving the overall cost; (2) token usage is highly variable and inherently stochastic: runs on the same task can differ by up to 30x in total tokens, and higher token usage does not translate into higher accuracy; instead, accuracy often peaks at intermediate cost and saturates at higher costs; (3) models vary substantially in token efficiency: on the same tasks, Kimi-K2 and Claude-Sonnet-4.5, on average, consume over 1.5 million more tokens than GPT-5; (4) task difficulty rated by human experts only weakly aligns with actual token costs, revealing a fundamental gap between human-perceived complexity and the computational effort agents actually expend; and (5) frontier models fail to accurately predict their own token usage (with weak-to-moderate correlations, up to 0.39) and systematically underestimate real token costs. Our study offers new insights into the economics of AI agents and can inspire future research in this direction.

Jiao Sun

The evolution of biological morphology is critical for understanding the diversity of the natural world, yet traditional analyses often involve subjective biases in the selection and coding of morphological traits. This study employs deep learning techniques, utilising a ResNet34 model capable of recognising over 10,000 bird species, to explore avian morphological evolution. We extract weights from the model's final fully connected (fc) layer and investigate the semantic alignment between the high-dimensional embedding space learned by the model and biological phenotypes. The results demonstrate that the high-dimensional embedding space encodes phenotypic convergence. Subsequently, we assess the morphological disparity among various taxa and evaluate the association between morphological disparity and species richness, demonstrating that species richness is the primary driver of morphospace expansion. Moreover, the disparity-through-time analysis reveals a visual "early burst" after the K-Pg extinction. While mainly aimed at evolutionary analysis, this study also provides insights into the interpretability of Deep Neural Networks. We demonstrate that hierarchical semantic structures (biological taxonomy) emerged in the high-dimensional embedding space despite being trained on flat labels. Furthermore, through adversarial examples, we provide evidence that our model in this task can overcome texture bias and learn holistic shape representations (body plans), challenging the prevailing view that CNNs rely primarily on local textures.

Ming Zhong, Xiang Zhou, Ting-Yun Chang et al.

Large Language Models (LLMs) have catalyzed vibe coding, where users leverage LLMs to generate and iteratively refine code through natural language interactions until it passes their vibe check. Vibe check is tied to real-world human preference and goes beyond functionality: the solution should feel right, read cleanly, preserve intent, and remain correct. However, current code evaluation remains anchored to pass@k and captures only functional correctness, overlooking the non-functional instructions that users routinely apply. In this paper, we hypothesize that instruction following is the missing piece underlying vibe check that represents human preference in coding besides functional correctness. To quantify models' code instruction following capabilities with measurable signals, we present VeriCode, a taxonomy of 30 verifiable code instructions together with corresponding deterministic verifiers. We use the taxonomy to augment established evaluation suites, resulting in Vibe Checker, a testbed to assess both code instruction following and functional correctness. Upon evaluating 31 leading LLMs, we show that even the strongest models struggle to comply with multiple instructions and exhibit clear functional regression. Most importantly, a composite score of functional correctness and instruction following correlates the best with human preference, with the latter emerging as the primary differentiator on real-world programming tasks. Our work identifies core factors of the vibe check, providing a concrete path for benchmarking and developing models that better align with user preferences in coding.

Yufei Tian, Jiao Sun, Nanyun Peng et al.

As language models evolve to tackle complex, multifaceted tasks, their evaluation must adapt to capture this intricacy. A granular, skill-specific understanding of model capabilities can empower researchers to make informed model development plans. In this paper, we introduce SkillVerse, an unsupervised tree-structured diagnosis framework for understanding model proficiency in specific abilities. With LLM as a judge, SkillVerse first critiques the model responses, and then organizes them into a hierarchical structure termed dendrogram. Given proficiency at arbitrary levels of granularity, SkillVerse is flexible to produce insights of behaviors of modern large models. We also demonstrate its efficacy in two downstream tasks: 1) improving model in-context learning by 25% using a tree-search algorithm to select more informative few-shot demonstrations, and 2) accurately predicting new model weaknesses with a 55% success rate, 22% higher than without SkillVerse.

Shenzhe Zhu, Jiao Sun, Yi Nian et al. · mit

AI agents are increasingly used in consumer-facing applications to assist with tasks such as product search, negotiation, and transaction execution. In this paper, we explore a future scenario where both consumers and merchants authorize AI agents to fully automate negotiations and transactions. We aim to answer two key questions: (1) Do different LLM agents vary in their ability to secure favorable deals for users? (2) What risks arise from fully automating deal-making with AI agents in consumer markets? To address these questions, we develop an experimental framework that evaluates the performance of various LLM agents in real-world negotiation and transaction settings. Our findings reveal that AI-mediated deal-making is an inherently imbalanced game -- different agents achieve significantly different outcomes for their users. Moreover, behavioral anomalies in LLMs can result in financial losses for both consumers and merchants, such as overspending or accepting unreasonable deals. These results underscore that while automation can improve efficiency, it also introduces substantial risks. Users should exercise caution when delegating business decisions to AI agents.

Mohsen Fayyaz, Fan Yin, Jiao Sun et al.

We study how well large language models (LLMs) explain their generations through rationales -- a set of tokens extracted from the input text that reflect the decision-making process of LLMs. Specifically, we systematically study rationales derived using two approaches: (1) popular prompting-based methods, where prompts are used to guide LLMs in generating rationales, and (2) technical attribution-based methods, which leverage attention or gradients to identify important tokens. Our analysis spans three classification datasets with annotated rationales, encompassing tasks with varying performance levels. While prompting-based self-explanations are widely used, our study reveals that these explanations are not always as "aligned" with the human rationale as attribution-based explanations. Even more so, fine-tuning LLMs to enhance classification task accuracy does not enhance the alignment of prompting-based rationales. Still, it does considerably improve the alignment of attribution-based methods (e.g., InputXGradient). More importantly, we show that prompting-based self-explanation is also less "faithful" than attribution-based explanations, failing to provide a reliable account of the model's decision-making process. To evaluate faithfulness, unlike prior studies that excluded misclassified examples, we evaluate all instances and also examine the impact of fine-tuning and accuracy on alignment and faithfulness. Our findings suggest that inconclusive faithfulness results reported in earlier studies may stem from low classification accuracy. These findings underscore the importance of more rigorous and comprehensive evaluations of LLM rationales.

Liangchen Luo, Yinxiao Liu, Rosanne Liu et al.

Complex multi-step reasoning tasks, such as solving mathematical problems or generating code, remain a significant hurdle for even the most advanced large language models (LLMs). Verifying LLM outputs with an Outcome Reward Model (ORM) is a standard inference-time technique aimed at enhancing the reasoning performance of LLMs. However, this still proves insufficient for reasoning tasks with a lengthy or multi-hop reasoning chain, where the intermediate outcomes are neither properly rewarded nor penalized. Process supervision addresses this limitation by assigning intermediate rewards during the reasoning process. To date, the methods used to collect process supervision data have relied on either human annotation or per-step Monte Carlo estimation, both prohibitively expensive to scale, thus hindering the broad application of this technique. In response to this challenge, we propose a novel divide-and-conquer style Monte Carlo Tree Search (MCTS) algorithm named \textit{OmegaPRM} for the efficient collection of high-quality process supervision data. This algorithm swiftly identifies the first error in the Chain of Thought (CoT) with binary search and balances the positive and negative examples, thereby ensuring both efficiency and quality. As a result, we are able to collect over 1.5 million process supervision annotations to train Process Reward Models (PRMs). This fully automated process supervision alongside the weighted self-consistency algorithm is able to enhance LLMs' math reasoning performances. We improved the success rates of the instruction-tuned Gemini Pro model from 51\% to 69.4\% on MATH500 and from 86.4\% to 93.6\% on GSM8K. Similarly, we boosted the success rates of Gemma2 27B from 42.3\% to 58.2\% on MATH500 and from 74.0\% to 92.2\% on GSM8K. The entire process operates without any human intervention or supervision, making our method both financially and ...

Chunting Zhou, Pengfei Liu, Puxin Xu et al.

Large language models are trained in two stages: (1) unsupervised pretraining from raw text, to learn general-purpose representations, and (2) large scale instruction tuning and reinforcement learning, to better align to end tasks and user preferences. We measure the relative importance of these two stages by training LIMA, a 65B parameter LLaMa language model fine-tuned with the standard supervised loss on only 1,000 carefully curated prompts and responses, without any reinforcement learning or human preference modeling. LIMA demonstrates remarkably strong performance, learning to follow specific response formats from only a handful of examples in the training data, including complex queries that range from planning trip itineraries to speculating about alternate history. Moreover, the model tends to generalize well to unseen tasks that did not appear in the training data. In a controlled human study, responses from LIMA are either equivalent or strictly preferred to GPT-4 in 43% of cases; this statistic is as high as 58% when compared to Bard and 65% versus DaVinci003, which was trained with human feedback. Taken together, these results strongly suggest that almost all knowledge in large language models is learned during pretraining, and only limited instruction tuning data is necessary to teach models to produce high quality output.

Jiao Sun, Q. Vera Liao, Michael Muller et al.

What does it mean for a generative AI model to be explainable? The emergent discipline of explainable AI (XAI) has made great strides in helping people understand discriminative models. Less attention has been paid to generative models that produce artifacts, rather than decisions, as output. Meanwhile, generative AI (GenAI) technologies are maturing and being applied to application domains such as software engineering. Using scenario-based design and question-driven XAI design approaches, we explore users' explainability needs for GenAI in three software engineering use cases: natural language to code, code translation, and code auto-completion. We conducted 9 workshops with 43 software engineers in which real examples from state-of-the-art generative AI models were used to elicit users' explainability needs. Drawing from prior work, we also propose 4 types of XAI features for GenAI for code and gathered additional design ideas from participants. Our work explores explainability needs for GenAI for code and demonstrates how human-centered approaches can drive the technical development of XAI in novel domains.

Jiao Sun, Tongshuang Wu, Yue Jiang et al.

People write personalized greeting cards on various occasions. While prior work has studied gender roles in greeting card messages, systematic analysis at scale and tools for raising the awareness of gender stereotyping remain under-investigated. To this end, we collect a large greeting card message corpus covering three different occasions (birthday, Valentine's Day and wedding) from three sources (exemplars from greeting message websites, real-life greetings from social media and language model generated ones). We uncover a wide range of gender stereotypes in this corpus via topic modeling, odds ratio and Word Embedding Association Test (WEAT). We further conduct a survey to understand people's perception of gender roles in messages from this corpus and if gender stereotyping is a concern. The results show that people want to be aware of gender roles in the messages, but remain unconcerned unless the perceived gender roles conflict with the recipient's true personality. In response, we developed GreetA, an interactive visualization and writing assistant tool to visualize fine-grained topics in greeting card messages drafted by the users and the associated gender perception scores, but without suggesting text changes as an intervention.

Chenyu Wang, Zongyu Lin, Xiaochen Yang et al.

The crime forecasting is an important problem as it greatly contributes to urban safety. Typically, the goal of the problem is to predict different types of crimes for each geographical region (like a neighborhood or censor tract) in the near future. Since nearby regions usually have similar socioeconomic characteristics which indicate similar crime patterns, recent state-of-the-art solutions constructed a distance-based region graph and utilized Graph Neural Network (GNN) techniques for crime forecasting, because the GNN techniques could effectively exploit the latent relationships between neighboring region nodes in the graph. However, this distance-based pre-defined graph cannot fully capture crime correlation between regions that are far from each other but share similar crime patterns. Hence, to make an accurate crime prediction, the main challenge is to learn a better graph that reveals the dependencies between regions in crime occurrences and meanwhile captures the temporal patterns from historical crime records. To address these challenges, we propose an end-to-end graph convolutional recurrent network called HAGEN with several novel designs for crime prediction. Specifically, our framework could jointly capture the crime correlation between regions and the temporal crime dynamics by combining an adaptive region graph learning module with the Diffusion Convolution Gated Recurrent Unit (DCGRU). Based on the homophily assumption of GNN, we propose a homophily-aware constraint to regularize the optimization of the region graph so that neighboring region nodes on the learned graph share similar crime patterns, thus fitting the mechanism of diffusion convolution. It also incorporates crime embedding to model the interdependencies between regions and crime categories. Empirical experiments and comprehensive analysis on two real-world datasets showcase the effectiveness of HAGEN.

Sunipa Dev, Emily Sheng, Jieyu Zhao et al.

Recent studies show that Natural Language Processing (NLP) technologies propagate societal biases about demographic groups associated with attributes such as gender, race, and nationality. To create interventions and mitigate these biases and associated harms, it is vital to be able to detect and measure such biases. While existing works propose bias evaluation and mitigation methods for various tasks, there remains a need to cohesively understand the biases and the specific harms they measure, and how different measures compare with each other. To address this gap, this work presents a practical framework of harms and a series of questions that practitioners can answer to guide the development of bias measures. As a validation of our framework and documentation questions, we also present several case studies of how existing bias measures in NLP -- both intrinsic measures of bias in representations and extrinsic measures of bias of downstream applications -- can be aligned with different harms and how our proposed documentation questions facilitates more holistic understanding of what bias measures are measuring.

Jiao Sun, Nanyun Peng

Human activities can be seen as sequences of events, which are crucial to understanding societies. Disproportional event distribution for different demographic groups can manifest and amplify social stereotypes, and potentially jeopardize the ability of members in some groups to pursue certain goals. In this paper, we present the first event-centric study of gender biases in a Wikipedia corpus. To facilitate the study, we curate a corpus of career and personal life descriptions with demographic information consisting of 7,854 fragments from 10,412 celebrities. Then we detect events with a state-of-the-art event detection model, calibrate the results using strategically generated templates, and extract events that have asymmetric associations with genders. Our study discovers that the Wikipedia pages tend to intermingle personal life events with professional events for females but not for males, which calls for the awareness of the Wikipedia community to formalize guidelines and train the editors to mind the implicit biases that contributors carry. Our work also lays the foundation for future works on quantifying and discovering event biases at the corpus level.

Rujun Han, I-Hung Hsu, Jiao Sun et al.

Understanding how events are semantically related to each other is the essence of reading comprehension. Recent event-centric reading comprehension datasets focus mostly on event arguments or temporal relations. While these tasks partially evaluate machines' ability of narrative understanding, human-like reading comprehension requires the capability to process event-based information beyond arguments and temporal reasoning. For example, to understand causality between events, we need to infer motivation or purpose; to establish event hierarchy, we need to understand the composition of events. To facilitate these tasks, we introduce ESTER, a comprehensive machine reading comprehension (MRC) dataset for Event Semantic Relation Reasoning. The dataset leverages natural language queries to reason about the five most common event semantic relations, provides more than 6K questions and captures 10.1K event relation pairs. Experimental results show that the current SOTA systems achieve 22.1%, 63.3%, and 83.5% for token-based exact-match, F1, and event-based HIT@1 scores, which are all significantly below human performances (36.0%, 79.6%, 100% respectively), highlighting our dataset as a challenging benchmark.

Mingyu Derek Ma, Jiao Sun, Mu Yang et al.

We present EventPlus, a temporal event understanding pipeline that integrates various state-of-the-art event understanding components including event trigger and type detection, event argument detection, event duration and temporal relation extraction. Event information, especially event temporal knowledge, is a type of common sense knowledge that helps people understand how stories evolve and provides predictive hints for future events. EventPlus as the first comprehensive temporal event understanding pipeline provides a convenient tool for users to quickly obtain annotations about events and their temporal information for any user-provided document. Furthermore, we show EventPlus can be easily adapted to other domains (e.g., biomedical domain). We make EventPlus publicly available to facilitate event-related information extraction and downstream applications.

Lu Chen, Jiao Sun, Wei Xu

Deep neural networks (DNNs) significantly improved the accuracy of optical character recognition (OCR) and inspired many important applications. Unfortunately, OCRs also inherit the vulnerabilities of DNNs under adversarial examples. Different from colorful vanilla images, text images usually have clear backgrounds. Adversarial examples generated by most existing adversarial attacks are unnatural and pollute the background severely. To address this issue, we propose the Fast Adversarial Watermark Attack (FAWA) against sequence-based OCR models in the white-box manner. By disguising the perturbations as watermarks, we can make the resulting adversarial images appear natural to human eyes and achieve a perfect attack success rate. FAWA works with either gradient-based or optimization-based perturbation generation. In both letter-level and word-level attacks, our experiments show that in addition to natural appearance, FAWA achieves a 100% attack success rate with 60% less perturbations and 78% fewer iterations on average. In addition, we further extend FAWA to support full-color watermarks, other languages, and even the OCR accuracy-enhancing mechanism.

Yikun Ban, Jiao Sun, Xin Liu

Fraud has severely detrimental impacts on the business of social networks and other online applications. A user can become a fake celebrity by purchasing "zombie followers" on Twitter. A merchant can boost his reputation through fake reviews on Amazon. This phenomenon also conspicuously exists on Facebook, Yelp and TripAdvisor, etc. In all the cases, fraudsters try to manipulate the platform's ranking mechanism by faking interactions between the fake accounts they control and the target customers.