Multimodal Large Language Models (MLLMs) have recently shown remarkable perceptual capability in answering visual questions, however, little is known about the limits of their perception. In particular, while prior works have provided anecdotal evidence of MLLMs' sensitivity to object size, this phenomenon and its underlying causes have not been explored comprehensively. In this work, we quantitatively study the perception of small visual objects in several state-of-the-art MLLMs and reveal a pervasive limitation in answering questions about small objects in images. Next, we identify four independent factors that can contribute to this limitation -- object quality, size, distractors, and location -- and conduct controlled intervention studies to measure the effect of each factor on MLLMs' perception. In particular, we find that lower object quality and smaller object size can both independently reduce MLLMs' ability to answer visual questions. More surprisingly, we find that the location of the object in the image and the presence of visual distractors can also significantly reduce MLLMs' question answering accuracy. Our study provides a better understanding of the perceptual limitation of MLLMs and contributes new evaluation protocols for analyzing the perception of future MLLMs. To facilitate further investigations, we release our code and data.

Laughter is one of the most expressive and natural aspects of human speech, conveying emotions, social cues, and humor. However, most text-to-speech (TTS) systems lack the ability to produce realistic and appropriate laughter sounds, limiting their applications and user experience. While there have been prior works to generate natural laughter, they fell short in terms of controlling the timing and variety of the laughter to be generated. In this work, we propose ELaTE, a zero-shot TTS that can generate natural laughing speech of any speaker based on a short audio prompt with precise control of laughter timing and expression. Specifically, ELaTE works on the audio prompt to mimic the voice characteristic, the text prompt to indicate the contents of the generated speech, and the input to control the laughter expression, which can be either the start and end times of laughter, or the additional audio prompt that contains laughter to be mimicked. We develop our model based on the foundation of conditional flow-matching-based zero-shot TTS, and fine-tune it with frame-level representation from a laughter detector as additional conditioning. With a simple scheme to mix small-scale laughter-conditioned data with large-scale pre-training data, we demonstrate that a pre-trained zero-shot TTS model can be readily fine-tuned to generate natural laughter with precise controllability, without losing any quality of the pre-trained zero-shot TTS model. Through the evaluations, we show that ELaTE can generate laughing speech with significantly higher quality and controllability compared to conventional models. See https://aka.ms/elate/ for demo samples.

This study evaluates the ability of Large Language Model (LLM)-based Subpopulation Representative Models (SRMs) to generalize from empirical data, utilizing in-context learning with data from the 2016 and 2020 American National Election Studies. We explore generalization across response variables and demographic subgroups. While conditioning with empirical data improves performance on the whole, the benefit of in-context learning varies considerably across demographics, sometimes hurting performance for one demographic while helping performance for others. The inequitable benefits of in-context learning for SRM present a challenge for practitioners implementing SRMs, and for decision-makers who might come to rely on them. Our work highlights a need for fine-grained benchmarks captured from diverse subpopulations that test not only fidelity but generalization.

The vast majority of discourse around AI development assumes that subservient, "moral" models aligned with "human values" are universally beneficial -- in short, that good AI is sycophantic AI. We explore the shadow of the sycophantic paradigm, a design space we term antagonistic AI: AI systems that are disagreeable, rude, interrupting, confrontational, challenging, etc. -- embedding opposite behaviors or values. Far from being "bad" or "immoral," we consider whether antagonistic AI systems may sometimes have benefits to users, such as forcing users to confront their assumptions, build resilience, or develop healthier relational boundaries. Drawing from formative explorations and a speculative design workshop where participants designed fictional AI technologies that employ antagonism, we lay out a design space for antagonistic AI, articulating potential benefits, design techniques, and methods of embedding antagonistic elements into user experience. Finally, we discuss the many ethical challenges of this space and identify three dimensions for the responsible design of antagonistic AI -- consent, context, and framing.

We ideate a future design workflow that involves AI technology. Drawing from activity and communication theory, we attempt to isolate the new value large AI models can provide design compared to past technologies. We arrive at three affordances -- dynamic grounding, constructive negotiation, and sustainable motivation -- that summarize latent qualities of natural language-enabled foundation models that, if explicitly designed for, can support the process of design. Through design fiction, we then imagine a future interface as a diegetic prototype, the story of Squirrel Game, that demonstrates each of our three affordances in a realistic usage scenario. Our design process, terminology, and diagrams aim to contribute to future discussions about the relative affordances of AI technology with regard to collaborating with human designers.

This position paper investigates the integration of Differential Privacy (DP) in the training of Mixture of Experts (MoE) models within the field of natural language processing. As Large Language Models (LLMs) scale to billions of parameters, leveraging expansive datasets, they exhibit enhanced linguistic capabilities and emergent abilities. However, this growth raises significant computational and privacy concerns. Our study addresses these issues by exploring the potential of MoE models, known for their computational efficiency, and the application of DP, a standard for privacy preservation. We present the first known attempt to train MoE models under the constraints of DP, addressing the unique challenges posed by their architecture and the complexities of DP integration. Our initial experimental studies demonstrate that MoE models can be effectively trained with DP, achieving performance that is competitive with their non-private counterparts. This initial study aims to provide valuable insights and ignite further research in the domain of privacy-preserving MoE models, softly laying the groundwork for prospective developments in this evolving field.

How do transformer-based large language models (LLMs) store and retrieve knowledge? We focus on the most basic form of this task -- factual recall, where the model is tasked with explicitly surfacing stored facts in prompts of form `Fact: The Colosseum is in the country of'. We find that the mechanistic story behind factual recall is more complex than previously thought. It comprises several distinct, independent, and qualitatively different mechanisms that additively combine, constructively interfering on the correct attribute. We term this generic phenomena the additive motif: models compute through summing up multiple independent contributions. Each mechanism's contribution may be insufficient alone, but summing results in constructive interfere on the correct answer. In addition, we extend the method of direct logit attribution to attribute an attention head's output to individual source tokens. We use this technique to unpack what we call `mixed heads' -- which are themselves a pair of two separate additive updates from different source tokens.

In this work, we study the issue of reward hacking on the response length, a challenge emerging in Reinforcement Learning from Human Feedback (RLHF) on LLMs. A well-formatted, verbose but less helpful response from the LLMs can often deceive LLMs or even human evaluators to achieve high scores. The same issue also holds for some reward models in RL. To address the challenges in both training and evaluation, we establish a more reliable evaluation protocol for comparing different training configurations, which inspects the trade-off between LLM evaluation score and response length obtained by varying training hyperparameters. Based on this evaluation, we conduct large-scale studies, where the results shed insights into the efficacy of hyperparameters and tricks used in RL on mitigating length bias. We further propose to improve the reward model by jointly training two linear heads on shared feature representations to predict the rewards, one trained to correlate with length, and the other trained to decorrelate with length and therefore focus more on the actual content. We then discard the length head in RL to prevent reward hacking on length. Experiments demonstrate that our approach almost eliminates the reward correlation with length, and improves the obtained policy by a significant margin.

Reinforcement Learning from Human Feedback (RLHF) learns from the preference signal provided by a probabilistic preference model, which takes a prompt and two responses as input, and produces a score indicating the preference of one response against another. So far, the most popular RLHF paradigm is reward-based, which starts with an initial step of reward modeling, and the constructed reward is then used to provide a reward signal for the subsequent reward optimization stage. However, the existence of a reward function is a strong assumption and the reward-based RLHF is limited in expressivity and cannot capture the real-world complicated human preference. In this work, we provide theoretical insights for a recently proposed learning paradigm, Nash learning from human feedback (NLHF), which considered a general preference model and formulated the alignment process as a game between two competitive LLMs. The learning objective is to find a policy that consistently generates responses preferred over any competing policy while staying close to the initial model. The objective is defined as the Nash equilibrium (NE) of the KL-regularized preference model. We aim to make the first attempt to study the theoretical learnability of the KL-regularized NLHF by considering both offline and online settings. For the offline learning from a pre-collected dataset, we propose algorithms that are efficient under suitable coverage conditions of the dataset. For batch online learning from iterative interactions with a preference oracle, our proposed algorithm enjoys a finite sample guarantee under the structural condition of the underlying preference model. Our results connect the new NLHF paradigm with traditional RL theory, and validate the potential of reward-model-free learning under general preference.

In day-to-day communication, people often approximate the truth -- for example, rounding the By treating these objectives as separate, standard approaches time or omitting details -- in order to be maximally implicitly assume that honesty and helpfulness are jointly helpful to the listener. How do large language achievable. In everyday conversation, however, they can be models (LLMs) handle such nuanced tradeoffs? in tension (Wilson & Sperber, 2002). People regularly round To address this question, we use psychological numerical values like time (Van Der Henst et al., 2002), models and experiments designed to characterize distance (Krifka, 2007), and monetary values (Kao et al., human behavior to analyze LLMs. We test a 2014), or endorse false generalizations (van Rooij & Schulz, range of LLMs and explore how optimization for 2020) when they believe these approximations will help the human preferences or inference-time reasoning listener. People also regularly trade literal honesty for other affects these trade-offs. We find that reinforcement communicative goals, as implicit in figures of speech like learning from human feedback improves metaphors (Tendahl & Gibbs Jr, 2008), hyperbole (Carston both honesty and helpfulness, while chain-ofthought & Wearing, 2011) and irony (Popa-Wyatt, 2019).