Hiromu Yakura

Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao et al. · allen-ai, amazon-science

Language models demonstrate both quantitative improvement and new qualitative capabilities with increasing scale. Despite their potentially transformative impact, these new capabilities are as yet poorly characterized. In order to inform future research, prepare for disruptive new model capabilities, and ameliorate socially harmful effects, it is vital that we understand the present and near-future capabilities and limitations of language models. To address this challenge, we introduce the Beyond the Imitation Game benchmark (BIG-bench). BIG-bench currently consists of 204 tasks, contributed by 450 authors across 132 institutions. Task topics are diverse, drawing problems from linguistics, childhood development, math, common-sense reasoning, biology, physics, social bias, software development, and beyond. BIG-bench focuses on tasks that are believed to be beyond the capabilities of current language models. We evaluate the behavior of OpenAI's GPT models, Google-internal dense transformer architectures, and Switch-style sparse transformers on BIG-bench, across model sizes spanning millions to hundreds of billions of parameters. In addition, a team of human expert raters performed all tasks in order to provide a strong baseline. Findings include: model performance and calibration both improve with scale, but are poor in absolute terms (and when compared with rater performance); performance is remarkably similar across model classes, though with benefits from sparsity; tasks that improve gradually and predictably commonly involve a large knowledge or memorization component, whereas tasks that exhibit "breakthrough" behavior at a critical scale often involve multiple steps or components, or brittle metrics; social bias typically increases with scale in settings with ambiguous context, but this can be improved with prompting.

Hiromu Yakura, Ezequiel Lopez-Lopez, Levin Brinkmann et al.

From the invention of writing and the printing press, to television and social media, human history is punctuated by major innovations in communication technology, which fundamentally altered how ideas spread and reshaped our culture. Recent chatbots powered by generative artificial intelligence constitute a novel medium that encodes cultural patterns in their neural representations and disseminates them in conversations with hundreds of millions of people. Understanding whether these patterns transmit into human language, and ultimately shape human culture, is a fundamental question. While fully quantifying the causal impact of a chatbot like ChatGPT on human culture is very challenging, lexicographic shift in human spoken communication may offer an early indicator of such broad phenomenon. Here, we apply econometric causal inference techniques to 740,249 hours of human discourse from 360,445 YouTube academic talks and 771,591 conversational podcast episodes across multiple disciplines. We detect a measurable and abrupt increase in the use of words preferentially generated by ChatGPT, such as delve, comprehend, boast, swift, and meticulous, after its release. These findings suggest a scenario where machines, originally trained on human data and subsequently exhibiting their own cultural traits, can, in turn, measurably reshape human culture. This marks the beginning of a closed cultural feedback loop in which cultural traits circulate bidirectionally between humans and machines. Our results motivate further research into the evolution of human-machine culture, and raise concerns over the erosion of linguistic and cultural diversity, and the risks of scalable manipulation.

Riku Arakawa, Hiromu Yakura, Masataka Goto

CatAlyst uses generative models to help workers' progress by influencing their task engagement instead of directly contributing to their task outputs. It prompts distracted workers to resume their tasks by generating a continuation of their work and presenting it as an intervention that is more context-aware than conventional (predetermined) feedback. The prompt can function by drawing their interest and lowering the hurdle for resumption even when the generated continuation is insufficient to substitute their work, while recent human-AI collaboration research aiming at work substitution depends on a stable high accuracy. This frees CatAlyst from domain-specific model-tuning and makes it applicable to various tasks. Our studies involving writing and slide-editing tasks demonstrated CatAlyst's effectiveness in helping workers swiftly resume tasks with a lowered cognitive load. The results suggest a new form of human-AI collaboration where large generative models publicly available but imperfect for each individual domain can contribute to workers' digital well-being.

Riku Arakawa, Hiromu Yakura, Mayank Goel

We routinely perform procedures (such as cooking) that include a set of atomic steps. Often, inadvertent omission or misordering of a single step can lead to serious consequences, especially for those experiencing cognitive challenges such as dementia. This paper introduces PrISM-Observer, a smartwatch-based, context-aware, real-time intervention system designed to support daily tasks by preventing errors. Unlike traditional systems that require users to seek out information, the agent observes user actions and intervenes proactively. This capability is enabled by the agent's ability to continuously update its belief in the user's behavior in real-time through multimodal sensing and forecast optimal intervention moments and methods. We first validated the steps-tracking performance of our framework through evaluations across three datasets with different complexities. Then, we implemented a real-time agent system using a smartwatch and conducted a user study in a cooking task scenario. The system generated helpful interventions, and we gained positive feedback from the participants. The general applicability of PrISM-Observer to daily tasks promises broad applications, for instance, including support for users requiring more involved interventions, such as people with dementia or post-surgical patients.

Riku Arakawa, Hiromu Yakura

Recent organizations have started to adopt AI-based decision support tools to optimize human resource development practices, while facing various challenges of using AIs in highly contextual and sensitive domains. We present our case study that aims to help professional assessors make decisions in human assessment, in which they conduct interviews with assessees and evaluate their suitability for certain job roles. Our workshop with two industrial assessors elucidated troubles they face (i.e., maintaining stable and non-subjective observation of assessees' behaviors) and derived requirements of AI systems (i.e., extracting their nonverbal cues from interview videos in an interpretable manner). In response, we employed an unsupervised anomaly detection algorithm using multimodal behavioral features such as facial keypoints, body and head pose, and gaze. The algorithm extracts outlier scenes from the video based on behavioral features as well as informing which feature contributes to the outlierness. We first evaluated how the assessors would perceive the extracted cues and discovered that the algorithm is useful in suggesting scenes to which assessors would pay attention, thanks to its interpretability. Then, we developed an interface prototype incorporating the algorithm and had six assessors use it for their actual assessment. Their comments revealed the effectiveness of introducing unsupervised anomaly detection to enhance their feeling of confidence and objectivity of the assessment along with potential use scenarios of such AI-based systems in human assessment. Our approach, which builds on top of the idea of separating observation and interpretation in human-AI collaboration, will facilitate human decision making in highly contextual domains, such as human assessment, while keeping their trust in the system.

Yuki Koike, Hiroyuki Katsura, Hiromu Yakura et al.

Mutation-based fuzzing has become one of the most common vulnerability discovery solutions over the last decade. Fuzzing can be optimized when targeting specific programs, and given that, some studies have employed online optimization methods to do it automatically, i.e., tuning fuzzers for any given program in a program-agnostic manner. However, previous studies have neither fully explored mutation schemes suitable for online optimization methods, nor online optimization methods suitable for mutation schemes. In this study, we propose an optimization framework called SLOPT that encompasses both a bandit-friendly mutation scheme and mutation-scheme-friendly bandit algorithms. The advantage of SLOPT is that it can generally be incorporated into existing fuzzers, such as AFL and Honggfuzz. As a proof of concept, we implemented SLOPT-AFL++ by integrating SLOPT into AFL++ and showed that the program-agnostic optimization delivered by SLOPT enabled SLOPT-AFL++ to achieve higher code coverage than AFL++ in all of ten real-world FuzzBench programs. Moreover, we ran SLOPT-AFL++ against several real-world programs from OSS-Fuzz and successfully identified three previously unknown vulnerabilities, even though these programs have been fuzzed by AFL++ for a considerable number of CPU days on OSS-Fuzz.

Hiromu Yakura, Masataka Goto

Recent text-to-audio generation techniques have the potential to allow novice users to freely generate music audio. Even if they do not have musical knowledge, such as about chord progressions and instruments, users can try various text prompts to generate audio. However, compared to the image domain, gaining a clear understanding of the space of possible music audios is difficult because users cannot listen to the variations of the generated audios simultaneously. We therefore facilitate users in exploring not only text prompts but also audio priors that constrain the text-to-audio music generation process. This dual-sided exploration enables users to discern the impact of different text prompts and audio priors on the generation results through iterative comparison of them. Our developed interface, IteraTTA, is specifically designed to aid users in refining text prompts and selecting favorable audio priors from the generated audios. With this, users can progressively reach their loosely-specified goals while understanding and exploring the space of possible results. Our implementation and discussions highlight design considerations that are specifically required for text-to-audio models and how interaction techniques can contribute to their effectiveness.

Riku Arakawa, Hiromu Yakura

In this paper, we discuss the potential of applying unsupervised anomaly detection in constructing AI-based interactive systems that deal with highly contextual situations, i.e., human-human communication, in collaboration with domain experts. We reached this approach of utilizing unsupervised anomaly detection through our experience of developing a computational support tool for executive coaching, which taught us the importance of providing interpretable results so that expert coaches can take both the results and contexts into account. The key idea behind this approach is to leave room for expert coaches to unleash their open-ended interpretations, rather than simplifying the nature of social interactions to well-defined problems that are tractable by conventional supervised algorithms. In addition, we found that this approach can be extended to nurturing novice coaches; by prompting them to interpret the results from the system, it can provide the coaches with educational opportunities. Although the applicability of this approach should be validated in other domains, we believe that the idea of leveraging unsupervised anomaly detection to construct AI-based interactive systems would shed light on another direction of human-AI communication.

Hiromu Yakura

Metaphors and sarcasm are precious fruits of our highly evolved social communication skills. However, children with the condition then known as Asperger syndrome are known to have difficulties in comprehending sarcasm, even if they possess adequate verbal IQs for understanding metaphors. Accordingly, researchers had employed a screening test that assesses metaphor and sarcasm comprehension to distinguish Asperger syndrome from other conditions with similar external behaviors (e.g., attention-deficit/hyperactivity disorder). This study employs a standardized test to evaluate recent large language models' (LLMs) understanding of nuanced human communication. The results indicate improved metaphor comprehension with increased model parameters; however, no similar improvement was observed for sarcasm comprehension. Considering that a human's ability to grasp sarcasm has been associated with the amygdala, a pivotal cerebral region for emotional learning, a distinctive strategy for training LLMs would be imperative to imbue them with the ability in a cognitively grounded manner.

Riku Arakawa, Hiromu Yakura

Chatbots' role in fostering self-reflection is now widely recognized, especially in inducing users' behavior change. While the benefits of 24/7 availability, scalability, and consistent responses have been demonstrated in contexts such as healthcare and tutoring to help one form a new habit, their utilization in coaching necessitating deeper introspective dialogue to induce leadership growth remains unexplored. This paper explores the potential of such a chatbot powered by recent Large Language Models (LLMs) in collaboration with professional coaches in the field of executive coaching. Through a design workshop with them and two weeks of user study involving ten coach-client pairs, we explored the feasibility and nuances of integrating chatbots to complement human coaches. Our findings highlight the benefits of chatbots' ubiquity and reasoning capabilities enabled by LLMs while identifying their limitations and design necessities for effective collaboration between human coaches and chatbots. By doing so, this work contributes to the foundation for augmenting one's self-reflective process with prevalent conversational agents through the human-in-the-loop approach.

Alejandro H. Artiles, Hiromu Yakura, Levin Brinkmann et al.

In open-ended domains like art, autonomous agents must generate ideas that are both original and internally coherent, yet current Large Language Models (LLMs) either default to familiar cultural patterns or sacrifice coherence when pushed toward novelty. We address this by introducing the Cultural Alien Sampler (CAS), a concept-selection method that explicitly separates compositional fit from cultural typicality. CAS uses two GPT-2 models fine-tuned on WikiArt concepts: a Concept Coherence Model that scores whether concepts plausibly co-occur within artworks, and a Cultural Context Model that estimates how typical those combinations are within individual artists' bodies of work. CAS targets combinations that are high in coherence and low in typicality, yielding ideas that maintain internal consistency while deviating from learned conventions and embedded cultural context. In a human evaluation (N = 100), our approach outperforms random selection and GPT-4o baselines and achieves performance comparable to human art students in both perceived originality and harmony. Additionally, a quantitative study shows that our method produces more diverse outputs and explores a broader conceptual space than its GPT-4o counterpart, demonstrating that artificial cultural alienness can unlock creative potential in autonomous agents.

Prateek Gupta, Qiankun Zhong, Hiromu Yakura et al.

A growing body of multi-agent studies with Large Language Models (LLMs) explores how norms and cooperation emerge in mixed-motive scenarios, where pursuing individual gain can undermine the collective good. While prior work has explored these dynamics in both richly contextualized simulations and simplified game-theoretic environments, most LLM systems featuring common-pool resource (CPR) games provide agents with explicit reward functions directly tied to their actions. In contrast, human cooperation often emerges without full visibility into payoffs and population, relying instead on heuristics, communication, and punishment. We introduce a CPR simulation framework that removes explicit reward signals and embeds cultural-evolutionary mechanisms: social learning (adopting strategies and beliefs from successful peers) and norm-based punishment, grounded in Ostrom's principles of resource governance. Agents also individually learn from the consequences of harvesting, monitoring, and punishing via environmental feedback, enabling norms to emerge endogenously. We establish the validity of our simulation by reproducing key findings from existing studies on human behavior. Building on this, we examine norm evolution across a $2\times2$ grid of environmental and social initialisations (resource-rich vs. resource-scarce; altruistic vs. selfish) and benchmark how agentic societies comprised of different LLMs perform under these conditions. Our results reveal systematic model differences in sustaining cooperation and norm formation, positioning the framework as a rigorous testbed for studying emergent norms in mixed-motive LLM societies. Such analysis can inform the design of AI systems deployed in social and organizational contexts, where alignment with cooperative norms is critical for stability, fairness, and effective governance of AI-mediated environments.

Raluca Rilla, Tobias Werner, Hiromu Yakura et al.

Online behavioural research faces an emerging threat as participants increasingly turn to large language models (LLMs) for advice, translation, or task delegation: LLM Pollution. We identify three interacting variants through which LLM Pollution threatens the validity and integrity of online behavioural research. First, Partial LLM Mediation occurs when participants make selective use of LLMs for specific aspects of a task, such as translation or wording support, leading researchers to (mis)interpret LLM-shaped outputs as human ones. Second, Full LLM Delegation arises when agentic LLMs complete studies with little to no human oversight, undermining the central premise of human-subject research at a more foundational level. Third, LLM Spillover signifies human participants altering their behaviour as they begin to anticipate LLM presence in online studies, even when none are involved. While Partial Mediation and Full Delegation form a continuum of increasing automation, LLM Spillover reflects second-order reactivity effects. Together, these variants interact and generate cascading distortions that compromise sample authenticity, introduce biases that are difficult to detect post hoc, and ultimately undermine the epistemic grounding of online research on human cognition and behaviour. Crucially, the threat of LLM Pollution is already co-evolving with advances in generative AI, creating an escalating methodological arms race. To address this, we propose a multi-layered response spanning researcher practices, platform accountability, and community efforts. As the challenge evolves, coordinated adaptation will be essential to safeguard methodological integrity and preserve the validity of online behavioural research.

Alejandro Hernandez, Levin Brinkmann, Ignacio Serna et al.

While AI models have demonstrated remarkable capabilities in constrained domains like game strategy, their potential for genuine creativity in open-ended domains like art remains debated. We explore this question by examining how AI can transcend human cognitive limitations in visual art creation. Our research hypothesizes that visual art contains a vast unexplored space of conceptual combinations, constrained not by inherent incompatibility, but by cognitive limitations imposed by artists' cultural, temporal, geographical and social contexts. To test this hypothesis, we present the Alien Recombination method, a novel approach utilizing fine-tuned large language models to identify and generate concept combinations that lie beyond human cognitive availability. The system models and deliberately counteracts human availability bias, the tendency to rely on immediately accessible examples, to discover novel artistic combinations. This system not only produces combinations that have never been attempted before within our dataset but also identifies and generates combinations that are cognitively unavailable to all artists in the domain. Furthermore, we translate these combinations into visual representations, enabling the exploration of subjective perceptions of novelty. Our findings suggest that cognitive unavailability is a promising metric for optimizing artistic novelty, outperforming merely temperature scaling without additional evaluation criteria. This approach uses generative models to connect previously unconnected ideas, providing new insight into the potential of framing AI-driven creativity as a combinatorial problem.

Riku Arakawa, Kiyosu Maeda, Hiromu Yakura

Multimodal scene search of conversations is essential for unlocking valuable insights into social dynamics and enhancing our communication. While experts in conversational analysis have their own knowledge and skills to find key scenes, a lack of comprehensive, user-friendly tools that streamline the processing of diverse multimodal queries impedes efficiency and objectivity. To solve it, we developed Providence, a visual-programming-based tool based on design considerations derived from a formative study with experts. It enables experts to combine various machine learning algorithms to capture human behavioral cues without writing code. Our study showed its preferable usability and satisfactory output with less cognitive load imposed in accomplishing scene search tasks of conversations, verifying the importance of its customizability and transparency. Furthermore, through the in-the-wild trial, we confirmed the objectivity and reusability of the tool transform experts' workflow, suggesting the advantage of expert-AI teaming in a highly human-contextual domain.

Hiromu Yakura, Yuki Koyama, Masataka Goto

Current deep learning techniques for style transfer would not be optimal for design support since their "one-shot" transfer does not fit exploratory design processes. To overcome this gap, we propose parametric transcription, which transcribes an end-to-end style transfer effect into parameter values of specific transformations available in an existing content editing tool. With this approach, users can imitate the style of a reference sample in the tool that they are familiar with and thus can easily continue further exploration by manipulating the parameters. To enable this, we introduce a framework that utilizes an existing pretrained model for style transfer to calculate a perceptual style distance to the reference sample and uses black-box optimization to find the parameters that minimize this distance. Our experiments with various third-party tools, such as Instagram and Blender, show that our framework can effectively leverage deep learning techniques for computational design support.

Riku Arakawa, Hiromu Yakura

A growing number of people are using catch-up TV services rather than watching simultaneously with other audience members at the time of broadcast. However, computational support for such catching-up users has not been well explored. In particular, we are observing an emerging phenomenon in online media consumption experiences in which speculation plays a vital role. As the phenomenon of speculation implicitly assumes simultaneity in media consumption, there is a gap for catching-up users, who cannot directly appreciate the consumption experiences. This conversely suggests that there is potential for computational support to enhance the consumption experiences of catching-up users. Accordingly, we conducted a series of studies to pave the way for developing computational support for catching-up users. First, we conducted semi-structured interviews to understand how people are engaging with speculation during media consumption. As a result, we discovered the distinctive aspects of speculation-based consumption experiences in contrast to social viewing experiences sharing immediate reactions that have been discussed in previous studies. We then designed two prototypes for supporting catching-up users based on our quantitative analysis of Twitter data in regard to reaction- and speculation-based media consumption. Lastly, we evaluated the prototypes in a user experiment and, based on its results, discussed ways to empower catching-up users with computational supports in response to recent transformations in media consumption.

Riku Arakawa, Hiromu Yakura

Explicitly alerting users is not always an optimal intervention, especially when they are not motivated to obey. For example, in video-based learning, learners who are distracted from the video would not follow an alert asking them to pay attention. Inspired by the concept of Mindless Computing, we propose a novel intervention approach, Mindless Attractor, that leverages the nature of human speech communication to help learners refocus their attention without relying on their motivation. Specifically, it perturbs the voice in the video to direct their attention without consuming their conscious awareness. Our experiments not only confirmed the validity of the proposed approach but also emphasized its advantages in combination with a machine learning-based sensing module. Namely, it would not frustrate users even though the intervention is activated by false-positive detection of their attentive state. Our intervention approach can be a reliable way to induce behavioral change in human-AI symbiosis.

Hiromu Yakura

In Japanese idol culture, meet-and-greet events where fans were allowed to handshake with an idol member for several seconds were regarded as its essential component until the spread of COVID-19. Now, idol groups are struggling in the transition of such events to computer-mediated communication because these events had emphasized meeting face-to-face over communicating, as we can infer from their length of time. I anticipated that investigating this emerging transition would provide implications because their communication has a unique characteristic that is distinct from well-studied situations, such as workplace communication and intimate relationships. Therefore, I first conducted a quantitative survey to develop a precise understanding of the transition, and based on its results, had semi-structured interviews with idol fans about their perceptions of the transition. The survey revealed distinctive approaches, including one where fans gathered at a venue but were isolated from the idol member by an acrylic plate and talked via a video call. Then the interviews not only provided answers to why such an approach would be reasonable but also suggested the existence of a large gap between conventional offline events and emerging online events in their perceptions. Based on the results, I discussed how we can develop interaction techniques to support this transition and how we can apply it to other situations outside idol culture, such as computer-mediated performing arts.

Hiromu Yakura, Youhei Akimoto, Jun Sakuma

In adversarial attacks intended to confound deep learning models, most studies have focused on limiting the magnitude of the modification so that humans do not notice the attack. On the other hand, during an attack against autonomous cars, for example, most drivers would not find it strange if a small insect image were placed on a stop sign, or they may overlook it. In this paper, we present a systematic approach to generate natural adversarial examples against classification models by employing such natural-appearing perturbations that imitate a certain object or signal. We first show the feasibility of this approach in an attack against an image classifier by employing generative adversarial networks that produce image patches that have the appearance of a natural object to fool the target model. We also introduce an algorithm to optimize placement of the perturbation in accordance with the input image, which makes the generation of adversarial examples fast and likely to succeed. Moreover, we experimentally show that the proposed approach can be extended to the audio domain, for example, to generate perturbations that sound like the chirping of birds to fool a speech classifier.

Riku Arakawa, Hiromu Yakura

Executive coaching has been drawing more and more attention for developing corporate managers. While conversing with managers, coach practitioners are also required to understand internal states of coachees through objective observations. In this paper, we present REsCUE, an automated system to aid coach practitioners in detecting unconscious behaviors of their clients. Using an unsupervised anomaly detection algorithm applied to multimodal behavior data such as the subject's posture and gaze, REsCUE notifies behavioral cues for coaches via intuitive and interpretive feedback in real-time. Our evaluation with actual coaching scenes confirms that REsCUE provides the informative cues to understand internal states of coachees. Since REsCUE is based on the unsupervised method and does not assume any prior knowledge, further applications beside executive coaching are conceivable using our framework.

Hiromu Yakura, Jun Sakuma

We propose a method to generate audio adversarial examples that can attack a state-of-the-art speech recognition model in the physical world. Previous work assumes that generated adversarial examples are directly fed to the recognition model, and is not able to perform such a physical attack because of reverberation and noise from playback environments. In contrast, our method obtains robust adversarial examples by simulating transformations caused by playback or recording in the physical world and incorporating the transformations into the generation process. Evaluation and a listening experiment demonstrated that our adversarial examples are able to attack without being noticed by humans. This result suggests that audio adversarial examples generated by the proposed method may become a real threat.