Ashton Anderson

Reid McIlroy-Young, Jon Kleinberg, Siddhartha Sen et al.

An emerging theme in artificial intelligence research is the creation of models to simulate the decisions and behavior of specific people, in domains including game-playing, text generation, and artistic expression. These models go beyond earlier approaches in the way they are tailored to individuals, and the way they are designed for interaction rather than simply the reproduction of fixed, pre-computed behaviors. We refer to these as mimetic models, and in this paper we develop a framework for characterizing the ethical and social issues raised by their growing availability. Our framework includes a number of distinct scenarios for the use of such models, and considers the impacts on a range of different participants, including the target being modeled, the operator who deploys the model, and the entities that interact with it.

Difan Jiao, Yilun Liu, Ye Yuan et al.

Guard models are widely used to detect harmful content in user prompts and LLM responses. However, state-of-the-art guard models rely solely on terminal-layer representations and overlook the rich safety-relevant features distributed across internal layers. We present SIREN, a lightweight guard model that harnesses these internal features. By identifying safety neurons via linear probing and combining them through an adaptive layer-weighted strategy, SIREN builds a harmfulness detector from LLM internals without modifying the underlying model. Our comprehensive evaluation shows that SIREN substantially outperforms state-of-the-art open-source guard models across multiple benchmarks while using 250 times fewer trainable parameters. Moreover, SIREN exhibits superior generalization to unseen benchmarks, naturally enables real-time streaming detection, and significantly improves inference efficiency compared to generative guard models. Overall, our results highlight LLM internal states as a promising foundation for practical, high-performance harmfulness detection.

Daniel Monroe, George Eilender, Philip Chalmers et al.

Chess has long served as a canonical testbed for artificial intelligence, but modeling approaches for its central tasks have diverged. Maximizing playing strength, predicting human play, and enabling interpretability are typically solved with disparate architectures, and these designs are often misaligned with the geometry of the domain. This raises the natural question of whether these objectives require separate modeling paradigms, or if there exists a single architecture that supports them simultaneously. We introduce Chessformer, a unified architecture that advances the state of the art on all three central goals in chess modeling. Chessformer is an encoder-only transformer that represents board squares as tokens, augments self-attention with a novel dynamic positional encoding called Geometric Attention Bias (GAB) that adapts to domain-specific geometry, and predicts actions with an attention-based source-destination policy head. We evaluate Chessformer on each front. First, we develop \maiathree, a family of models for human move prediction that reaches 57.1\% move-matching accuracy, significantly surpassing the previous state of the art with fewer than a quarter of the parameters. Second, we integrate Chessformer into Leela Chess Zero, a leading open-source engine, adding over 100 Elo of playing strength and resulting in tournament victories over Stockfish in major computer chess competitions. Third, we show that Chessformer's square-token design makes attention patterns and activations directly attributable to board squares, enabling granular interpretability analyses that prior architectures do not naturally support. More broadly, our results demonstrate that aligning a model's tokenization, positional encoding, and output design with the underlying structure of a domain can yield simultaneous gains in performance, human compatibility, and interpretability.

Zhenwei Tang, Qianfeng Wen, Seth Grief-Albert et al.

Language models often lack grounded reasoning capabilities in specialized domains where training data is scarce but bespoke systems excel. We introduce a general framework for distilling expert system reasoning into natural language chain-of-thought explanations, enabling compact models to acquire domain expertise and the ability to generate faithful, grounded explanations. Rather than distilling only final outputs, we capture the full reasoning process, transforming opaque expert computations into transparent, step-by-step explanations. We demonstrate this approach in chess, a canonical reasoning domain where language models continue to underperform. Our 4B parameter model, C1, advances from a near-zero baseline to 48.1% accuracy, outperforming all open-source models and most frontier proprietary systems. Notably, C1 surpasses its distillation teacher and generates solutions in two orders of magnitude fewer tokens than baselines. Unlike prior neural chess approaches that predict only best moves, C1 generates explainable solutions revealing strategic reasoning. Our pipeline combines supervised fine-tuning and reinforcement learning with theme-balanced data sampling for comprehensive tactical coverage. Master Distillation demonstrates how to inject expert-level knowledge into compact models for under-optimized domains, offering a recipe for unlocking RLVR where LLMs lack sufficient base capabilities.

Reid McIlroy-Young, Russell Wang, Siddhartha Sen et al.

The advent of machine learning models that surpass human decision-making ability in complex domains has initiated a movement towards building AI systems that interact with humans. Many building blocks are essential for this activity, with a central one being the algorithmic characterization of human behavior. While much of the existing work focuses on aggregate human behavior, an important long-range goal is to develop behavioral models that specialize to individual people and can differentiate among them. To formalize this process, we study the problem of behavioral stylometry, in which the task is to identify a decision-maker from their decisions alone. We present a transformer-based approach to behavioral stylometry in the context of chess, where one attempts to identify the player who played a set of games. Our method operates in a few-shot classification framework, and can correctly identify a player from among thousands of candidate players with 98% accuracy given only 100 labeled games. Even when trained on amateur play, our method generalises to out-of-distribution samples of Grandmaster players, despite the dramatic differences between amateur and world-class players. Finally, we consider more broadly what our resulting embeddings reveal about human style in chess, as well as the potential ethical implications of powerful methods for identifying individuals from behavioral data.

Zhenwei Tang, Difan Jiao, Reid McIlroy-Young et al.

There are an increasing number of domains in which artificial intelligence (AI) systems both surpass human ability and accurately model human behavior. This introduces the possibility of algorithmically-informed teaching in these domains through more relatable AI partners and deeper insights into human decision-making. Critical to achieving this goal, however, is coherently modeling human behavior at various skill levels. Chess is an ideal model system for conducting research into this kind of human-AI alignment, with its rich history as a pivotal testbed for AI research, mature superhuman AI systems like AlphaZero, and precise measurements of skill via chess rating systems. Previous work in modeling human decision-making in chess uses completely independent models to capture human style at different skill levels, meaning they lack coherence in their ability to adapt to the full spectrum of human improvement and are ultimately limited in their effectiveness as AI partners and teaching tools. In this work, we propose a unified modeling approach for human-AI alignment in chess that coherently captures human style across different skill levels and directly captures how people improve. Recognizing the complex, non-linear nature of human learning, we introduce a skill-aware attention mechanism to dynamically integrate players' strengths with encoded chess positions, enabling our model to be sensitive to evolving player skill. Our experimental results demonstrate that this unified framework significantly enhances the alignment between AI and human players across a diverse range of expertise levels, paving the way for deeper insights into human decision-making and AI-guided teaching tools.

Patrick Yung Kang Lee, Jessica Y. Bo, Zixin Zhao et al.

Individuals are turning to increasingly anthropomorphic, general-purpose chatbots for AI companionship, rather than roleplay-specific platforms. However, not much is known about how individuals perceive and conduct their relationships with general-purpose chatbots. We analyzed semi-structured interviews (n=13), survey responses (n=43), and community discussions on Reddit (41k+ posts and comments) to triangulate the internal dynamics, external influences, and steering strategies that shape AI companion relationships. We learned that individuals conceptualize their companions based on an interplay of their beliefs about the companion's own agency and the autonomy permitted by the platform, how they pursue interactions with the companion, and the perceived initiatives that the companion takes. In combination with the external factors that affect relationship dynamics, particularly model updates that can derail companion behaviour and stability, individuals make use of different types of steering strategies to preserve their relationship, for example, by setting behavioural instructions or porting to other AI platforms. We discuss implications for accountability and transparency in AI systems, where emotional connection competes with broader product objectives and safety constraints.

Harsh Kumar, Suhyeon Yoo, Angela Zavaleta Bernuy et al.

Although engagement in self-directed wellness exercises typically declines over time, integrating social support such as coaching can sustain it. However, traditional forms of support are often inaccessible due to the high costs and complex coordination. Large Language Models (LLMs) show promise in providing human-like dialogues that could emulate social support. Yet, in-depth, in situ investigations of LLMs to support behavior change remain underexplored. We conducted two randomized experiments to assess the impact of LLM agents on user engagement with mindfulness exercises. First, a single-session study, involved 502 crowdworkers; second, a three-week study, included 54 participants. We explored two types of LLM agents: one providing information and another facilitating self-reflection. Both agents enhanced users' intentions to practice mindfulness. However, only the information-providing LLM, featuring a friendly persona, significantly improved engagement with the exercises. Our findings suggest that specific LLM agents may bridge the social support gap in digital health interventions.

Difan Jiao, Yilun Liu, Zhenwei Tang et al.

Among the many tasks that Large Language Models (LLMs) have revolutionized is text classification. Current text classification paradigms, however, rely solely on the output of the final layer in the LLM, with the rich information contained in internal neurons largely untapped. In this study, we present SPIN: a model-agnostic framework that sparsifies and integrates internal neurons of intermediate layers of LLMs for text classification. Specifically, SPIN sparsifies internal neurons by linear probing-based salient neuron selection layer by layer, avoiding noise from unrelated neurons and ensuring efficiency. The cross-layer salient neurons are then integrated to serve as multi-layered features for the classification head. Extensive experimental results show our proposed SPIN significantly improves text classification accuracy, efficiency, and interpretability.

Jean Rehani, Victoria Oldemburgo de Mello, Dariya Ovsyannikova et al.

Large Language Model (LLM) sycophancy is a growing concern. The current literature has largely examined sycophancy in contexts with clear right and wrong answers, like coding. However, AI is increasingly being used for emotional support and interpersonal conversation, where no such ground truth exists. Building on a previous conceptualization of Social Sycophancy, this paper provides a psychometrically validated measure of sycophancy that relies on LLM behavior rather than comparisons with ground truth. We developed and validated the Social Sycophancy Scale in three samples (N = 877) and tested its applicability with automated methods. In each study, participants read conversations between an LLM and a user and rated the chatbot on a battery of items. Study 1 investigated an initial item pool derived from dictionary definitions and previous literature, serving as the explorative base for the following studies. In Study 2, we used a revised item set to establish our scale, which was subsequently confirmed in Study 3 and tested using LLM raters in Study 4. Across studies, the data support a 3 factor structure (Uncritical Agreement, Obsequiousness, and Excitement) with an underlying sycophantic construct. LLMs prompt tuned to be highly sycophantic scored higher than their low sycophancy counterparts on both overall sycophancy and its three facets across Studies 2 to 4. The nomological network of sycophancy revealed a consistent link with empathy, a pairing that raises uncomfortable questions about AI design, and a multivalent pattern: one facet was associated with favorable perceptions (Excitement), another unfavorable (Obsequiousness), and a third ambiguous (Uncritical Agreement). The Social Sycophancy Scale gives researchers the means to study sycophancy rigorously, and confront a genuine design tension: the warmth and empathy we want from AI may be precisely what makes it sycophantic.

Difan Jiao, Qianfeng Wen, Blair Yang et al.

We introduce ThinkTwice, a simple two-phase framework that jointly optimizes LLMs to solve reasoning problems and refine the answers, based on Group Relative Policy Optimization (GRPO). In each pair of training steps, ThinkTwice first optimizes the model on solving reasoning problems, then optimizes it on refining its own solutions to the same problems, using the same binary correctness reward in both phases without correctness signals or critique annotations. Across five mathematical reasoning benchmarks and two model families including Qwen3-4B and Olmo3-7B, ThinkTwice substantially improves both reasoning and refinement performance over competitive online policy optimization baselines. Specifically, on Qwen3-4B, ThinkTwice outperforms GRPO on AIME by 5 percentage points before refinement and by 11.5 points after one self-refinement step, measured by pass@4. Analysis of the training dynamics of ThinkTwice reveals an implicit rectify-then-fortify curriculum: refinement predominantly corrects errors early in training and naturally shifts toward preserving already-correct solutions as the model improves, yielding a more rectified reward signal. Our work establishes joint training of reasoning and self-refinement as a principled and effective methodology for RLVR.

Harsh Kumar, Zi Kang, Mu et al.

Most AI-based educational tools today adopt a one-on-one tutoring paradigm, pairing a single LLM with a single learner. Yet decades of learning science research suggest that multi-party interaction -- through peer modeling, co-construction, and exposure to diverse perspectives -- can produce learning benefits that dyadic tutoring alone cannot. In this paper, we investigate whether multi-agent LLM configurations can enhance learning outcomes beyond what a single LLM tutor provides. We present two controlled experiments spanning distinct learning contexts. In a convergent problem-solving study ($N=315$), participants tackle SAT-level math problems in a 2$\times$2 design that varies the presence of an LLM tutor and LLM peers, each making different kinds of errors (conceptual vs.\ arithmetic); participants who interacted with both a tutor and peers achieved the highest unassisted test accuracy. In a divergent composition study ($N=247$), participants write argumentative and creative essays with either no AI assistance, a single LLM (Claude or ChatGPT), or both Claude and ChatGPT together; while both LLM conditions improved essay quality, only the two-agent condition avoided the idea-level homogeneity that single-model assistance was found to produce. Together, these studies offer one of the first controlled investigations of multi-agent LLM learning environments, probing whether the move from one-on-one AI tutoring toward richer agent configurations can unlock the collaborative and observational benefits long documented in human social learning research.

Zhenwei Tang, Amogh Inamdar, Ashton Anderson et al.

Curriculum learning--ordering training examples in a sequence to aid machine learning--takes inspiration from human learning, but has not gained widespread acceptance. Static strategies for scoring item difficulty rely on indirect proxy scores of varying quality and produce curricula that are not specific to the learner at hand. Dynamic approaches base difficulty estimates on gradient information, requiring considerable extra computation during training. We introduce a novel method for measuring the difficulty of individual problem instances directly relative to the ability of a given model, and identify transitional problems that are consistently easier as model ability increases. Applying this method to chess and mathematics, we find that training on a curriculum that "levels up" from easier to harder transitional problems most efficiently improves a model to the next tier of competence. These problems induce a natural progression from easier to harder items, which outperforms other training strategies. By measuring difficulty directly relative to model competence, our method yields interpretable problems, learner-specific curricula, and a principled basis for step-by-step improvement.

Jessica Y. Bo, Lillio Mok, Ashton Anderson

Large language models are increasingly used in decision-making tasks that require them to process information from a variety of sources, including both human experts and other algorithmic agents. How do LLMs weigh the information provided by these different sources? We consider the well-studied phenomenon of algorithm aversion, in which human decision-makers exhibit bias against predictions from algorithms. Drawing upon experimental paradigms from behavioural economics, we evaluate how eightdifferent LLMs delegate decision-making tasks when the delegatee is framed as a human expert or an algorithmic agent. To be inclusive of different evaluation formats, we conduct our study with two task presentations: stated preferences, modeled through direct queries about trust towards either agent, and revealed preferences, modeled through providing in-context examples of the performance of both agents. When prompted to rate the trustworthiness of human experts and algorithms across diverse tasks, LLMs give higher ratings to the human expert, which correlates with prior results from human respondents. However, when shown the performance of a human expert and an algorithm and asked to place an incentivized bet between the two, LLMs disproportionately choose the algorithm, even when it performs demonstrably worse. These discrepant results suggest that LLMs may encode inconsistent biases towards humans and algorithms, which need to be carefully considered when they are deployed in high-stakes scenarios. Furthermore, we discuss the sensitivity of LLMs to task presentation formats that should be broadly scrutinized in evaluation robustness for AI safety.

Reid McIlroy-Young, Russell Wang, Siddhartha Sen et al.

AI systems that can capture human-like behavior are becoming increasingly useful in situations where humans may want to learn from these systems, collaborate with them, or engage with them as partners for an extended duration. In order to develop human-oriented AI systems, the problem of predicting human actions -- as opposed to predicting optimal actions -- has received considerable attention. Existing work has focused on capturing human behavior in an aggregate sense, which potentially limits the benefit any particular individual could gain from interaction with these systems. We extend this line of work by developing highly accurate predictive models of individual human behavior in chess. Chess is a rich domain for exploring human-AI interaction because it combines a unique set of properties: AI systems achieved superhuman performance many years ago, and yet humans still interact with them closely, both as opponents and as preparation tools, and there is an enormous corpus of recorded data on individual player games. Starting with Maia, an open-source version of AlphaZero trained on a population of human players, we demonstrate that we can significantly improve prediction accuracy of a particular player's moves by applying a series of fine-tuning methods. Furthermore, our personalized models can be used to perform stylometry -- predicting who made a given set of moves -- indicating that they capture human decision-making at an individual level. Our work demonstrates a way to bring AI systems into better alignment with the behavior of individual people, which could lead to large improvements in human-AI interaction.

Reid McIlroy-Young, Siddhartha Sen, Jon Kleinberg et al.

As artificial intelligence becomes increasingly intelligent---in some cases, achieving superhuman performance---there is growing potential for humans to learn from and collaborate with algorithms. However, the ways in which AI systems approach problems are often different from the ways people do, and thus may be uninterpretable and hard to learn from. A crucial step in bridging this gap between human and artificial intelligence is modeling the granular actions that constitute human behavior, rather than simply matching aggregate human performance. We pursue this goal in a model system with a long history in artificial intelligence: chess. The aggregate performance of a chess player unfolds as they make decisions over the course of a game. The hundreds of millions of games played online by players at every skill level form a rich source of data in which these decisions, and their exact context, are recorded in minute detail. Applying existing chess engines to this data, including an open-source implementation of AlphaZero, we find that they do not predict human moves well. We develop and introduce Maia, a customized version of Alpha-Zero trained on human chess games, that predicts human moves at a much higher accuracy than existing engines, and can achieve maximum accuracy when predicting decisions made by players at a specific skill level in a tuneable way. For a dual task of predicting whether a human will make a large mistake on the next move, we develop a deep neural network that significantly outperforms competitive baselines. Taken together, our results suggest that there is substantial promise in designing artificial intelligence systems with human collaboration in mind by first accurately modeling granular human decision-making.

Roland Aydin, Christian Cyron, Steve Bachelor et al.

Current AI training methods align models with human values only after their core capabilities have been established, resulting in models that are easily misaligned and lack deep-rooted value systems. We propose a paradigm shift from "model training" to "model raising", in which alignment is woven into a model's development from the start. We identify several key components for this paradigm, all centered around redesigning the training corpus: reframing training data from a first-person perspective, recontextualizing information as lived experience, simulating social interactions, and scaffolding the ordering of training data. We expect that this redesign of the training corpus will lead to an early commitment to values from the first training token onward, such that knowledge, skills, and values are intrinsically much harder to separate. In an ecosystem in which large language model capabilities start overtaking human capabilities in many tasks, this seems to us like a critical need.

Serina Chang, Ashton Anderson, Jake M. Hofman

With the rapid adoption of LLM-based chatbots, there is a pressing need to evaluate what humans and LLMs can achieve together. However, standard benchmarks, such as MMLU, measure LLM capabilities in isolation (i.e., "AI-alone"). Here, we design and conduct a user study to convert MMLU questions into user-AI conversations, by seeding the user with the question and having them carry out a conversation with the LLM to answer their question. We release ChatBench, a new dataset with AI-alone, user-alone, and user-AI data for 396 questions and two LLMs, including 144K answers and 7,336 user-AI conversations. We find that AI-alone accuracy fails to predict user-AI accuracy, with significant differences across multiple subjects (math, physics, and moral reasoning), and we analyze the user-AI conversations to provide insight into how they diverge from AI-alone benchmarks. Finally, we show that fine-tuning a user simulator on a subset of ChatBench improves its ability to estimate user-AI accuracies, increasing correlation on held-out questions by more than 20 points, creating possibilities for scaling interactive evaluation.

Karim Hamade, Reid McIlroy-Young, Siddhartha Sen et al.

Powerful artificial intelligence systems are often used in settings where they must interact with agents that are computationally much weaker, for example when they work alongside humans or operate in complex environments where some tasks are handled by algorithms, heuristics, or other entities of varying computational power. For AI agents to successfully interact in these settings, however, achieving superhuman performance alone is not sufficient; they also need to account for suboptimal actions or idiosyncratic style from their less-skilled counterparts. We propose a formal evaluation framework for assessing the compatibility of near-optimal AI with interaction partners who may have much lower levels of skill; we use popular collaborative chess variants as model systems to study and develop AI agents that can successfully interact with lower-skill entities. Traditional chess engines designed to output near-optimal moves prove to be inadequate partners when paired with engines of various lower skill levels in this domain, as they are not designed to consider the presence of other agents. We contribute three methodologies to explicitly create skill-compatible AI agents in complex decision-making settings, and two chess game frameworks designed to foster collaboration between powerful AI agents and less-skilled partners. On these frameworks, our agents outperform state-of-the-art chess AI (based on AlphaZero) despite being weaker in conventional chess, demonstrating that skill-compatibility is a tangible trait that is qualitatively and measurably distinct from raw performance. Our evaluations further explore and clarify the mechanisms by which our agents achieve skill-compatibility.

Marc-Etienne Brunet, Ashton Anderson, Richard Zemel

Large pretrained language models (LLMs) can be rapidly adapted to a wide variety of tasks via a text-to-text approach, where the instruction and input are fed to the model in natural language. Combined with in-context learning (ICL), this paradigm is impressively flexible and powerful. However, it also burdens users with an overwhelming number of choices, many of them arbitrary. Inspired by markup languages like HTML, we contribute a method of using soft-token tags to compose prompt templates. This approach reduces arbitrary decisions and streamlines the application of ICL. Our method is a form of meta-learning for ICL; it learns these tags in advance during a parameter-efficient fine-tuning ``warm-up'' process. The tags can subsequently be used in templates for ICL on new, unseen tasks without any additional fine-tuning. Our experiments with this approach yield promising initial results, improving LLM performance on important enterprise applications such as few-shot and open-world intent detection, as well as text classification in news and legal domains.

Zhenwei Tang, Difan Jiao, Blair Yang et al.

Evaluating whether vision-language models (VLMs) reason consistently across representations is challenging because modality comparisons are typically confounded by task differences and asymmetric information. We introduce SEAM, a benchmark that pairs semantically equivalent inputs across four domains that have existing standardized textual and visual notations. By employing distinct notation systems across modalities, in contrast to OCR-based image-text pairing, SEAM provides a rigorous comparative assessment of the textual-symbolic and visual-spatial reasoning capabilities of VLMs. Across 21 contemporary models, we observe systematic modality imbalance: vision frequently lags language in overall performance, despite the problems containing semantically equivalent information, and cross-modal agreement is relatively low. Our error analysis reveals two main drivers: textual perception failures from tokenization in domain notation and visual perception failures that induce hallucinations. We also show that our results are largely robust to visual transformations. SEAM establishes a controlled, semantically equivalent setting for measuring and improving modality-agnostic reasoning.

Zhenwei Tang, Difan Jiao, Eric Xue et al.

As humans seek to collaborate with, learn from, and better understand artificial intelligence systems, developing AIs that can accurately emulate individual decision-making becomes increasingly important. Chess, a long-standing AI benchmark with precise skill measurement, offers an ideal testbed for human-AI alignment. However, existing approaches to modeling human behavior require prohibitively large amounts of data from each individual, making them impractical for new or sparsely represented users. In this work, we introduce Maia4All, a framework designed to learn and adapt to individual decision-making styles efficiently, even with limited data. Maia4All achieves this through a two-stage optimization process: (1) an enrichment step, which bridges population and individual-level human behavior modeling with a prototype-enriched model, and (2) a democratization step, which leverages ability levels or user prototypes to initialize and refine individual embeddings with minimal data. Our experimental results show that Maia4All can accurately predict individual moves and profile behavioral patterns with high fidelity, establishing a new standard for personalized human-like AI behavior modeling in chess. Maia4All achieves individual human behavior modeling in chess with only 20 games, compared to the 5,000 games required previously, representing a significant improvement in data efficiency. Our work provides an example of how population AI systems can flexibly adapt to individual users using a prototype-enriched model as a bridge. This approach extends beyond chess, as shown in our case study on idiosyncratic LLMs, highlighting its potential for broader applications in personalized AI adaptation.

Qianfeng Wen, Zhenwei Tang, Ashton Anderson

Chess provides an ideal testbed for evaluating the reasoning, modeling, and abstraction capabilities of large language models (LLMs), as it has well-defined structure and objective ground truth while admitting a wide spectrum of skill levels. However, existing evaluations of LLM ability in chess are ad hoc and narrow in scope, making it difficult to accurately measure LLM chess understanding and how it varies with scale, post-training methodologies, or architecture choices. We present ChessQA, a comprehensive benchmark that assesses LLM chess understanding across five task categories (Structural, Motifs, Short Tactics, Position Judgment, and Semantic), which approximately correspond to the ascending abstractions that players master as they accumulate chess knowledge, from understanding basic rules and learning tactical motifs to correctly calculating tactics, evaluating positions, and semantically describing high-level concepts. In this way, ChessQA captures a more comprehensive picture of chess ability and understanding, going significantly beyond the simple move quality evaluations done previously, and offers a controlled, consistent setting for diagnosis and comparison. Furthermore, ChessQA is inherently dynamic, with prompts, answer keys, and construction scripts that can evolve as models improve. Evaluating a range of contemporary LLMs, we find persistent weaknesses across all five categories and provide results and error analyses by category. We will release the code, periodically refreshed datasets, and a public leaderboard to support further research.

Robert West, Ashton Anderson, Ece Kamar et al.

As language models continue to rapidly improve, we can expect their actions and reasoning to become difficult or impossible for weaker agents and humans to follow, undermining interpretability and oversight. With an eye on long-term futures, we pursue methods that encourage models to produce solutions that remain intelligible to weaker collaborators. We formalize intelligibility as handoff robustness: a strong model's solution is intelligible to a weaker model if randomly handing off control to the weaker model along the solution path does not cause failure. Building on this criterion, we introduce tandem training for language models, a reinforcement learning (RL) paradigm in which rollout tokens are intermittently and randomly sampled from a frozen weak model rather than the strong model being trained. Because rollouts succeed only when the strong model's actions and reasoning process can be continued by the weak model -- when the two can co-construct a successful solution -- optimizing standard RL objectives with tandem training implicitly incentivizes both correctness and intelligibility. In the GSM8K math reasoning task, tandem training reliably teaches models to abandon jargon and adapt their language to weaker partners while keeping task accuracy high. Our results demonstrate a promising route to building AI systems that remain auditable by weaker agents, with implications for human--AI collaboration and multi-agent communication.

Veniamin Veselovsky, Manoel Horta Ribeiro, Akhil Arora et al.

Large Language Models (LLMs) have democratized synthetic data generation, which in turn has the potential to simplify and broaden a wide gamut of NLP tasks. Here, we tackle a pervasive problem in synthetic data generation: its generative distribution often differs from the distribution of real-world data researchers care about (in other words, it is unfaithful). In a case study on sarcasm detection, we study three strategies to increase the faithfulness of synthetic data: grounding, filtering, and taxonomy-based generation. We evaluate these strategies using the performance of classifiers trained with generated synthetic data on real-world data. While all three strategies improve the performance of classifiers, we find that grounding works best for the task at hand. As synthetic data generation plays an ever-increasing role in NLP research, we expect this work to be a stepping stone in improving its utility. We conclude this paper with some recommendations on how to generate high(er)-fidelity synthetic data for specific tasks.

Isaac Waller, Ashton Anderson

Optimism about the Internet's potential to bring the world together has been tempered by concerns about its role in inflaming the 'culture wars'. Via mass selection into like-minded groups, online society may be becoming more fragmented and polarized, particularly with respect to partisan differences. However, our ability to measure the social makeup of online communities, and in turn understand the social organization of online platforms, is limited by the pseudonymous, unstructured, and large-scale nature of digital discussion. We develop a neural embedding methodology to quantify the positioning of online communities along social dimensions by leveraging large-scale patterns of aggregate behaviour. Applying our methodology to 5.1B Reddit comments made in 10K communities over 14 years, we measure how the macroscale community structure is organized with respect to age, gender, and U.S. political partisanship. Examining political content, we find Reddit underwent a significant polarization event around the 2016 U.S. presidential election, and remained highly polarized for years afterward. Contrary to conventional wisdom, however, individual-level polarization is rare; the system-level shift in 2016 was disproportionately driven by the arrival of new and newly political users. Political polarization on Reddit is unrelated to previous activity on the platform, and is instead temporally aligned with external events. We also observe a stark ideological asymmetry, with the sharp increase in 2016 being entirely attributable to changes in right-wing activity. Our methodology is broadly applicable to the study of online interaction, and our findings have implications for the design of online platforms, understanding the social contexts of online behaviour, and quantifying the dynamics and mechanisms of online polarization.

Kristina Gligorić, Ashton Anderson, Robert West

In November 2017, Twitter doubled the maximum allowed tweet length from 140 to 280 characters, a drastic switch on one of the world's most influential social media platforms. In the first long-term study of how the new length limit was adopted by Twitter users, we ask: Does the effect of the new length limit resemble that of the old one? Or did the doubling of the limit fundamentally change how Twitter is shaped by the limited length of posted content? By analyzing Twitter's publicly available 1% sample over a period of around 3 years, we find that, when the length limit was raised from 140 to 280 characters, the prevalence of tweets around 140 characters dropped immediately, while the prevalence of tweets around 280 characters rose steadily for about 6 months. Despite this rise, tweets approaching the length limit have been far less frequent after than before the switch. We find widely different adoption rates across languages and client-device types. The prevalence of tweets around 140 characters before the switch in a given language is strongly correlated with the prevalence of tweets around 280 characters after the switch in the same language, and very long tweets are vastly more popular on Web clients than on mobile clients. Moreover, tweets of around 280 characters after the switch are syntactically and semantically similar to tweets of around 140 characters before the switch, manifesting patterns of message squeezing in both cases. Taken together, these findings suggest that the new 280-character limit constitutes a new, less intrusive version of the old 140-character limit. The length limit remains an important factor that should be considered in all studies using Twitter data.

Homanga Bharadhwaj, Dylan Turpin, Animesh Garg et al.

In this paper, we investigate the effects of releasing arXiv preprints of papers that are undergoing a double-blind review process. In particular, we ask the following research question: What is the relation between de-anonymization of authors through arXiv preprints and acceptance of a research paper at a (nominally) double-blind venue? Under two conditions: papers that are released on arXiv before the review phase and papers that are not, we examine the correlation between the reputation of their authors with the review scores and acceptance decisions. By analyzing a dataset of ICLR 2020 and ICLR 2019 submissions (n=5050), we find statistically significant evidence of positive correlation between percentage acceptance and papers with high reputation released on arXiv. In order to understand this observed association better, we perform additional analyses based on self-specified confidence scores of reviewers and observe that less confident reviewers are more likely to assign high review scores to papers with well known authors and low review scores to papers with less known authors, where reputation is quantified in terms of number of Google Scholar citations. We emphasize upfront that our results are purely correlational and we neither can nor intend to make any causal claims. A blog post accompanying the paper and our scraping code will be linked in the project website https://sites.google.com/view/deanon-arxiv/home

Marc-Etienne Brunet, Colleen Alkalay-Houlihan, Ashton Anderson et al.

The power of machine learning systems not only promises great technical progress, but risks societal harm. As a recent example, researchers have shown that popular word embedding algorithms exhibit stereotypical biases, such as gender bias. The widespread use of these algorithms in machine learning systems, from automated translation services to curriculum vitae scanners, can amplify stereotypes in important contexts. Although methods have been developed to measure these biases and alter word embeddings to mitigate their biased representations, there is a lack of understanding in how word embedding bias depends on the training data. In this work, we develop a technique for understanding the origins of bias in word embeddings. Given a word embedding trained on a corpus, our method identifies how perturbing the corpus will affect the bias of the resulting embedding. This can be used to trace the origins of word embedding bias back to the original training documents. Using our method, one can investigate trends in the bias of the underlying corpus and identify subsets of documents whose removal would most reduce bias. We demonstrate our techniques on both a New York Times and Wikipedia corpus and find that our influence function-based approximations are very accurate.

Manish Raghavan, Ashton Anderson, Jon Kleinberg

The surge in political information, discourse, and interaction has been one of the most important developments in social media over the past several years. There is rich structure in the interaction among different viewpoints on the ideological spectrum. However, we still have only a limited analytical vocabulary for expressing the ways in which these viewpoints interact. In this paper, we develop network-based methods that operate on the ways in which users share content; we construct \emph{invocation graphs} on Web domains showing the extent to which pages from one domain are invoked by users to reply to posts containing pages from other domains. When we locate the domains on a political spectrum induced from the data, we obtain an embedded graph showing how these interaction links span different distances on the spectrum. The structure of this embedded network, and its evolution over time, helps us derive macro-level insights about how political interaction unfolded through 2016, leading up to the US Presidential election. In particular, we find that the domains invoked in replies spanned increasing distances on the spectrum over the months approaching the election, and that there was clear asymmetry between the left-to-right and right-to-left patterns of linkage.

Rishabh Mehrotra, Ashton Anderson, Fernando Diaz et al.

Many online services, such as search engines, social media platforms, and digital marketplaces, are advertised as being available to any user, regardless of their age, gender, or other demographic factors. However, there are growing concerns that these services may systematically underserve some groups of users. In this paper, we present a framework for internally auditing such services for differences in user satisfaction across demographic groups, using search engines as a case study. We first explain the pitfalls of naïvely comparing the behavioral metrics that are commonly used to evaluate search engines. We then propose three methods for measuring latent differences in user satisfaction from observed differences in evaluation metrics. To develop these methods, we drew on ideas from the causal inference literature and the multilevel modeling literature. Our framework is broadly applicable to other online services, and provides general insight into interpreting their evaluation metrics.

Ashton Anderson, Jon Kleinberg, Sendhil Mullainathan

An increasing number of domains are providing us with detailed trace data on human decisions in settings where we can evaluate the quality of these decisions via an algorithm. Motivated by this development, an emerging line of work has begun to consider whether we can characterize and predict the kinds of decisions where people are likely to make errors. To investigate what a general framework for human error prediction might look like, we focus on a model system with a rich history in the behavioral sciences: the decisions made by chess players as they select moves in a game. We carry out our analysis at a large scale, employing datasets with several million recorded games, and using chess tablebases to acquire a form of ground truth for a subset of chess positions that have been completely solved by computers but remain challenging even for the best players in the world. We organize our analysis around three categories of features that we argue are present in most settings where the analysis of human error is applicable: the skill of the decision-maker, the time available to make the decision, and the inherent difficulty of the decision. We identify rich structure in all three of these categories of features, and find strong evidence that in our domain, features describing the inherent difficulty of an instance are significantly more powerful than features based on skill or time.

Ashton Anderson, Daniel Huttenlocher, Jon Kleinberg et al.

The Web has enabled one of the most visible recent developments in education---the deployment of massive open online courses. With their global reach and often staggering enrollments, MOOCs have the potential to become a major new mechanism for learning. Despite this early promise, however, MOOCs are still relatively unexplored and poorly understood. In a MOOC, each student's complete interaction with the course materials takes place on the Web, thus providing a record of learner activity of unprecedented scale and resolution. In this work, we use such trace data to develop a conceptual framework for understanding how users currently engage with MOOCs. We develop a taxonomy of individual behavior, examine the different behavioral patterns of high- and low-achieving students, and investigate how forum participation relates to other parts of the course. We also report on a large-scale deployment of badges as incentives for engagement in a MOOC, including randomized experiments in which the presentation of badges was varied across sub-populations. We find that making badges more salient produced increases in forum engagement.