William Held

Caleb Ziems, William Held, Omar Shaikh et al. · gatech

Large Language Models (LLMs) are capable of successfully performing many language processing tasks zero-shot (without training data). If zero-shot LLMs can also reliably classify and explain social phenomena like persuasiveness and political ideology, then LLMs could augment the Computational Social Science (CSS) pipeline in important ways. This work provides a road map for using LLMs as CSS tools. Towards this end, we contribute a set of prompting best practices and an extensive evaluation pipeline to measure the zero-shot performance of 13 language models on 25 representative English CSS benchmarks. On taxonomic labeling tasks (classification), LLMs fail to outperform the best fine-tuned models but still achieve fair levels of agreement with humans. On free-form coding tasks (generation), LLMs produce explanations that often exceed the quality of crowdworkers' gold references. We conclude that the performance of today's LLMs can augment the CSS research pipeline in two ways: (1) serving as zero-shot data annotators on human annotation teams, and (2) bootstrapping challenging creative generation tasks (e.g., explaining the underlying attributes of a text). In summary, LLMs are posed to meaningfully participate in social science analysis in partnership with humans.

Omar Shaikh, Hongxin Zhang, William Held et al. · gatech

Generating a Chain of Thought (CoT) has been shown to consistently improve large language model (LLM) performance on a wide range of NLP tasks. However, prior work has mainly focused on logical reasoning tasks (e.g. arithmetic, commonsense QA); it remains unclear whether improvements hold for more diverse types of reasoning, especially in socially situated contexts. Concretely, we perform a controlled evaluation of zero-shot CoT across two socially sensitive domains: harmful questions and stereotype benchmarks. We find that zero-shot CoT reasoning in sensitive domains significantly increases a model's likelihood to produce harmful or undesirable output, with trends holding across different prompt formats and model variants. Furthermore, we show that harmful CoTs increase with model size, but decrease with improved instruction following. Our work suggests that zero-shot CoT should be used with caution on socially important tasks, especially when marginalized groups or sensitive topics are involved.

Caleb Ziems, William Held, Jingfeng Yang et al. · amazon-science, gatech

Dialect differences caused by regional, social, and economic factors cause performance discrepancies for many groups of language technology users. Inclusive and equitable language technology must critically be dialect invariant, meaning that performance remains constant over dialectal shifts. Current systems often fall short of this ideal since they are designed and tested on a single dialect: Standard American English (SAE). We introduce a suite of resources for evaluating and achieving English dialect invariance. The resource is called Multi-VALUE, a controllable rule-based translation system spanning 50 English dialects and 189 unique linguistic features. Multi-VALUE maps SAE to synthetic forms of each dialect. First, we use this system to stress tests question answering, machine translation, and semantic parsing. Stress tests reveal significant performance disparities for leading models on non-standard dialects. Second, we use this system as a data augmentation technique to improve the dialect robustness of existing systems. Finally, we partner with native speakers of Chicano and Indian English to release new gold-standard variants of the popular CoQA task. To execute the transformation code, run model checkpoints, and download both synthetic and gold-standard dialectal benchmark datasets, see http://value-nlp.org.

Omar Shaikh, Caleb Ziems, William Held et al. · gatech

Pragmatic reference enables efficient interpersonal communication. Prior work uses simple reference games to test models of pragmatic reasoning, often with unidentified speakers and listeners. In practice, however, speakers' sociocultural background shapes their pragmatic assumptions. For example, readers of this paper assume NLP refers to "Natural Language Processing," and not "Neuro-linguistic Programming." This work introduces the Cultural Codes dataset, which operationalizes sociocultural pragmatic inference in a simple word reference game. Cultural Codes is based on the multi-turn collaborative two-player game, Codenames Duet. Our dataset consists of 794 games with 7,703 turns, distributed across 153 unique players. Alongside gameplay, we collect information about players' personalities, values, and demographics. Utilizing theories of communication and pragmatics, we predict each player's actions via joint modeling of their sociocultural priors and the game context. Our experiments show that accounting for background characteristics significantly improves model performance for tasks related to both clue giving and guessing, indicating that sociocultural priors play a vital role in gameplay decisions.

William Held, Diyi Yang · gatech

Multilingual transformer-based models demonstrate remarkable zero and few-shot transfer across languages by learning and reusing language-agnostic features. However, as a fixed-size model acquires more languages, its performance across all languages degrades, a phenomenon termed interference. Often attributed to limited model capacity, interference is commonly addressed by adding additional parameters despite evidence that transformer-based models are overparameterized. In this work, we show that it is possible to reduce interference by instead identifying and pruning language-specific parameters. First, we use Shapley Values, a credit allocation metric from coalitional game theory, to identify attention heads that introduce interference. Then, we show that removing identified attention heads from a fixed model improves performance for a target language on both sentence classification and structural prediction, seeing gains as large as 24.7\%. Finally, we provide insights on language-agnostic and language-specific attention heads using attention visualization.

William Held, Christopher Hidey, Fei Liu et al. · gatech

Modern virtual assistants use internal semantic parsing engines to convert user utterances to actionable commands. However, prior work has demonstrated that semantic parsing is a difficult multilingual transfer task with low transfer efficiency compared to other tasks. In global markets such as India and Latin America, this is a critical issue as switching between languages is prevalent for bilingual users. In this work we dramatically improve the zero-shot performance of a multilingual and codeswitched semantic parsing system using two stages of multilingual alignment. First, we show that constrastive alignment pretraining improves both English performance and transfer efficiency. We then introduce a constrained optimization approach for hyperparameter-free adversarial alignment during finetuning. Our Doubly Aligned Multilingual Parser (DAMP) improves mBERT transfer performance by 3x, 6x, and 81x on the Spanglish, Hinglish and Multilingual Task Oriented Parsing benchmarks respectively and outperforms XLM-R and mT5-Large using 3.2x fewer parameters.

William Held, Camille Harris, Michael Best et al. · gatech

Coloniality, the continuation of colonial harms beyond "official" colonization, has pervasive effects across society and scientific fields. Natural Language Processing (NLP) is no exception to this broad phenomenon. In this work, we argue that coloniality is implicitly embedded in and amplified by NLP data, algorithms, and software. We formalize this analysis using Actor-Network Theory (ANT): an approach to understanding social phenomena through the network of relationships between human stakeholders and technology. We use our Actor-Network to guide a quantitative survey of the geography of different phases of NLP research, providing evidence that inequality along colonial boundaries increases as NLP builds on itself. Based on this, we argue that combating coloniality in NLP requires not only changing current values but also active work to remove the accumulation of colonial ideals in our foundational data and algorithms.

Zedian Xiao, William Held, Yanchen Liu et al. · gatech, harvard

Large Language Models (LLMs) are trained on corpora disproportionally weighted in favor of Standard American English. As a result, speakers of other dialects experience significantly more failures when interacting with these technologies. In practice, these speakers often accommodate their speech to be better understood. Our work shares the belief that language technologies should be designed to accommodate the diversity in English dialects and not the other way around. However, prior works on dialect struggle with generalizing to evolving and emerging dialects in a scalable manner. To fill this gap, our method, HyperLoRA, leverages expert linguistic knowledge to enable resource-efficient adaptation via hypernetworks. By disentangling dialect-specific and cross-dialectal information, HyperLoRA improves generalization to unseen dialects in a task-agnostic fashion. Not only is HyperLoRA more scalable in the number of parameters, but it also achieves the best or most competitive performance across 5 dialects in a zero-shot setting. In this way, our approach facilitates access to language technology for billions of English dialect speakers who are traditionally underrepresented.

Eric Czech, Zhiwei Xu, Yael Elmatad et al.

Chinchilla Approach 2 is among the most widely used methods for fitting neural scaling laws. Its parabolic approximation introduces systematic biases in compute-optimal allocation estimates, even on noise-free synthetic data. Applied to published Llama 3 IsoFLOP data at open frontier compute scales, these biases imply a parameter underallocation corresponding to 6.5% of the $3.8\times10^{25}$ FLOP training budget and \$1.4M (90% CI: \$412K-\$2.9M) in unnecessary compute at 50% H100 MFU. Simulated multimodal model misallocations show even greater opportunity costs due to higher loss surface asymmetry. Three sources of this error are examined: IsoFLOP sampling grid width (Taylor approximation accuracy), uncentered IsoFLOP sampling, and loss surface asymmetry ($α\neq β$). Chinchilla Approach 3 largely eliminates these biases but is often regarded as less data-efficient, numerically unstable, prone to local minima, and harder to implement. Each concern is shown to be unfounded or addressable, especially when the partially linear structure of the objective is exploited via Variable Projection, enabling unbiased inference on all five loss surface parameters through a two-dimensional optimization that is well-conditioned, analytically differentiable, and amenable to dense, or even exhaustive, grid search. It may serve as a more convenient replacement for Approach 2 or a more scalable alternative for adaptations of Approach 3 to richer scaling law formulations. See https://github.com/Open-Athena/vpnls for details and https://openathena.ai/scaling-law-analysis for other results from this study.

Potsawee Manakul, Woody Haosheng Gan, Martijn Bartelds et al. · gatech

Current audio language models are predominantly text-first, either extending pre-trained text LLM backbones or relying on semantic-only audio tokens, limiting general audio modeling. This paper presents a systematic empirical study of native audio foundation models that apply next-token prediction to audio at scale, jointly modeling semantic content, acoustic details, and text to support both general audio generation and cross-modal capabilities. We provide comprehensive empirical insights for building such models: (1) We systematically investigate design choices -- data sources, text mixture ratios, and token composition -- establishing a validated training recipe. (2) We conduct the first scaling law study for discrete audio models via IsoFLOP analysis on 64 models spanning $3{\times}10^{18}$ to $3{\times}10^{20}$ FLOPs, finding that optimal data grows 1.6$\times$ faster than optimal model size. (3) We apply these lessons to train SODA (Scaling Open Discrete Audio), a suite of models from 135M to 4B parameters on 500B tokens, comparing against our scaling predictions and existing models. SODA serves as a flexible backbone for diverse audio/text tasks -- we demonstrate this by fine-tuning for voice-preserving speech-to-speech translation, using the same unified architecture.

Michael J. Ryan, William Held, Diyi Yang · gatech, stanford

Before being deployed for user-facing applications, developers align Large Language Models (LLMs) to user preferences through a variety of procedures, such as Reinforcement Learning From Human Feedback (RLHF) and Direct Preference Optimization (DPO). Current evaluations of these procedures focus on benchmarks of instruction following, reasoning, and truthfulness. However, human preferences are not universal, and aligning to specific preference sets may have unintended effects. We explore how alignment impacts performance along three axes of global representation: English dialects, multilingualism, and opinions from and about countries worldwide. Our results show that current alignment procedures create disparities between English dialects and global opinions. We find alignment improves capabilities in several languages. We conclude by discussing design decisions that led to these unintended impacts and recommendations for more equitable preference tuning. We make our code and data publicly available on Github.

Caleb Ziems, William Held, Jane Yu et al.

To serve global users safely and productively, LLMs need culture-specific knowledge that might not be learned during pre-training. How do we find such knowledge that is (1) salient to in-group users, but (2) unknown to LLMs? The most common solutions are single-initiative: either researchers define challenging questions that users passively answer (traditional annotation), or users actively produce data that researchers structure as benchmarks (knowledge extraction). The process would benefit from mixed-initiative collaboration, where users guide the process to meaningfully reflect their cultures, and LLMs steer the process towards more challenging questions that meet the researcher's goals. We propose a mixed-initiative methodology called CultureCartography. Here, an LLM initializes annotation with questions for which it has low-confidence answers, making explicit both its prior knowledge and the gaps therein. This allows a human respondent to fill these gaps and steer the model towards salient topics through direct edits. We implement this methodology as a tool called CultureExplorer. Compared to a baseline where humans answer LLM-proposed questions, we find that CultureExplorer more effectively produces knowledge that leading models like DeepSeek R1 and GPT-4o are missing, even with web search. Fine-tuning on this data boosts the accuracy of Llama-3.1-8B by up to 19.2% on related culture benchmarks.

Woody Haosheng Gan, William Held, Diyi Yang

The rapid proliferation of large audio models (LAMs) demands efficient approaches for model comparison, yet comprehensive benchmarks are costly. To fill this gap, we investigate whether minimal subsets can reliably evaluate LAMs while reducing costs and data redundancy. Analyzing 10 subset selection methods with 18 audio models across 40 tasks covering major LAM evaluation dimensions, we show that subsets of just 50 examples (0.3% of data) can achieve over 0.93 Pearson correlation with full benchmark scores. To understand how well these scores align with what practitioners ultimately care about, user satisfaction, we collect 776 human preference ratings from realistic voice assistant conversations, finding that both subsets and full benchmark achieve only 0.85 correlation with human. To better predict preferences, we trained regression models on these selected subsets, achieving 0.98 correlation -- outperforming regression models trained on both random subsets and the full benchmark. This demonstrates that in regression modeling, well-curated subsets outpredict the full benchmark, showing quality over quantity. We open-source these regression-weighted subsets as the HUMANS benchmark, an efficient proxy for LAM evaluation that captures both benchmark performance and user preferences.

Long Phan, Alice Gatti, Ziwen Han et al. · amazon-science, apple-ml

Benchmarks are important tools for tracking the rapid advancements in large language model (LLM) capabilities. However, benchmarks are not keeping pace in difficulty: LLMs now achieve over 90\% accuracy on popular benchmarks like MMLU, limiting informed measurement of state-of-the-art LLM capabilities. In response, we introduce Humanity's Last Exam (HLE), a multi-modal benchmark at the frontier of human knowledge, designed to be the final closed-ended academic benchmark of its kind with broad subject coverage. HLE consists of 2,500 questions across dozens of subjects, including mathematics, humanities, and the natural sciences. HLE is developed globally by subject-matter experts and consists of multiple-choice and short-answer questions suitable for automated grading. Each question has a known solution that is unambiguous and easily verifiable, but cannot be quickly answered via internet retrieval. State-of-the-art LLMs demonstrate low accuracy and calibration on HLE, highlighting a significant gap between current LLM capabilities and the expert human frontier on closed-ended academic questions. To inform research and policymaking upon a clear understanding of model capabilities, we publicly release HLE at https://lastexam.ai.

Diyi Yang, Caleb Ziems, William Held et al. · gatech

People rely on social skills like conflict resolution to communicate effectively and to thrive in both work and personal life. However, practice environments for social skills are typically out of reach for most people. How can we make social skill training more available, accessible, and inviting? Drawing upon interdisciplinary research from communication and psychology, this perspective paper identifies social skill barriers to enter specialized fields. Then we present a solution that leverages large language models for social skill training via a generic framework. Our AI Partner, AI Mentor framework merges experiential learning with realistic practice and tailored feedback. This work ultimately calls for cross-disciplinary innovation to address the broader implications for workforce development and social equality.

William Held, Bhargavi Paranjape, Punit Singh Koura et al. · gatech

Large Language Models improve with increasing amounts of high-quality training data. However, leveraging larger datasets requires balancing quality, quantity, and diversity across sources. After evaluating nine baseline methods under both compute- and data-constrained scenarios, we find token-count heuristics outperform manual and learned mixes, indicating that simple approaches accounting for dataset size and diversity are surprisingly effective. Building on this insight, we propose two complementary approaches: UtiliMax, which extends token-based heuristics by incorporating utility estimates from reduced-scale ablations, achieving up to a 10.6x speedup over manual baselines; and Model Estimated Data Utility (MEDU), which leverages LLMs to estimate data utility from small samples, matching ablation-based performance while reducing computational requirements by $\sim$200x. Together, these approaches establish a new framework for automated, compute-efficient data mixing that is robust across training regimes.

Potsawee Manakul, Woody Haosheng Gan, Michael J. Ryan et al. · gatech

Current speech evaluation suffers from two critical limitations: the need and difficulty of designing specialized systems targeting individual audio characteristics, and poor correlation between automatic evaluation methods and human preferences. This work presents a systematic study of Large Audio Model (LAM) as a Judge, AudioJudge, investigating whether it can provide a unified evaluation framework that addresses both challenges. We systematically explore AudioJudge across audio characteristic detection tasks, including pronunciation, speaking rate, speaker identification and speech quality, and system-level human preference simulation for automated benchmarking. We investigate different prompt engineering strategies, finding that audio concatenation combined with in-context learning significantly improves performance across both audio characteristic detection and human preference simulation tasks. We further introduce a multi-aspect ensemble AudioJudge to enable general-purpose multi-aspect audio evaluation. This method decomposes speech assessment into specialized judges for lexical content, speech quality, and paralinguistic features, achieving up to 0.91 Spearman correlation with human preferences on our system ranking benchmark. Robustness analysis reveals that while LAMs maintain strong performance under acoustic noise, they exhibit significant verbosity and positional biases that require careful mitigation.

Michael J Ryan, Omar Shaikh, Aditri Bhagirath et al. · gatech

Recent calls for pluralistic alignment of Large Language Models (LLMs) encourage adapting models to diverse user preferences. However, most prior work on personalized reward models heavily rely on additional identity information, such as demographic details or a predefined set of preference categories. To this end, we introduce SynthesizeMe, an approach to inducing synthetic user personas from user interactions for personalized reward modeling. SynthesizeMe first generates and verifies reasoning to explain user preferences, then induces synthetic user personas from that reasoning, and finally filters to informative prior user interactions in order to build personalized prompts for a particular user. We show that using SynthesizeMe induced prompts improves personalized LLM-as-a-judge accuracy by 4.4% on Chatbot Arena. Combining SynthesizeMe derived prompts with a reward model achieves top performance on PersonalRewardBench: a new curation of user-stratified interactions with chatbots collected from 854 users of Chatbot Arena and PRISM.

William Held, David Hall, Percy Liang et al. · gatech

Scaling laws describe how language models improve with additional data, parameters, and compute. While widely used, they are typically measured on aggregate test sets. Aggregate evaluations yield clean trends but average over heterogeneous subpopulations, obscuring performance disparities. We introduce relative scaling laws, which track how performance gaps between test distributions evolve with scale rather than focusing solely on absolute error. Using 255 decoder-only Transformers trained under matched-compute (IsoFLOP) budgets from $10^{18}$--$10^{20}$ FLOPs on standard pretraining datasets, we find diverse trajectories: academic domains on MMLU converge toward parity; regional English dialects shift depending on population size; and clusters of AI risk behaviours split, with capability- and influence-related risks increasing during pretraining while adversarial risks do not. These results show that although scaling improves overall performance, it is not a universal equalizer. To support further study, we release all model checkpoints from this work to enable practitioners to measure relative alongside traditional scaling laws, in order to better prioritize robustness challenges in light of the bitter lesson.

Caleb Ziems, William Held, Jane Dwivedi-Yu et al.

Information Retrieval (IR) systems are designed to deliver relevant content, but traditional systems may not optimize rankings for fairness, neutrality, or the balance of ideas. Consequently, IR can often introduce indexical biases, or biases in the positional order of documents. Although indexical bias can demonstrably affect people's opinion, voting patterns, and other behaviors, these issues remain understudied as the field lacks reliable metrics and procedures for automatically measuring indexical bias. Towards this end, we introduce the PAIR framework, which supports automatic bias audits for ranked documents or entire IR systems. After introducing DUO, the first general-purpose automatic bias metric, we run an extensive evaluation of 8 IR systems on a new corpus of 32k synthetic and 4.7k natural documents, with 4k queries spanning 1.4k controversial issue topics. A human behavioral study validates our approach, showing that our bias metric can help predict when and how indexical bias will shift a reader's opinion.

Yanchen Liu, William Held, Diyi Yang

Existing large language models (LLMs) that mainly focus on Standard American English (SAE) often lead to significantly worse performance when being applied to other English dialects. While existing mitigations tackle discrepancies for individual target dialects, they assume access to high-accuracy dialect identification systems. The boundaries between dialects are inherently flexible, making it difficult to categorize language into discrete predefined categories. In this paper, we propose DADA (Dialect Adaptation via Dynamic Aggregation), a modular approach to imbue SAE-trained models with multi-dialectal robustness by composing adapters which handle specific linguistic features. The compositional architecture of DADA allows for both targeted adaptation to specific dialect variants and simultaneous adaptation to various dialects. We show that DADA is effective for both single task and instruction finetuned language models, offering an extensible and interpretable framework for adapting existing LLMs to different English dialects.

William Held, Dan Iter, Dan Jurafsky

Performing event and entity coreference resolution across documents vastly increases the number of candidate mentions, making it intractable to do the full $n^2$ pairwise comparisons. Existing approaches simplify by considering coreference only within document clusters, but this fails to handle inter-cluster coreference, common in many applications. As a result cross-document coreference algorithms are rarely applied to downstream tasks. We draw on an insight from discourse coherence theory: potential coreferences are constrained by the reader's discourse focus. We model the entities/events in a reader's focus as a neighborhood within a learned latent embedding space which minimizes the distance between mentions and the centroids of their gold coreference clusters. We then use these neighborhoods to sample only hard negatives to train a fine-grained classifier on mention pairs and their local discourse features. Our approach achieves state-of-the-art results for both events and entities on the ECB+, Gun Violence, Football Coreference, and Cross-Domain Cross-Document Coreference corpora. Furthermore, training on multiple corpora improves average performance across all datasets by 17.2 F1 points, leading to a robust coreference resolution model for use in downstream tasks where link distribution is unknown.