Showing 1–45 of 45 results for author: Pang, R Y

MolmoWeb: Open Visual Web Agent and Open Data for the Open Web

Authors: Tanmay Gupta, Piper Wolters, Zixian Ma, Peter Sushko, Rock Yuren Pang, Diego Llanes, Yue Yang, Taira Anderson, Boyuan Zheng, Zhongzheng Ren, Harsh Trivedi, Taylor Blanton, Caleb Ouellette, Winson Han, Ali Farhadi, Ranjay Krishna

Abstract: Web agents--autonomous systems that navigate and execute tasks on the web on behalf of users--have the potential to transform how people interact with the digital world. However, the most capable web agents today rely on proprietary models with undisclosed training data and recipes, limiting scientific understanding, reproducibility, and community-driven progress. We believe agents for the open… ▽ More Web agents--autonomous systems that navigate and execute tasks on the web on behalf of users--have the potential to transform how people interact with the digital world. However, the most capable web agents today rely on proprietary models with undisclosed training data and recipes, limiting scientific understanding, reproducibility, and community-driven progress. We believe agents for the open web should be built in the open. To this end, we introduce (1) MolmoWebMix, a large and diverse mixture of browser task demonstrations and web-GUI perception data and (2) MolmoWeb, a family of fully open multimodal web agents. Specifically, MolmoWebMix combines over 100K synthetic task trajectories from multiple complementary generation pipelines with 30K+ human demonstrations, atomic web-skill trajectories, and GUI perception data, including referring expression grounding and screenshot question answering. MolmoWeb agents operate as instruction-conditioned visual-language action policies: given a task instruction and a webpage screenshot, they predict the next browser action, requiring no access to HTML, accessibility trees, or specialized APIs. Available in 4B and 8B size, on browser-use benchmarks like WebVoyager, Online-Mind2Web, and DeepShop, MolmoWeb agents achieve state-of-the-art results outperforming similar scale open-weight-only models such as Fara-7B, UI-Tars-1.5-7B, and Holo1-7B. MolmoWeb-8B also surpasses set-of-marks (SoM) agents built on much larger closed frontier models like GPT-4o. We further demonstrate consistent gains through test-time scaling via parallel rollouts with best-of-N selection, achieving 94.7% and 60.5% pass@4 (compared to 78.2% and 35.3% pass@1) on WebVoyager and Online-Mind2Web respectively. We will release model checkpoints, training data, code, and a unified evaluation harness to enable reproducibility and accelerate open research on web agents. △ Less

Submitted 9 April, 2026; originally announced April 2026.

Comments: https://allenai.org/blog/molmoweb

Interrogating Design Homogenization in Web Vibe Coding

Authors: Donghoon Shin, Alice Gao, Rock Yuren Pang, Jaewook Lee, Katharina Reinecke, Emily Tseng

Abstract: Generative AI is known for its tendency to homogenize, often reproducing dominant style conventions found in training data. However, it remains unclear how these homogenizing effects extend to complex structural tasks like web design. As lay creators increasingly turn to LLMs to 'vibe-code' websites -- prompting for aesthetic and functional goals rather than writing code -- they may inadvertently… ▽ More Generative AI is known for its tendency to homogenize, often reproducing dominant style conventions found in training data. However, it remains unclear how these homogenizing effects extend to complex structural tasks like web design. As lay creators increasingly turn to LLMs to 'vibe-code' websites -- prompting for aesthetic and functional goals rather than writing code -- they may inadvertently narrow the diversity of their designs, and limit creative expression throughout the internet. In this paper, we interrogate the possibility of design homogenization in web vibe coding. We first characterize the vibe coding lifecycle, pinpointing stages where homogenization risks may arise. We then conduct a sociotechnical risk analysis unpacking the potential harms of web vibe coding and their interaction with design homogenization. We identify that the push for frictionless generation can exacerbate homogenization and its harms. Finally, we propose a mitigation framework centered on the idea of productive friction. Through case studies at the micro, meso, and macro levels, we show how centering productive friction can empower creators to challenge default outputs and preserve diverse expression in AI-mediated web design. △ Less

Submitted 13 March, 2026; originally announced March 2026.

ACM Class: H.5.2; I.2.7

GeoVisA11y: An AI-based Geovisualization Question-Answering System for Screen-Reader Users

Authors: Chu Li, Rock Yuren Pang, Arnavi Chheda-Kothary, Ather Sharif, Henok Assalif, Jeffrey Heer, Jon E. Froehlich

Abstract: Geovisualizations are powerful tools for communicating spatial information, but are inaccessible to screen-reader users. To address this limitation, we present GeoVisA11y, an LLM-based question-answering system that makes geovisualizations accessible through natural language interaction. The system supports map reading, analysis, interpretation and navigation by handling analytical, geospatial, vi… ▽ More Geovisualizations are powerful tools for communicating spatial information, but are inaccessible to screen-reader users. To address this limitation, we present GeoVisA11y, an LLM-based question-answering system that makes geovisualizations accessible through natural language interaction. The system supports map reading, analysis, interpretation and navigation by handling analytical, geospatial, visual and contextual queries. Through user studies with 12 screen-reader users and sighted participants, we demonstrate that GeoVisA11y effectively bridges accessibility gaps while revealing distinct interaction patterns between user groups. We contribute: (1) an open-source, accessible geovisualization system, (2) empirical findings on query and navigation differences, and (3) a dataset of geospatial queries to inform future research on accessible data visualization. △ Less

Submitted 7 March, 2026; originally announced March 2026.

Comments: This manuscript has been accepted at CHI'26

On the Regulatory Potential of User Interfaces for AI Agent Governance

Authors: K. J. Kevin Feng, Tae Soo Kim, Rock Yuren Pang, Faria Huq, Tal August, Amy X. Zhang

Abstract: AI agents that take actions in their environment autonomously over extended time horizons require robust governance interventions to curb their potentially consequential risks. Prior proposals for governing AI agents primarily target system-level safeguards (e.g., prompt injection monitors) or agent infrastructure (e.g., agent IDs). In this work, we explore a complementary approach: regulating use… ▽ More AI agents that take actions in their environment autonomously over extended time horizons require robust governance interventions to curb their potentially consequential risks. Prior proposals for governing AI agents primarily target system-level safeguards (e.g., prompt injection monitors) or agent infrastructure (e.g., agent IDs). In this work, we explore a complementary approach: regulating user interfaces of AI agents as a way of enforcing transparency and behavioral requirements that then demand changes at the system and/or infrastructure levels. Specifically, we analyze 22 existing agentic systems to identify UI elements that play key roles in human-agent interaction and communication. We then synthesize those elements into six high-level interaction design patterns that hold regulatory potential (e.g., requiring agent memory to be editable). We conclude with policy recommendations based on our analysis. Our work exposes a new surface for regulatory action that supplements previous proposals for practical AI agent governance. △ Less

Submitted 30 November, 2025; originally announced December 2025.

Comments: RegML workshop at NeurIPS 2025 (oral)

AdvancedIF: Rubric-Based Benchmarking and Reinforcement Learning for Advancing LLM Instruction Following

Authors: Yun He, Wenzhe Li, Hejia Zhang, Songlin Li, Karishma Mandyam, Sopan Khosla, Yuanhao Xiong, Nanshu Wang, Xiaoliang Peng, Beibin Li, Shengjie Bi, Shishir G. Patil, Qi Qi, Shengyu Feng, Julian Katz-Samuels, Richard Yuanzhe Pang, Sujan Gonugondla, Hunter Lang, Yue Yu, Yundi Qian, Maryam Fazel-Zarandi, Licheng Yu, Amine Benhalloum, Hany Awadalla, Manaal Faruqui

Abstract: Recent progress in large language models (LLMs) has led to impressive performance on a range of tasks, yet advanced instruction following (IF)-especially for complex, multi-turn, and system-prompted instructions-remains a significant challenge. Rigorous evaluation and effective training for such capabilities are hindered by the lack of high-quality, human-annotated benchmarks and reliable, interpr… ▽ More Recent progress in large language models (LLMs) has led to impressive performance on a range of tasks, yet advanced instruction following (IF)-especially for complex, multi-turn, and system-prompted instructions-remains a significant challenge. Rigorous evaluation and effective training for such capabilities are hindered by the lack of high-quality, human-annotated benchmarks and reliable, interpretable reward signals. In this work, we introduce AdvancedIF (we will release this benchmark soon), a comprehensive benchmark featuring over 1,600 prompts and expert-curated rubrics that assess LLMs ability to follow complex, multi-turn, and system-level instructions. We further propose RIFL (Rubric-based Instruction-Following Learning), a novel post-training pipeline that leverages rubric generation, a finetuned rubric verifier, and reward shaping to enable effective reinforcement learning for instruction following. Extensive experiments demonstrate that RIFL substantially improves the instruction-following abilities of LLMs, achieving a 6.7% absolute gain on AdvancedIF and strong results on public benchmarks. Our ablation studies confirm the effectiveness of each component in RIFL. This work establishes rubrics as a powerful tool for both training and evaluating advanced IF in LLMs, paving the way for more capable and reliable AI systems. △ Less

Submitted 26 November, 2025; v1 submitted 13 November, 2025; originally announced November 2025.

Prompt Curriculum Learning for Efficient LLM Post-Training

Authors: Zhaolin Gao, Joongwon Kim, Wen Sun, Thorsten Joachims, Sid Wang, Richard Yuanzhe Pang, Liang Tan

Abstract: We introduce Prompt Curriculum Learning (PCL), a lightweight reinforcement learning (RL) algorithm that selects intermediate-difficulty prompts using a learned value model to post-train language models. Since post-training LLMs via RL remains sensitive to batching and prompt selection strategies, we first conduct a series of systematic experiments where we (1) determine the optimal training batch… ▽ More We introduce Prompt Curriculum Learning (PCL), a lightweight reinforcement learning (RL) algorithm that selects intermediate-difficulty prompts using a learned value model to post-train language models. Since post-training LLMs via RL remains sensitive to batching and prompt selection strategies, we first conduct a series of systematic experiments where we (1) determine the optimal training batch size that balances generation efficiency and gradient quality and (2) establish the importance of focusing on prompts of intermediate difficulty for the policy. We build upon these results to design PCL, which identifies prompts of intermediate difficulty for the current policy in an on-policy manner by using a value model that is concurrently updated based on the current policy. By focusing on informative prompts that yield high effective ratios, PCL achieves either the highest performance or requires significantly less time to reach comparable performance to its counterparts. Compared to rollout-based filtering methods, PCL avoids costly rollouts and achieves $12.1\times$ and $16.9\times$ faster speed on identifying intermediate-difficulty prompts when training on MATH and DeepScaleR, respectively. We further demonstrate that our value model accurately predicts prompt difficulty and allows PCL to focus on progressively more challenging prompts during RL. Our results present a new methodology that delivers improved tradeoff between upper-bound performance and efficiency for reasoning-focused RL. △ Less

Submitted 1 October, 2025; originally announced October 2025.

Advancing Data Equity: Practitioner Responsibility and Accountability in NLP Data Practices

Authors: Jay L. Cunningham, Kevin Zhongyang Shao, Rock Yuren Pang, Nathaniel Mengist

Abstract: While research has focused on surfacing and auditing algorithmic bias to ensure equitable AI development, less is known about how NLP practitioners - those directly involved in dataset development, annotation, and deployment - perceive and navigate issues of NLP data equity. This study is among the first to center practitioners' perspectives, linking their experiences to a multi-scalar AI governan… ▽ More While research has focused on surfacing and auditing algorithmic bias to ensure equitable AI development, less is known about how NLP practitioners - those directly involved in dataset development, annotation, and deployment - perceive and navigate issues of NLP data equity. This study is among the first to center practitioners' perspectives, linking their experiences to a multi-scalar AI governance framework and advancing participatory recommendations that bridge technical, policy, and community domains. Drawing on a 2024 questionnaire and focus group, we examine how U.S.-based NLP data practitioners conceptualize fairness, contend with organizational and systemic constraints, and engage emerging governance efforts such as the U.S. AI Bill of Rights. Findings reveal persistent tensions between commercial objectives and equity commitments, alongside calls for more participatory and accountable data workflows. We critically engage debates on data diversity and diversity washing, arguing that improving NLP equity requires structural governance reforms that support practitioner agency and community consent. △ Less

Submitted 16 August, 2025; v1 submitted 13 August, 2025; originally announced August 2025.

Comments: 10 pages, 6 Pages (References and Appendices). The archival version has been accepted to AAAI (AIES 2025) without the extended Appendices. This extended version includes Appendices

Interactive Reasoning: Visualizing and Controlling Chain-of-Thought Reasoning in Large Language Models

Authors: Rock Yuren Pang, K. J. Kevin Feng, Shangbin Feng, Chu Li, Weijia Shi, Yulia Tsvetkov, Jeffrey Heer, Katharina Reinecke

Abstract: The output quality of large language models (LLMs) can be improved via "reasoning": generating segments of chain-of-thought (CoT) content to further condition the model prior to producing user-facing output. While these chains contain valuable information, they are verbose and lack explicit organization, making them tedious to review. Moreover, they lack opportunities for user feedback, such as to… ▽ More The output quality of large language models (LLMs) can be improved via "reasoning": generating segments of chain-of-thought (CoT) content to further condition the model prior to producing user-facing output. While these chains contain valuable information, they are verbose and lack explicit organization, making them tedious to review. Moreover, they lack opportunities for user feedback, such as to remove unwanted considerations, add desired ones, or clarify unclear assumptions. We introduce Interactive Reasoning, an interaction design that visualizes chain-of-thought outputs as a hierarchy of topics and enables user review and modification. We implement interactive reasoning in Hippo, a prototype for AI-assisted decision making in the face of uncertain trade-offs. In a user study with 16 participants, we find that interactive reasoning in Hippo allows users to quickly identify and interrupt erroneous generations, efficiently steer the model towards customized responses, and better understand both model reasoning and model outputs. Our work contributes to a new paradigm that incorporates user oversight into LLM reasoning processes. △ Less

Submitted 30 June, 2025; originally announced June 2025.

Passing the Buck to AI: How Individuals' Decision-Making Patterns Affect Reliance on AI

Authors: Katelyn Xiaoying Mei, Rock Yuren Pang, Alex Lyford, Lucy Lu Wang, Katharina Reinecke

Abstract: Psychological research has identified different patterns individuals have while making decisions, such as vigilance (making decisions after thorough information gathering), hypervigilance (rushed and anxious decision-making), and buckpassing (deferring decisions to others). We examine whether these decision-making patterns shape peoples' likelihood of seeking out or relying on AI. In an online exp… ▽ More Psychological research has identified different patterns individuals have while making decisions, such as vigilance (making decisions after thorough information gathering), hypervigilance (rushed and anxious decision-making), and buckpassing (deferring decisions to others). We examine whether these decision-making patterns shape peoples' likelihood of seeking out or relying on AI. In an online experiment with 810 participants tasked with distinguishing food facts from myths, we found that a higher buckpassing tendency was positively correlated with both seeking out and relying on AI suggestions, while being negatively correlated with the time spent reading AI explanations. In contrast, the higher a participant tended towards vigilance, the more carefully they scrutinized the AI's information, as indicated by an increased time spent looking through the AI's explanations. These findings suggest that a person's decision-making pattern plays a significant role in their adoption and reliance on AI, which provides a new understanding of individual differences in AI-assisted decision-making. △ Less

Submitted 22 January, 2026; v1 submitted 2 May, 2025; originally announced May 2025.

OLMoTrace: Tracing Language Model Outputs Back to Trillions of Training Tokens

Authors: Jiacheng Liu, Taylor Blanton, Yanai Elazar, Sewon Min, YenSung Chen, Arnavi Chheda-Kothary, Huy Tran, Byron Bischoff, Eric Marsh, Michael Schmitz, Cassidy Trier, Aaron Sarnat, Jenna James, Jon Borchardt, Bailey Kuehl, Evie Cheng, Karen Farley, Sruthi Sreeram, Taira Anderson, David Albright, Carissa Schoenick, Luca Soldaini, Dirk Groeneveld, Rock Yuren Pang, Pang Wei Koh , et al. (6 additional authors not shown)

Abstract: We present OLMoTrace, the first system that traces the outputs of language models back to their full, multi-trillion-token training data in real time. OLMoTrace finds and shows verbatim matches between segments of language model output and documents in the training text corpora. Powered by an extended version of infini-gram (Liu et al., 2024), our system returns tracing results within a few second… ▽ More We present OLMoTrace, the first system that traces the outputs of language models back to their full, multi-trillion-token training data in real time. OLMoTrace finds and shows verbatim matches between segments of language model output and documents in the training text corpora. Powered by an extended version of infini-gram (Liu et al., 2024), our system returns tracing results within a few seconds. OLMoTrace can help users understand the behavior of language models through the lens of their training data. We showcase how it can be used to explore fact checking, hallucination, and the creativity of language models. OLMoTrace is publicly available and fully open-source. △ Less

Submitted 7 July, 2025; v1 submitted 9 April, 2025; originally announced April 2025.

Comments: ACL 2025 demo track

Accessibility for Whom? Perceptions of Sidewalk Barriers Across Disability Groups and Implications for Designing Personalized Maps

Authors: Chu Li, Rock Yuren Pang, Delphine Labbé, Yochai Eisenberg, Maryam Hosseini, Jon E. Froehlich

Abstract: Despite diverse mobility needs worldwide, existing mapping tools fail to address the varied experiences of different mobility device users. This paper presents a large-scale online survey exploring how five mobility groups -- users of canes, walkers, mobility scooters, manual wheelchairs, and motorized wheelchairs -- perceive sidewalk barriers. Using 52 sidewalk barrier images, respondents evaluat… ▽ More Despite diverse mobility needs worldwide, existing mapping tools fail to address the varied experiences of different mobility device users. This paper presents a large-scale online survey exploring how five mobility groups -- users of canes, walkers, mobility scooters, manual wheelchairs, and motorized wheelchairs -- perceive sidewalk barriers. Using 52 sidewalk barrier images, respondents evaluated their confidence in navigating each scenario. Our findings (N=190) reveal variations in barrier perceptions across groups, while also identifying shared concerns. To further demonstrate the value of this data, we showcase its use in two custom prototypes: a visual analytics tool and a personalized routing tool. Our survey findings and open dataset advance work in accessibility-focused maps, routing algorithms, and urban planning. △ Less

Submitted 27 February, 2025; originally announced February 2025.

Comments: Manuscript accepted at CHI'25

Understanding the LLM-ification of CHI: Unpacking the Impact of LLMs at CHI through a Systematic Literature Review

Authors: Rock Yuren Pang, Hope Schroeder, Kynnedy Simone Smith, Solon Barocas, Ziang Xiao, Emily Tseng, Danielle Bragg

Abstract: Large language models (LLMs) have been positioned to revolutionize HCI, by reshaping not only the interfaces, design patterns, and sociotechnical systems that we study, but also the research practices we use. To-date, however, there has been little understanding of LLMs' uptake in HCI. We address this gap via a systematic literature review of 153 CHI papers from 2020-24 that engage with LLMs. We t… ▽ More Large language models (LLMs) have been positioned to revolutionize HCI, by reshaping not only the interfaces, design patterns, and sociotechnical systems that we study, but also the research practices we use. To-date, however, there has been little understanding of LLMs' uptake in HCI. We address this gap via a systematic literature review of 153 CHI papers from 2020-24 that engage with LLMs. We taxonomize: (1) domains where LLMs are applied; (2) roles of LLMs in HCI projects; (3) contribution types; and (4) acknowledged limitations and risks. We find LLM work in 10 diverse domains, primarily via empirical and artifact contributions. Authors use LLMs in five distinct roles, including as research tools or simulated users. Still, authors often raise validity and reproducibility concerns, and overwhelmingly study closed models. We outline opportunities to improve HCI research with and on LLMs, and provide guiding questions for researchers to consider the validity and appropriateness of LLM-related work. △ Less

Submitted 21 January, 2025; originally announced January 2025.

Comments: This is a preprint version of the paper conditionally accepted to CHI'25

Transformers Struggle to Learn to Search

Authors: Abulhair Saparov, Srushti Pawar, Shreyas Pimpalgaonkar, Nitish Joshi, Richard Yuanzhe Pang, Vishakh Padmakumar, Seyed Mehran Kazemi, Najoung Kim, He He

Abstract: Search is an ability foundational in many important tasks, and recent studies have shown that large language models (LLMs) struggle to perform search robustly. It is unknown whether this inability is due to a lack of data, insufficient model parameters, or fundamental limitations of the transformer architecture. In this work, we use the foundational graph connectivity problem as a testbed to gener… ▽ More Search is an ability foundational in many important tasks, and recent studies have shown that large language models (LLMs) struggle to perform search robustly. It is unknown whether this inability is due to a lack of data, insufficient model parameters, or fundamental limitations of the transformer architecture. In this work, we use the foundational graph connectivity problem as a testbed to generate effectively limitless high-coverage data to train small transformers and test whether they can learn to perform search. We find that, when given the right training distribution, the transformer is able to learn to search. We analyze the algorithm that the transformer has learned through a novel mechanistic interpretability technique that enables us to extract the computation graph from the trained model. We find that transformers perform search at every vertex in parallel: For each vertex in the input graph, transformers compute the set of vertices reachable from that vertex. Each layer then progressively expands these sets, allowing the model to search over a number of vertices exponential in $n_{\text{layers}}$. However, we find that as the input graph size increases, the transformer has greater difficulty in learning the task. This difficulty is not resolved even as the number of parameters is increased, suggesting that increasing model scale will not lead to robust search abilities. We also find that performing search in-context (i.e., chain-of-thought) does not resolve this inability to learn to search on larger graphs. △ Less

Submitted 16 March, 2025; v1 submitted 5 December, 2024; originally announced December 2024.

Comments: Published as a conference paper at ICLR 2025

Self-Generated Critiques Boost Reward Modeling for Language Models

Authors: Yue Yu, Zhengxing Chen, Aston Zhang, Liang Tan, Chenguang Zhu, Richard Yuanzhe Pang, Yundi Qian, Xuewei Wang, Suchin Gururangan, Chao Zhang, Melanie Kambadur, Dhruv Mahajan, Rui Hou

Abstract: Reward modeling is crucial for aligning large language models (LLMs) with human preferences, especially in reinforcement learning from human feedback (RLHF). However, current reward models mainly produce scalar scores and struggle to incorporate critiques in a natural language format. We hypothesize that predicting both critiques and the scalar reward would improve reward modeling ability. Motivat… ▽ More Reward modeling is crucial for aligning large language models (LLMs) with human preferences, especially in reinforcement learning from human feedback (RLHF). However, current reward models mainly produce scalar scores and struggle to incorporate critiques in a natural language format. We hypothesize that predicting both critiques and the scalar reward would improve reward modeling ability. Motivated by this, we propose Critic-RM, a framework that improves reward models using self-generated critiques without extra supervision. Critic-RM employs a two-stage process: generating and filtering high-quality critiques, followed by joint fine-tuning on reward prediction and critique generation. Experiments across benchmarks show that Critic-RM improves reward modeling accuracy by 3.7%-7.3% compared to standard reward models and LLM judges, demonstrating strong performance and data efficiency. Additional studies further validate the effectiveness of generated critiques in rectifying flawed reasoning steps with 2.5%-3.2% gains in improving reasoning accuracy. △ Less

Submitted 9 February, 2025; v1 submitted 25 November, 2024; originally announced November 2024.

Comments: Accepted to NAACL 2025 (Main Conference)

Journal ref: NAACL 2025

Self-Consistency Preference Optimization

Authors: Archiki Prasad, Weizhe Yuan, Richard Yuanzhe Pang, Jing Xu, Maryam Fazel-Zarandi, Mohit Bansal, Sainbayar Sukhbaatar, Jason Weston, Jane Yu

Abstract: Self-alignment, whereby models learn to improve themselves without human annotation, is a rapidly growing research area. However, existing techniques often fail to improve complex reasoning tasks due to the difficulty of assigning correct rewards. An orthogonal approach that is known to improve correctness is self-consistency, a method applied at inference time based on multiple sampling in order… ▽ More Self-alignment, whereby models learn to improve themselves without human annotation, is a rapidly growing research area. However, existing techniques often fail to improve complex reasoning tasks due to the difficulty of assigning correct rewards. An orthogonal approach that is known to improve correctness is self-consistency, a method applied at inference time based on multiple sampling in order to find the most consistent answer. In this work, we extend the self-consistency concept to help train models. We thus introduce self-consistency preference optimization (ScPO), which iteratively trains consistent answers to be preferred over inconsistent ones on unsupervised new problems. We show ScPO leads to large improvements over conventional reward model training on reasoning tasks such as GSM8K and MATH, closing the gap with supervised training with gold answers or preferences, and that combining ScPO with standard supervised learning improves results even further. On ZebraLogic, ScPO finetunes Llama-3 8B to be superior to Llama-3 70B, Gemma-2 27B, and Claude-3 Haiku. △ Less

Submitted 6 July, 2025; v1 submitted 6 November, 2024; originally announced November 2024.

Comments: ICML 2025 (camera-ready)

Self-Taught Evaluators

Authors: Tianlu Wang, Ilia Kulikov, Olga Golovneva, Ping Yu, Weizhe Yuan, Jane Dwivedi-Yu, Richard Yuanzhe Pang, Maryam Fazel-Zarandi, Jason Weston, Xian Li

Abstract: Model-based evaluation is at the heart of successful model development -- as a reward model for training, and as a replacement for human evaluation. To train such evaluators, the standard approach is to collect a large amount of human preference judgments over model responses, which is costly and the data becomes stale as models improve. In this work, we present an approach that aims to im-prove e… ▽ More Model-based evaluation is at the heart of successful model development -- as a reward model for training, and as a replacement for human evaluation. To train such evaluators, the standard approach is to collect a large amount of human preference judgments over model responses, which is costly and the data becomes stale as models improve. In this work, we present an approach that aims to im-prove evaluators without human annotations, using synthetic training data only. Starting from unlabeled instructions, our iterative self-improvement scheme generates contrasting model outputs and trains an LLM-as-a-Judge to produce reasoning traces and final judgments, repeating this training at each new iteration using the improved predictions. Without any labeled preference data, our Self-Taught Evaluator can improve a strong LLM (Llama3-70B-Instruct) from 75.4 to 88.3 (88.7 with majority vote) on RewardBench. This outperforms commonly used LLM judges such as GPT-4 and matches the performance of the top-performing reward models trained with labeled examples. △ Less

Submitted 8 August, 2024; v1 submitted 5 August, 2024; originally announced August 2024.

AltGeoViz: Facilitating Accessible Geovisualization

Authors: Chu Li, Rock Yuren Pang, Ather Sharif, Arnavi Chheda-Kothary, Jeffrey Heer, Jon E. Froehlich

Abstract: Geovisualizations are powerful tools for exploratory spatial analysis, enabling sighted users to discern patterns, trends, and relationships within geographic data. However, these visual tools have remained largely inaccessible to screen-reader users. We present AltGeoViz, a new system we designed to facilitate geovisualization exploration for these users. AltGeoViz dynamically generates alt-text… ▽ More Geovisualizations are powerful tools for exploratory spatial analysis, enabling sighted users to discern patterns, trends, and relationships within geographic data. However, these visual tools have remained largely inaccessible to screen-reader users. We present AltGeoViz, a new system we designed to facilitate geovisualization exploration for these users. AltGeoViz dynamically generates alt-text descriptions based on the user's current map view, providing summaries of spatial patterns and descriptive statistics. In a study of five screen-reader users, we found that AltGeoViz enabled them to interact with geovisualizations in previously infeasible ways. Participants demonstrated a clear understanding of data summaries and their location context, and they could synthesize spatial understandings of their explorations. Moreover, we identified key areas for improvement, such as the addition of intuitive spatial navigation controls and comparative analysis features. △ Less

Submitted 9 December, 2024; v1 submitted 19 June, 2024; originally announced June 2024.

Comments: IEEE VIS 2024

An Introduction to Vision-Language Modeling

Authors: Florian Bordes, Richard Yuanzhe Pang, Anurag Ajay, Alexander C. Li, Adrien Bardes, Suzanne Petryk, Oscar Mañas, Zhiqiu Lin, Anas Mahmoud, Bargav Jayaraman, Mark Ibrahim, Melissa Hall, Yunyang Xiong, Jonathan Lebensold, Candace Ross, Srihari Jayakumar, Chuan Guo, Diane Bouchacourt, Haider Al-Tahan, Karthik Padthe, Vasu Sharma, Hu Xu, Xiaoqing Ellen Tan, Megan Richards, Samuel Lavoie , et al. (16 additional authors not shown)

Abstract: Following the recent popularity of Large Language Models (LLMs), several attempts have been made to extend them to the visual domain. From having a visual assistant that could guide us through unfamiliar environments to generative models that produce images using only a high-level text description, the vision-language model (VLM) applications will significantly impact our relationship with technol… ▽ More Following the recent popularity of Large Language Models (LLMs), several attempts have been made to extend them to the visual domain. From having a visual assistant that could guide us through unfamiliar environments to generative models that produce images using only a high-level text description, the vision-language model (VLM) applications will significantly impact our relationship with technology. However, there are many challenges that need to be addressed to improve the reliability of those models. While language is discrete, vision evolves in a much higher dimensional space in which concepts cannot always be easily discretized. To better understand the mechanics behind mapping vision to language, we present this introduction to VLMs which we hope will help anyone who would like to enter the field. First, we introduce what VLMs are, how they work, and how to train them. Then, we present and discuss approaches to evaluate VLMs. Although this work primarily focuses on mapping images to language, we also discuss extending VLMs to videos. △ Less

Submitted 27 May, 2024; originally announced May 2024.

BLIP: Facilitating the Exploration of Undesirable Consequences of Digital Technologies

Authors: Rock Yuren Pang, Sebastin Santy, René Just, Katharina Reinecke

Abstract: Digital technologies have positively transformed society, but they have also led to undesirable consequences not anticipated at the time of design or development. We posit that insights into past undesirable consequences can help researchers and practitioners gain awareness and anticipate potential adverse effects. To test this assumption, we introduce BLIP, a system that extracts real-world undes… ▽ More Digital technologies have positively transformed society, but they have also led to undesirable consequences not anticipated at the time of design or development. We posit that insights into past undesirable consequences can help researchers and practitioners gain awareness and anticipate potential adverse effects. To test this assumption, we introduce BLIP, a system that extracts real-world undesirable consequences of technology from online articles, summarizes and categorizes them, and presents them in an interactive, web-based interface. In two user studies with 15 researchers in various computer science disciplines, we found that BLIP substantially increased the number and diversity of undesirable consequences they could list in comparison to relying on prior knowledge or searching online. Moreover, BLIP helped them identify undesirable consequences relevant to their ongoing projects, made them aware of undesirable consequences they "had never considered," and inspired them to reflect on their own experiences with technology. △ Less

Submitted 10 May, 2024; originally announced May 2024.

Comments: To appear in the Proceedings of the CHI Conference on Human Factors in Computing Systems (CHI '24), May 11--16, 2024, Honolulu, HI, USA

Iterative Reasoning Preference Optimization

Authors: Richard Yuanzhe Pang, Weizhe Yuan, Kyunghyun Cho, He He, Sainbayar Sukhbaatar, Jason Weston

Abstract: Iterative preference optimization methods have recently been shown to perform well for general instruction tuning tasks, but typically make little improvement on reasoning tasks (Yuan et al., 2024, Chen et al., 2024). In this work we develop an iterative approach that optimizes the preference between competing generated Chain-of-Thought (CoT) candidates by optimizing for winning vs. losing reasoni… ▽ More Iterative preference optimization methods have recently been shown to perform well for general instruction tuning tasks, but typically make little improvement on reasoning tasks (Yuan et al., 2024, Chen et al., 2024). In this work we develop an iterative approach that optimizes the preference between competing generated Chain-of-Thought (CoT) candidates by optimizing for winning vs. losing reasoning steps that lead to the correct answer. We train using a modified DPO loss (Rafailov et al., 2023) with an additional negative log-likelihood term, which we find to be crucial. We show reasoning improves across repeated iterations of this scheme. While only relying on examples in the training set, our approach results in increasing accuracy on GSM8K, MATH, and ARC-Challenge for Llama-2-70B-Chat, outperforming other Llama-2-based models not relying on additionally sourced datasets. For example, we see a large improvement from 55.6% to 81.6% on GSM8K and an accuracy of 88.7% with majority voting out of 32 samples. △ Less

Submitted 25 June, 2024; v1 submitted 30 April, 2024; originally announced April 2024.

Self-Rewarding Language Models

Authors: Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, Jason Weston

Abstract: We posit that to achieve superhuman agents, future models require superhuman feedback in order to provide an adequate training signal. Current approaches commonly train reward models from human preferences, which may then be bottlenecked by human performance level, and secondly these separate frozen reward models cannot then learn to improve during LLM training. In this work, we study Self-Rewardi… ▽ More We posit that to achieve superhuman agents, future models require superhuman feedback in order to provide an adequate training signal. Current approaches commonly train reward models from human preferences, which may then be bottlenecked by human performance level, and secondly these separate frozen reward models cannot then learn to improve during LLM training. In this work, we study Self-Rewarding Language Models, where the language model itself is used via LLM-as-a-Judge prompting to provide its own rewards during training. We show that during Iterative DPO training that not only does instruction following ability improve, but also the ability to provide high-quality rewards to itself. Fine-tuning Llama 2 70B on three iterations of our approach yields a model that outperforms many existing systems on the AlpacaEval 2.0 leaderboard, including Claude 2, Gemini Pro, and GPT-4 0613. While there is much left still to explore, this work opens the door to the possibility of models that can continually improve in both axes. △ Less

Submitted 27 March, 2025; v1 submitted 18 January, 2024; originally announced January 2024.

Comments: ICML 2024

GPQA: A Graduate-Level Google-Proof Q&A Benchmark

Authors: David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, Samuel R. Bowman

Abstract: We present GPQA, a challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry. We ensure that the questions are high-quality and extremely difficult: experts who have or are pursuing PhDs in the corresponding domains reach 65% accuracy (74% when discounting clear mistakes the experts identified in retrospect), while highly skilled non-expert v… ▽ More We present GPQA, a challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry. We ensure that the questions are high-quality and extremely difficult: experts who have or are pursuing PhDs in the corresponding domains reach 65% accuracy (74% when discounting clear mistakes the experts identified in retrospect), while highly skilled non-expert validators only reach 34% accuracy, despite spending on average over 30 minutes with unrestricted access to the web (i.e., the questions are "Google-proof"). The questions are also difficult for state-of-the-art AI systems, with our strongest GPT-4 based baseline achieving 39% accuracy. If we are to use future AI systems to help us answer very hard questions, for example, when developing new scientific knowledge, we need to develop scalable oversight methods that enable humans to supervise their outputs, which may be difficult even if the supervisors are themselves skilled and knowledgeable. The difficulty of GPQA both for skilled non-experts and frontier AI systems should enable realistic scalable oversight experiments, which we hope can help devise ways for human experts to reliably get truthful information from AI systems that surpass human capabilities. △ Less

Submitted 20 November, 2023; originally announced November 2023.

Comments: 28 pages, 5 figures, 7 tables

The Case for Anticipating Undesirable Consequences of Computing Innovations Early, Often, and Across Computer Science

Authors: Rock Yuren Pang, Dan Grossman, Tadayoshi Kohno, Katharina Reinecke

Abstract: From smart sensors that infringe on our privacy to neural nets that portray realistic imposter deepfakes, our society increasingly bears the burden of negative, if unintended, consequences of computing innovations. As the experts in the technology we create, Computer Science (CS) researchers must do better at anticipating and addressing these undesirable consequences proactively. Our prior work sh… ▽ More From smart sensors that infringe on our privacy to neural nets that portray realistic imposter deepfakes, our society increasingly bears the burden of negative, if unintended, consequences of computing innovations. As the experts in the technology we create, Computer Science (CS) researchers must do better at anticipating and addressing these undesirable consequences proactively. Our prior work showed that many of us recognize the value of thinking preemptively about the perils our research can pose, yet we tend to address them only in hindsight. How can we change the culture in which considering undesirable consequences of digital technology is deemed as important, but is not commonly done? △ Less

Submitted 8 September, 2023; originally announced September 2023.

Comments: More details at NSF #2315937: https://www.nsf.gov/awardsearch/showAward?AWD_ID=2315937&HistoricalAwards=false

Leveraging Implicit Feedback from Deployment Data in Dialogue

Authors: Richard Yuanzhe Pang, Stephen Roller, Kyunghyun Cho, He He, Jason Weston

Abstract: We study improving social conversational agents by learning from natural dialogue between users and a deployed model, without extra annotations. To implicitly measure the quality of a machine-generated utterance, we leverage signals like user response length, sentiment and reaction of the future human utterances in the collected dialogue episodes. Our experiments use the publicly released deployme… ▽ More We study improving social conversational agents by learning from natural dialogue between users and a deployed model, without extra annotations. To implicitly measure the quality of a machine-generated utterance, we leverage signals like user response length, sentiment and reaction of the future human utterances in the collected dialogue episodes. Our experiments use the publicly released deployment data from BlenderBot (Xu et al., 2023). Human evaluation indicates improvements in our new models over baseline responses; however, we find that some proxy signals can lead to more generations with undesirable properties as well. For example, optimizing for conversation length can lead to more controversial or unfriendly generations compared to the baseline, whereas optimizing for positive sentiment or reaction can decrease these behaviors. △ Less

Submitted 31 January, 2024; v1 submitted 26 July, 2023; originally announced July 2023.

Comments: EACL 2024

Testing the General Deductive Reasoning Capacity of Large Language Models Using OOD Examples

Authors: Abulhair Saparov, Richard Yuanzhe Pang, Vishakh Padmakumar, Nitish Joshi, Seyed Mehran Kazemi, Najoung Kim, He He

Abstract: Given the intractably large size of the space of proofs, any model that is capable of general deductive reasoning must generalize to proofs of greater complexity. Recent studies have shown that large language models (LLMs) possess some abstract deductive reasoning ability given chain-of-thought prompts. However, they have primarily been tested on proofs using modus ponens or of a specific size, an… ▽ More Given the intractably large size of the space of proofs, any model that is capable of general deductive reasoning must generalize to proofs of greater complexity. Recent studies have shown that large language models (LLMs) possess some abstract deductive reasoning ability given chain-of-thought prompts. However, they have primarily been tested on proofs using modus ponens or of a specific size, and from the same distribution as the in-context examples. To measure the general deductive reasoning ability of LLMs, we test on a broad set of deduction rules and measure their ability to generalize to more complex proofs from simpler demonstrations from multiple angles: depth-, width-, and compositional generalization. To facilitate systematic exploration, we construct a new synthetic and programmable reasoning dataset that enables control over deduction rules and proof complexity. Our experiments on four LLMs of various sizes and training objectives show that they are able to generalize to compositional proofs. However, they have difficulty generalizing to longer proofs, and they require explicit demonstrations to produce hypothetical subproofs, specifically in proof by cases and proof by contradiction. △ Less

Submitted 3 November, 2023; v1 submitted 24 May, 2023; originally announced May 2023.

Comments: Published as a conference paper at NeurIPS 2023

Auditing Cross-Cultural Consistency of Human-Annotated Labels for Recommendation Systems

Authors: Rock Yuren Pang, Jack Cenatempo, Franklyn Graham, Bridgette Kuehn, Maddy Whisenant, Portia Botchway, Katie Stone Perez, Allison Koenecke

Abstract: Recommendation systems increasingly depend on massive human-labeled datasets; however, the human annotators hired to generate these labels increasingly come from homogeneous backgrounds. This poses an issue when downstream predictive models -- based on these labels -- are applied globally to a heterogeneous set of users. We study this disconnect with respect to the labels themselves, asking whethe… ▽ More Recommendation systems increasingly depend on massive human-labeled datasets; however, the human annotators hired to generate these labels increasingly come from homogeneous backgrounds. This poses an issue when downstream predictive models -- based on these labels -- are applied globally to a heterogeneous set of users. We study this disconnect with respect to the labels themselves, asking whether they are ``consistently conceptualized'' across annotators of different demographics. In a case study of video game labels, we conduct a survey on 5,174 gamers, identify a subset of inconsistently conceptualized game labels, perform causal analyses, and suggest both cultural and linguistic reasons for cross-country differences in label annotation. We further demonstrate that predictive models of game annotations perform better on global train sets as opposed to homogeneous (single-country) train sets. Finally, we provide a generalizable framework for practitioners to audit their own data annotation processes for consistent label conceptualization, and encourage practitioners to consider global inclusivity in recommendation systems starting from the early stages of annotator recruitment and data-labeling. △ Less

Submitted 10 May, 2023; originally announced May 2023.

Comments: Accepted at FAccT 2023

Anticipating Unintended Consequences of Technology Using Insights from Creativity Support Tools

Authors: Rock Yuren Pang, Katharina Reinecke

Abstract: Our society has been increasingly witnessing a number of negative, unintended consequences of digital technologies. While post-hoc policy regulation is crucial in addressing these issues, reasonably anticipating the consequences before deploying technology can help mitigate potential harm to society in the first place. Yet, the quest to anticipate potential harms can be difficult without seeing di… ▽ More Our society has been increasingly witnessing a number of negative, unintended consequences of digital technologies. While post-hoc policy regulation is crucial in addressing these issues, reasonably anticipating the consequences before deploying technology can help mitigate potential harm to society in the first place. Yet, the quest to anticipate potential harms can be difficult without seeing digital technologies deployed in the real world. In this position paper, we argue that anticipating unintended consequences of technology can be facilitated through creativity-enhancing interventions, such as by building on existing knowledge and insights from diverse stakeholders. Using lessons learned from prior work on creativity-support tools, the HCI community is uniquely equipped to design novel systems that aid in anticipating negative unintended consequences of technology on society. △ Less

Submitted 12 April, 2023; originally announced April 2023.

Comments: In CHI '23 Workshop on Designing Technology and Policy Simultaneously: Towards A Research Agenda and New Practice, April 23, 2023

"That's important, but...": How Computer Science Researchers Anticipate Unintended Consequences of Their Research Innovations

Authors: Kimberly Do, Rock Yuren Pang, Jiachen Jiang, Katharina Reinecke

Abstract: Computer science research has led to many breakthrough innovations but has also been scrutinized for enabling technology that has negative, unintended consequences for society. Given the increasing discussions of ethics in the news and among researchers, we interviewed 20 researchers in various CS sub-disciplines to identify whether and how they consider potential unintended consequences of their… ▽ More Computer science research has led to many breakthrough innovations but has also been scrutinized for enabling technology that has negative, unintended consequences for society. Given the increasing discussions of ethics in the news and among researchers, we interviewed 20 researchers in various CS sub-disciplines to identify whether and how they consider potential unintended consequences of their research innovations. We show that considering unintended consequences is generally seen as important but rarely practiced. Principal barriers are a lack of formal process and strategy as well as the academic practice that prioritizes fast progress and publications. Drawing on these findings, we discuss approaches to support researchers in routinely considering unintended consequences, from bringing diverse perspectives through community participation to increasing incentives to investigate potential consequences. We intend for our work to pave the way for routine explorations of the societal implications of technological innovations before, during, and after the research process. △ Less

Submitted 27 March, 2023; originally announced March 2023.

Comments: Corresponding author: Rock Yuren Pang, email provided below. Kimberly Do and Rock Yuren Pang contributed equally to this research. The author order is listed alphabetically. To appear in CHI Conference on Human Factors in Computing Systems (CHI '23), April 23-April 28, 2023, Hamburg, Germany. ACM, New York, NY, USA, 16 pages

Extrapolative Controlled Sequence Generation via Iterative Refinement

Authors: Vishakh Padmakumar, Richard Yuanzhe Pang, He He, Ankur P. Parikh

Abstract: We study the problem of extrapolative controlled generation, i.e., generating sequences with attribute values beyond the range seen in training. This task is of significant importance in automated design, especially drug discovery, where the goal is to design novel proteins that are \textit{better} (e.g., more stable) than existing sequences. Thus, by definition, the target sequences and their att… ▽ More We study the problem of extrapolative controlled generation, i.e., generating sequences with attribute values beyond the range seen in training. This task is of significant importance in automated design, especially drug discovery, where the goal is to design novel proteins that are \textit{better} (e.g., more stable) than existing sequences. Thus, by definition, the target sequences and their attribute values are out of the training distribution, posing challenges to existing methods that aim to directly generate the target sequence. Instead, in this work, we propose Iterative Controlled Extrapolation (ICE) which iteratively makes local edits to a sequence to enable extrapolation. We train the model on synthetically generated sequence pairs that demonstrate small improvement in the attribute value. Results on one natural language task (sentiment analysis) and two protein engineering tasks (ACE2 stability and AAV fitness) show that ICE considerably outperforms state-of-the-art approaches despite its simplicity. Our code and models are available at: https://github.com/vishakhpk/iter-extrapolation. △ Less

Submitted 7 June, 2023; v1 submitted 8 March, 2023; originally announced March 2023.

Comments: ICML 2023 - Camera Ready Version

Reward Gaming in Conditional Text Generation

Authors: Richard Yuanzhe Pang, Vishakh Padmakumar, Thibault Sellam, Ankur P. Parikh, He He

Abstract: To align conditional text generation model outputs with desired behaviors, there has been an increasing focus on training the model using reinforcement learning (RL) with reward functions learned from human annotations. Under this framework, we identify three common cases where high rewards are incorrectly assigned to undesirable patterns: noise-induced spurious correlation, naturally occurring sp… ▽ More To align conditional text generation model outputs with desired behaviors, there has been an increasing focus on training the model using reinforcement learning (RL) with reward functions learned from human annotations. Under this framework, we identify three common cases where high rewards are incorrectly assigned to undesirable patterns: noise-induced spurious correlation, naturally occurring spurious correlation, and covariate shift. We show that even though learned metrics achieve high performance on the distribution of the data used to train the reward function, the undesirable patterns may be amplified during RL training of the text generation model. While there has been discussion about reward gaming in the RL or safety community, in this discussion piece, we would like to highlight reward gaming in the natural language generation (NLG) community using concrete conditional text generation examples and discuss potential fixes and areas for future work. △ Less

Submitted 1 June, 2023; v1 submitted 16 November, 2022; originally announced November 2022.

Comments: ACL 2023

How Do Data Science Workers Communicate Intermediate Results?

Authors: Rock Yuren Pang, Ruotong Wang, Joely Nelson, Leilani Battle

Abstract: Data science workers increasingly collaborate on large-scale projects before communicating insights to a broader audience in the form of visualization. While prior work has modeled how data science teams, oftentimes with distinct roles and work processes, communicate knowledge to outside stakeholders, we have little knowledge of how data science workers communicate intermediately before delivering… ▽ More Data science workers increasingly collaborate on large-scale projects before communicating insights to a broader audience in the form of visualization. While prior work has modeled how data science teams, oftentimes with distinct roles and work processes, communicate knowledge to outside stakeholders, we have little knowledge of how data science workers communicate intermediately before delivering the final products. In this work, we contribute a nuanced description of the intermediate communication process within data science teams. By analyzing interview data with 8 self-identified data science workers, we characterized the data science intermediate communication process with four factors, including the types of audience, communication goals, shared artifacts, and mode of communication. We also identified overarching challenges in the current communication process. We also discussed design implications that might inform better tools that facilitate intermediate communication within data science teams. △ Less

Submitted 6 October, 2022; originally announced October 2022.

Comments: This paper was accepted for presentation as part of the eighth Symposium on Visualization in Data Science (VDS) at ACM KDD 2022 as well as IEEE VIS 2022. http://www.visualdatascience.org/2022/index.html

What Do NLP Researchers Believe? Results of the NLP Community Metasurvey

Authors: Julian Michael, Ari Holtzman, Alicia Parrish, Aaron Mueller, Alex Wang, Angelica Chen, Divyam Madaan, Nikita Nangia, Richard Yuanzhe Pang, Jason Phang, Samuel R. Bowman

Abstract: We present the results of the NLP Community Metasurvey. Run from May to June 2022, the survey elicited opinions on controversial issues, including industry influence in the field, concerns about AGI, and ethics. Our results put concrete numbers to several controversies: For example, respondents are split almost exactly in half on questions about the importance of artificial general intelligence, w… ▽ More We present the results of the NLP Community Metasurvey. Run from May to June 2022, the survey elicited opinions on controversial issues, including industry influence in the field, concerns about AGI, and ethics. Our results put concrete numbers to several controversies: For example, respondents are split almost exactly in half on questions about the importance of artificial general intelligence, whether language models understand language, and the necessity of linguistic structure and inductive bias for solving NLP problems. In addition, the survey posed meta-questions, asking respondents to predict the distribution of survey responses. This allows us not only to gain insight on the spectrum of beliefs held by NLP researchers, but also to uncover false sociological beliefs where the community's predictions don't match reality. We find such mismatches on a wide range of issues. Among other results, the community greatly overestimates its own belief in the usefulness of benchmarks and the potential for scaling to solve real-world problems, while underestimating its own belief in the importance of linguistic structure, inductive bias, and interdisciplinary science. △ Less

Submitted 26 August, 2022; originally announced August 2022.

Comments: 31 pages, 19 figures, 3 tables; more information at https://nlpsurvey.net

ACM Class: I.2.7

SQuALITY: Building a Long-Document Summarization Dataset the Hard Way

Authors: Alex Wang, Richard Yuanzhe Pang, Angelica Chen, Jason Phang, Samuel R. Bowman

Abstract: Summarization datasets are often assembled either by scraping naturally occurring public-domain summaries -- which are nearly always in difficult-to-work-with technical domains -- or by using approximate heuristics to extract them from everyday text -- which frequently yields unfaithful summaries. In this work, we turn to a slower but more straightforward approach to developing summarization bench… ▽ More Summarization datasets are often assembled either by scraping naturally occurring public-domain summaries -- which are nearly always in difficult-to-work-with technical domains -- or by using approximate heuristics to extract them from everyday text -- which frequently yields unfaithful summaries. In this work, we turn to a slower but more straightforward approach to developing summarization benchmark data: We hire highly-qualified contractors to read stories and write original summaries from scratch. To amortize reading time, we collect five summaries per document, with the first giving an overview and the subsequent four addressing specific questions. We use this protocol to collect SQuALITY, a dataset of question-focused summaries built on the same public-domain short stories as the multiple-choice dataset QuALITY (Pang et al., 2021). Experiments with state-of-the-art summarization systems show that our dataset is challenging and that existing automatic evaluation metrics are weak indicators of quality. △ Less

Submitted 23 May, 2022; originally announced May 2022.

Token Dropping for Efficient BERT Pretraining

Authors: Le Hou, Richard Yuanzhe Pang, Tianyi Zhou, Yuexin Wu, Xinying Song, Xiaodan Song, Denny Zhou

Abstract: Transformer-based models generally allocate the same amount of computation for each token in a given sequence. We develop a simple but effective "token dropping" method to accelerate the pretraining of transformer models, such as BERT, without degrading its performance on downstream tasks. In short, we drop unimportant tokens starting from an intermediate layer in the model to make the model focus… ▽ More Transformer-based models generally allocate the same amount of computation for each token in a given sequence. We develop a simple but effective "token dropping" method to accelerate the pretraining of transformer models, such as BERT, without degrading its performance on downstream tasks. In short, we drop unimportant tokens starting from an intermediate layer in the model to make the model focus on important tokens; the dropped tokens are later picked up by the last layer of the model so that the model still produces full-length sequences. We leverage the already built-in masked language modeling (MLM) loss to identify unimportant tokens with practically no computational overhead. In our experiments, this simple approach reduces the pretraining cost of BERT by 25% while achieving similar overall fine-tuning performance on standard downstream tasks. △ Less

Submitted 24 March, 2022; originally announced March 2022.

Comments: ACL 2022

Amortized Noisy Channel Neural Machine Translation

Authors: Richard Yuanzhe Pang, He He, Kyunghyun Cho

Abstract: Noisy channel models have been especially effective in neural machine translation (NMT). However, recent approaches like "beam search and rerank" (BSR) incur significant computation overhead during inference, making real-world application infeasible. We aim to study if it is possible to build an amortized noisy channel NMT model such that when we do greedy decoding during inference, the translatio… ▽ More Noisy channel models have been especially effective in neural machine translation (NMT). However, recent approaches like "beam search and rerank" (BSR) incur significant computation overhead during inference, making real-world application infeasible. We aim to study if it is possible to build an amortized noisy channel NMT model such that when we do greedy decoding during inference, the translation accuracy matches that of BSR in terms of reward (based on the source-to-target log probability and the target-to-source log probability) and quality (based on BLEU and BLEURT). We attempt three approaches to train the new model: knowledge distillation, one-step-deviation imitation learning, and Q learning. The first approach obtains the noisy channel signal from a pseudo-corpus, and the latter two approaches aim to optimize toward a noisy-channel MT reward directly. For all three approaches, the generated translations fail to achieve rewards comparable to BSR, but the translation quality approximated by BLEU and BLEURT is similar to the quality of BSR-produced translations. Additionally, all three approaches speed up inference by 1-2 orders of magnitude. △ Less

Submitted 18 July, 2022; v1 submitted 16 December, 2021; originally announced December 2021.

Comments: INLG 2022

QuALITY: Question Answering with Long Input Texts, Yes!

Authors: Richard Yuanzhe Pang, Alicia Parrish, Nitish Joshi, Nikita Nangia, Jason Phang, Angelica Chen, Vishakh Padmakumar, Johnny Ma, Jana Thompson, He He, Samuel R. Bowman

Abstract: To enable building and testing models on long-document comprehension, we introduce QuALITY, a multiple-choice QA dataset with context passages in English that have an average length of about 5,000 tokens, much longer than typical current models can process. Unlike in prior work with passages, our questions are written and validated by contributors who have read the entire passage, rather than rely… ▽ More To enable building and testing models on long-document comprehension, we introduce QuALITY, a multiple-choice QA dataset with context passages in English that have an average length of about 5,000 tokens, much longer than typical current models can process. Unlike in prior work with passages, our questions are written and validated by contributors who have read the entire passage, rather than relying on summaries or excerpts. In addition, only half of the questions are answerable by annotators working under tight time constraints, indicating that skimming and simple search are not enough to consistently perform well. Our baseline models perform poorly on this task (55.4%) and significantly lag behind human performance (93.5%). △ Less

Submitted 11 May, 2022; v1 submitted 15 December, 2021; originally announced December 2021.

Comments: NAACL 2022

AgreeSum: Agreement-Oriented Multi-Document Summarization

Authors: Richard Yuanzhe Pang, Adam D. Lelkes, Vinh Q. Tran, Cong Yu

Abstract: We aim to renew interest in a particular multi-document summarization (MDS) task which we call AgreeSum: agreement-oriented multi-document summarization. Given a cluster of articles, the goal is to provide abstractive summaries that represent information common and faithful to all input articles. Given the lack of existing datasets, we create a dataset for AgreeSum, and provide annotations on arti… ▽ More We aim to renew interest in a particular multi-document summarization (MDS) task which we call AgreeSum: agreement-oriented multi-document summarization. Given a cluster of articles, the goal is to provide abstractive summaries that represent information common and faithful to all input articles. Given the lack of existing datasets, we create a dataset for AgreeSum, and provide annotations on article-summary entailment relations for a subset of the clusters in the dataset. We aim to create strong baselines for the task by applying the top-performing pretrained single-document summarization model PEGASUS onto AgreeSum, leveraging both annotated clusters by supervised losses, and unannotated clusters by T5-based entailment-related and language-related losses. Compared to other baselines, both automatic evaluation and human evaluation show better article-summary and cluster-summary entailment in generated summaries. On a separate note, we hope that our article-summary entailment annotations contribute to the community's effort in improving abstractive summarization faithfulness. △ Less

Submitted 4 June, 2021; originally announced June 2021.

Comments: Findings of ACL 2021

Comparing Test Sets with Item Response Theory

Authors: Clara Vania, Phu Mon Htut, William Huang, Dhara Mungra, Richard Yuanzhe Pang, Jason Phang, Haokun Liu, Kyunghyun Cho, Samuel R. Bowman

Abstract: Recent years have seen numerous NLP datasets introduced to evaluate the performance of fine-tuned models on natural language understanding tasks. Recent results from large pretrained models, though, show that many of these datasets are largely saturated and unlikely to be able to detect further progress. What kind of datasets are still effective at discriminating among strong models, and what kind… ▽ More Recent years have seen numerous NLP datasets introduced to evaluate the performance of fine-tuned models on natural language understanding tasks. Recent results from large pretrained models, though, show that many of these datasets are largely saturated and unlikely to be able to detect further progress. What kind of datasets are still effective at discriminating among strong models, and what kind of datasets should we expect to be able to detect future improvements? To measure this uniformly across datasets, we draw on Item Response Theory and evaluate 29 datasets using predictions from 18 pretrained Transformer models on individual test examples. We find that Quoref, HellaSwag, and MC-TACO are best suited for distinguishing among state-of-the-art models, while SNLI, MNLI, and CommitmentBank seem to be saturated for current strong models. We also observe span selection task format, which is used for QA datasets like QAMR or SQuAD2.0, is effective in differentiating between strong and weak models. △ Less

Submitted 1 June, 2021; originally announced June 2021.

Comments: ACL 2021

Text Generation by Learning from Demonstrations

Authors: Richard Yuanzhe Pang, He He

Abstract: Current approaches to text generation largely rely on autoregressive models and maximum likelihood estimation. This paradigm leads to (i) diverse but low-quality samples due to mismatched learning objective and evaluation metric (likelihood vs. quality) and (ii) exposure bias due to mismatched history distributions (gold vs. model-generated). To alleviate these problems, we frame text generation a… ▽ More Current approaches to text generation largely rely on autoregressive models and maximum likelihood estimation. This paradigm leads to (i) diverse but low-quality samples due to mismatched learning objective and evaluation metric (likelihood vs. quality) and (ii) exposure bias due to mismatched history distributions (gold vs. model-generated). To alleviate these problems, we frame text generation as an offline reinforcement learning (RL) problem with expert demonstrations (i.e., the reference), where the goal is to maximize quality given model-generated histories. We propose GOLD (generation by off-policy learning from demonstrations): an easy-to-optimize algorithm that learns from the demonstrations by importance weighting. Intuitively, GOLD upweights confident tokens and downweights unconfident ones in the reference during training, avoiding optimization issues faced by prior RL approaches that rely on online data collection. According to both automatic and human evaluation, models trained by GOLD outperform those trained by MLE and policy gradient on summarization, question generation, and machine translation. Further, our models are less sensitive to decoding algorithms and alleviate exposure bias. △ Less

Submitted 2 March, 2021; v1 submitted 16 September, 2020; originally announced September 2020.

Comments: ICLR 2021

ENGINE: Energy-Based Inference Networks for Non-Autoregressive Machine Translation

Authors: Lifu Tu, Richard Yuanzhe Pang, Sam Wiseman, Kevin Gimpel

Abstract: We propose to train a non-autoregressive machine translation model to minimize the energy defined by a pretrained autoregressive model. In particular, we view our non-autoregressive translation system as an inference network (Tu and Gimpel, 2018) trained to minimize the autoregressive teacher energy. This contrasts with the popular approach of training a non-autoregressive model on a distilled cor… ▽ More We propose to train a non-autoregressive machine translation model to minimize the energy defined by a pretrained autoregressive model. In particular, we view our non-autoregressive translation system as an inference network (Tu and Gimpel, 2018) trained to minimize the autoregressive teacher energy. This contrasts with the popular approach of training a non-autoregressive model on a distilled corpus consisting of the beam-searched outputs of such a teacher model. Our approach, which we call ENGINE (ENerGy-based Inference NEtworks), achieves state-of-the-art non-autoregressive results on the IWSLT 2014 DE-EN and WMT 2016 RO-EN datasets, approaching the performance of autoregressive models. △ Less

Submitted 12 May, 2020; v1 submitted 2 May, 2020; originally announced May 2020.

Comments: ACL 2020 camera-ready version

Intermediate-Task Transfer Learning with Pretrained Models for Natural Language Understanding: When and Why Does It Work?

Authors: Yada Pruksachatkun, Jason Phang, Haokun Liu, Phu Mon Htut, Xiaoyi Zhang, Richard Yuanzhe Pang, Clara Vania, Katharina Kann, Samuel R. Bowman

Abstract: While pretrained models such as BERT have shown large gains across natural language understanding tasks, their performance can be improved by further training the model on a data-rich intermediate task, before fine-tuning it on a target task. However, it is still poorly understood when and why intermediate-task training is beneficial for a given target task. To investigate this, we perform a large… ▽ More While pretrained models such as BERT have shown large gains across natural language understanding tasks, their performance can be improved by further training the model on a data-rich intermediate task, before fine-tuning it on a target task. However, it is still poorly understood when and why intermediate-task training is beneficial for a given target task. To investigate this, we perform a large-scale study on the pretrained RoBERTa model with 110 intermediate-target task combinations. We further evaluate all trained models with 25 probing tasks meant to reveal the specific skills that drive transfer. We observe that intermediate tasks requiring high-level inference and reasoning abilities tend to work best. We also observe that target task performance is strongly correlated with higher-level abilities such as coreference resolution. However, we fail to observe more granular correlations between probing and target task performance, highlighting the need for further work on broad-coverage probing benchmarks. We also observe evidence that the forgetting of knowledge learned during pretraining may limit our analysis, highlighting the need for further work on transfer learning methods in these settings. △ Less

Submitted 9 May, 2020; v1 submitted 1 May, 2020; originally announced May 2020.

Comments: ACL 2020

Consistency of a Recurrent Language Model With Respect to Incomplete Decoding

Authors: Sean Welleck, Ilia Kulikov, Jaedeok Kim, Richard Yuanzhe Pang, Kyunghyun Cho

Abstract: Despite strong performance on a variety of tasks, neural sequence models trained with maximum likelihood have been shown to exhibit issues such as length bias and degenerate repetition. We study the related issue of receiving infinite-length sequences from a recurrent language model when using common decoding algorithms. To analyze this issue, we first define inconsistency of a decoding algorithm,… ▽ More Despite strong performance on a variety of tasks, neural sequence models trained with maximum likelihood have been shown to exhibit issues such as length bias and degenerate repetition. We study the related issue of receiving infinite-length sequences from a recurrent language model when using common decoding algorithms. To analyze this issue, we first define inconsistency of a decoding algorithm, meaning that the algorithm can yield an infinite-length sequence that has zero probability under the model. We prove that commonly used incomplete decoding algorithms - greedy search, beam search, top-k sampling, and nucleus sampling - are inconsistent, despite the fact that recurrent language models are trained to produce sequences of finite length. Based on these insights, we propose two remedies which address inconsistency: consistent variants of top-k and nucleus sampling, and a self-terminating recurrent language model. Empirical results show that inconsistency occurs in practice, and that the proposed methods prevent inconsistency. △ Less

Submitted 2 October, 2020; v1 submitted 6 February, 2020; originally announced February 2020.

Comments: EMNLP 2020

Improving Joint Training of Inference Networks and Structured Prediction Energy Networks

Authors: Lifu Tu, Richard Yuanzhe Pang, Kevin Gimpel

Abstract: Deep energy-based models are powerful, but pose challenges for learning and inference (Belanger and McCallum, 2016). Tu and Gimpel (2018) developed an efficient framework for energy-based models by training "inference networks" to approximate structured inference instead of using gradient descent. However, their alternating optimization approach suffers from instabilities during training, requirin… ▽ More Deep energy-based models are powerful, but pose challenges for learning and inference (Belanger and McCallum, 2016). Tu and Gimpel (2018) developed an efficient framework for energy-based models by training "inference networks" to approximate structured inference instead of using gradient descent. However, their alternating optimization approach suffers from instabilities during training, requiring additional loss terms and careful hyperparameter tuning. In this paper, we contribute several strategies to stabilize and improve this joint training of energy functions and inference networks for structured prediction. We design a compound objective to jointly train both cost-augmented and test-time inference networks along with the energy function. We propose joint parameterizations for the inference networks that encourage them to capture complementary functionality during learning. We empirically validate our strategies on two sequence labeling tasks, showing easier paths to strong performance than prior work, as well as further improvements with global energy terms. △ Less

Submitted 10 October, 2020; v1 submitted 7 November, 2019; originally announced November 2019.

Comments: EMNLP 2020 Workshop on Structured Prediction for NLP (SPNLP)

The Daunting Task of Real-World Textual Style Transfer Auto-Evaluation

Authors: Richard Yuanzhe Pang

Abstract: The difficulty of textual style transfer lies in the lack of parallel corpora. Numerous advances have been proposed for the unsupervised generation. However, significant problems remain with the auto-evaluation of style transfer tasks. Based on the summary of Pang and Gimpel (2018) and Mir et al. (2019), style transfer evaluations rely on three criteria: style accuracy of transferred sentences, co… ▽ More The difficulty of textual style transfer lies in the lack of parallel corpora. Numerous advances have been proposed for the unsupervised generation. However, significant problems remain with the auto-evaluation of style transfer tasks. Based on the summary of Pang and Gimpel (2018) and Mir et al. (2019), style transfer evaluations rely on three criteria: style accuracy of transferred sentences, content similarity between original and transferred sentences, and fluency of transferred sentences. We elucidate the problematic current state of style transfer research. Given that current tasks do not represent real use cases of style transfer, current auto-evaluation approach is flawed. This discussion aims to bring researchers to think about the future of style transfer and style transfer evaluation research. △ Less

Submitted 9 October, 2019; v1 submitted 8 October, 2019; originally announced October 2019.

Comments: Extended abstract in EMNLP Workshop on Neural Generation and Translation (WNGT 2019)

Unsupervised Evaluation Metrics and Learning Criteria for Non-Parallel Textual Transfer

Authors: Richard Yuanzhe Pang, Kevin Gimpel

Abstract: We consider the problem of automatically generating textual paraphrases with modified attributes or properties, focusing on the setting without parallel data (Hu et al., 2017; Shen et al., 2017). This setting poses challenges for evaluation. We show that the metric of post-transfer classification accuracy is insufficient on its own, and propose additional metrics based on semantic preservation and… ▽ More We consider the problem of automatically generating textual paraphrases with modified attributes or properties, focusing on the setting without parallel data (Hu et al., 2017; Shen et al., 2017). This setting poses challenges for evaluation. We show that the metric of post-transfer classification accuracy is insufficient on its own, and propose additional metrics based on semantic preservation and fluency as well as a way to combine them into a single overall score. We contribute new loss functions and training strategies to address the different metrics. Semantic preservation is addressed by adding a cyclic consistency loss and a loss based on paraphrase pairs, while fluency is improved by integrating losses based on style-specific language models. We experiment with a Yelp sentiment dataset and a new literature dataset that we propose, using multiple models that extend prior work (Shen et al., 2017). We demonstrate that our metrics correlate well with human judgments, at both the sentence-level and system-level. Automatic and manual evaluation also show large improvements over the baseline method of Shen et al. (2017). We hope that our proposed metrics can speed up system development for new textual transfer tasks while also encouraging the community to address our three complementary aspects of transfer quality. △ Less

Submitted 30 September, 2019; v1 submitted 28 October, 2018; originally announced October 2018.

Comments: EMNLP 2019 Workshop on Neural Generation and Translation (WNGT)