Raphaël Millière

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.

Dimitri Coelho Mollo, Raphaël Millière

Large language models (LLMs) produce seemingly meaningful outputs, yet they are trained on text alone without direct interaction with the world. This leads to a modern variant of the classical symbol grounding problem in AI: can LLMs' internal states and outputs be about extra-linguistic reality, independently of the meaning human interpreters project onto them? We argue that they can. We first distinguish referential grounding -- the connection between a representation and its worldly referent -- from other forms of grounding and argue it is the only kind essential to solving the problem. We contend that referential grounding is achieved when a system's internal states satisfy two conditions derived from teleosemantic theories of representation: (1) they stand in appropriate causal-informational relations to the world, and (2) they have a history of selection that has endowed them with the function of carrying this information. We argue that LLMs can meet both conditions, even without multimodality or embodiment.

Maijunxian Wang, Ruisi Wang, Juyi Lin et al.

Rapid progress in video models has largely focused on visual quality, leaving their reasoning capabilities underexplored. Video reasoning grounds intelligence in spatiotemporally consistent visual environments that go beyond what text can naturally capture, enabling intuitive reasoning over spatiotemporal structure such as continuity, interaction, and causality. However, systematically studying video reasoning and its scaling behavior is hindered by the lack of large-scale training data. To address this gap, we introduce the Very Big Video Reasoning (VBVR) Dataset, an unprecedentedly large-scale resource spanning 200 curated reasoning tasks following a principled taxonomy and over one million video clips, approximately three orders of magnitude larger than existing datasets. We further present VBVR-Bench, a verifiable evaluation framework that moves beyond model-based judging by incorporating rule-based, human-aligned scorers, enabling reproducible and interpretable diagnosis of video reasoning capabilities. Leveraging the VBVR suite, we conduct one of the first large-scale scaling studies of video reasoning and observe early signs of emergent generalization to unseen reasoning tasks. Together, VBVR lays a foundation for the next stage of research in generalizable video reasoning. The data, benchmark toolkit, and models are publicly available at https://video-reason.com/ .

Raphaël Millière

Text-guided image generation models can be prompted to generate images using nonce words adversarially designed to robustly evoke specific visual concepts. Two approaches for such generation are introduced: macaronic prompting, which involves designing cryptic hybrid words by concatenating subword units from different languages; and evocative prompting, which involves designing nonce words whose broad morphological features are similar enough to that of existing words to trigger robust visual associations. The two methods can also be combined to generate images associated with more specific visual concepts. The implications of these techniques for the circumvention of existing approaches to content moderation, and particularly the generation of offensive or harmful images, are discussed.

Raphaël Millière

This chapter critically examines the potential contributions of modern language models to theoretical linguistics. Despite their focus on engineering goals, these models' ability to acquire sophisticated linguistic knowledge from mere exposure to data warrants a careful reassessment of their relevance to linguistic theory. I review a growing body of empirical evidence suggesting that language models can learn hierarchical syntactic structure and exhibit sensitivity to various linguistic phenomena, even when trained on developmentally plausible amounts of data. While the competence/performance distinction has been invoked to dismiss the relevance of such models to linguistic theory, I argue that this assessment may be premature. By carefully controlling learning conditions and making use of causal intervention methods, experiments with language models can potentially constrain hypotheses about language acquisition and competence. I conclude that closer collaboration between theoretical linguists and computational researchers could yield valuable insights, particularly in advancing debates about linguistic nativism.

Safoora Yousefi, Leo Betthauser, Hosein Hasanbeig et al.

Large language models (LLMs) exhibit remarkable performance improvement through in-context learning (ICL) by leveraging task-specific examples in the input. However, the mechanisms behind this improvement remain elusive. In this work, we investigate how LLM embeddings and attention representations change following in-context-learning, and how these changes mediate improvement in behavior. We employ neuroscience-inspired techniques such as representational similarity analysis (RSA) and propose novel methods for parameterized probing and measuring ratio of attention to relevant vs. irrelevant information in Llama-2 70B and Vicuna 13B. We designed two tasks with a priori relationships among their conditions: linear regression and reading comprehension. We formed hypotheses about expected similarities in task representations and measured hypothesis alignment of LLM representations before and after ICL as well as changes in attention. Our analyses revealed a meaningful correlation between improvements in behavior after ICL and changes in both embeddings and attention weights across LLM layers. This empirical framework empowers a nuanced understanding of how latent representations shape LLM behavior, offering valuable tools and insights for future research and practical applications.

Raphaël Millière

A core challenge in the development of increasingly capable AI systems is to make them safe and reliable by ensuring their behaviour is consistent with human values. This challenge, known as the alignment problem, does not merely apply to hypothetical future AI systems that may pose catastrophic risks; it already applies to current systems, such as large language models, whose potential for harm is rapidly increasing. In this paper, I assess whether we are on track to solve the alignment problem for large language models, and what that means for the safety of future AI systems. I argue that existing strategies for alignment are insufficient, because large language models remain vulnerable to adversarial attacks that can reliably elicit unsafe behaviour. I offer an explanation of this lingering vulnerability on which it is not simply a contingent limitation of current language models, but has deep technical ties to a crucial aspect of what makes these models useful and versatile in the first place -- namely, their remarkable aptitude to learn "in context" directly from user instructions. It follows that the alignment problem is not only unsolved for current AI systems, but may be intrinsically difficult to solve without severely undermining their capabilities. Furthermore, this assessment raises concerns about the prospect of ensuring the safety of future and more capable AI systems.

Raphaël Millière

Deep learning algorithms are rapidly changing the way in which audiovisual media can be produced. Synthetic audiovisual media generated with deep learning - often subsumed colloquially under the label "deepfakes" - have a number of impressive characteristics; they are increasingly trivial to produce, and can be indistinguishable from real sounds and images recorded with a sensor. Much attention has been dedicated to ethical concerns raised by this technological development. Here, I focus instead on a set of issues related to the notion of synthetic audiovisual media, its place within a broader taxonomy of audiovisual media, and how deep learning techniques differ from more traditional approaches to media synthesis. After reviewing important etiological features of deep learning pipelines for media manipulation and generation, I argue that "deepfakes" and related synthetic media produced with such pipelines do not merely offer incremental improvements over previous methods, but challenge traditional taxonomical distinctions, and pave the way for genuinely novel kinds of audiovisual media.

Raphaël Millière, Charles Rathkopf

Evaluating the cognitive capacities of large language models (LLMs) requires overcoming not only anthropomorphic but also anthropocentric biases. This article identifies two types of anthropocentric bias that have been neglected: overlooking how auxiliary factors can impede LLM performance despite competence ("auxiliary oversight"), and dismissing LLM mechanistic strategies that differ from those of humans as not genuinely competent ("mechanistic chauvinism"). Mitigating these biases necessitates an empirically-driven, iterative approach to mapping cognitive tasks to LLM-specific capacities and mechanisms, which can be done by supplementing carefully designed behavioral experiments with mechanistic studies.

Raphaël Millière, Cameron Buckner

Large language models like GPT-4 have achieved remarkable proficiency in a broad spectrum of language-based tasks, some of which are traditionally associated with hallmarks of human intelligence. This has prompted ongoing disagreements about the extent to which we can meaningfully ascribe any kind of linguistic or cognitive competence to language models. Such questions have deep philosophical roots, echoing longstanding debates about the status of artificial neural networks as cognitive models. This article -- the first part of two companion papers -- serves both as a primer on language models for philosophers, and as an opinionated survey of their significance in relation to classic debates in the philosophy cognitive science, artificial intelligence, and linguistics. We cover topics such as compositionality, language acquisition, semantic competence, grounding, world models, and the transmission of cultural knowledge. We argue that the success of language models challenges several long-held assumptions about artificial neural networks. However, we also highlight the need for further empirical investigation to better understand their internal mechanisms. This sets the stage for the companion paper (Part II), which turns to novel empirical methods for probing the inner workings of language models, and new philosophical questions prompted by their latest developments.

Yiwei Wu, Atticus Geiger, Raphaël Millière

Variable binding -- the ability to associate variables with values -- is fundamental to symbolic computation and cognition. Although classical architectures typically implement variable binding via addressable memory, it is not well understood how modern neural networks lacking built-in binding operations may acquire this capacity. We investigate this by training a Transformer to dereference queried variables in symbolic programs where variables are assigned either numerical constants or other variables. Each program requires following chains of variable assignments up to four steps deep to find the queried value, and also contains irrelevant chains of assignments acting as distractors. Our analysis reveals a developmental trajectory with three distinct phases during training: (1) random prediction of numerical constants, (2) a shallow heuristic prioritizing early variable assignments, and (3) the emergence of a systematic mechanism for dereferencing assignment chains. Using causal interventions, we find that the model learns to exploit the residual stream as an addressable memory space, with specialized attention heads routing information across token positions. This mechanism allows the model to dynamically track variable bindings across layers, resulting in accurate dereferencing. Our results show how Transformer models can learn to implement systematic variable binding without explicit architectural support, bridging connectionist and symbolic approaches. To facilitate reproducible research, we developed Variable Scope, an interactive web platform for exploring our findings at https://variablescope.org

Raphaël Millière

Deep learning has enabled major advances across most areas of artificial intelligence research. This remarkable progress extends beyond mere engineering achievements and holds significant relevance for the philosophy of cognitive science. Deep neural networks have made significant strides in overcoming the limitations of older connectionist models that once occupied the centre stage of philosophical debates about cognition. This development is directly relevant to long-standing theoretical debates in the philosophy of cognitive science. Furthermore, ongoing methodological challenges related to the comparative evaluation of deep neural networks stand to benefit greatly from interdisciplinary collaboration with philosophy and cognitive science. The time is ripe for philosophers to explore foundational issues related to deep learning and cognition; this perspective paper surveys key areas where their contributions can be especially fruitful.

Raphaël Millière

The progress of AI systems such as large language models (LLMs) raises increasingly pressing concerns about their safe deployment. This paper examines the value alignment problem for LLMs, arguing that current alignment strategies are fundamentally inadequate to prevent misuse. Despite ongoing efforts to instill norms such as helpfulness, honesty, and harmlessness in LLMs through fine-tuning based on human preferences, they remain vulnerable to adversarial attacks that exploit conflicts between these norms. I argue that this vulnerability reflects a fundamental limitation of existing alignment methods: they reinforce shallow behavioral dispositions rather than endowing LLMs with a genuine capacity for normative deliberation. Drawing from on research in moral psychology, I show how humans' ability to engage in deliberative reasoning enhances their resilience against similar adversarial tactics. LLMs, by contrast, lack a robust capacity to detect and rationally resolve normative conflicts, leaving them susceptible to manipulation; even recent advances in reasoning-focused LLMs have not addressed this vulnerability. This ``shallow alignment'' problem carries significant implications for AI safety and regulation, suggesting that current approaches are insufficient for mitigating potential harms posed by increasingly capable AI systems.

Sam Musker, Alex Duchnowski, Raphaël Millière et al.

Analogical reasoning -- the capacity to identify and map structural relationships between different domains -- is fundamental to human cognition and learning. Recent studies have shown that large language models (LLMs) can sometimes match humans in analogical reasoning tasks, opening the possibility that analogical reasoning might emerge from domain-general processes. However, it is still debated whether these emergent capacities are largely superficial and limited to simple relations seen during training or whether they encompass the flexible representational and mapping capabilities which are the focus of leading cognitive models of analogy. In this study, we introduce novel analogical reasoning tasks that require participants to map between semantically contentful words and sequences of letters and other abstract characters. This task necessitates the ability to flexibly re-represent rich semantic information -- an ability which is known to be central to human analogy but which is thus far not well captured by existing cognitive theories and models. We assess the performance of both human participants and LLMs on tasks focusing on reasoning from semantic structure and semantic content, introducing variations that test the robustness of their analogical inferences. Advanced LLMs match human performance across several conditions, though humans and LLMs respond differently to certain task variations and semantic distractors. Our results thus provide new evidence that LLMs might offer a how-possibly explanation of human analogical reasoning in contexts that are not yet well modeled by existing theories, but that even today's best models are unlikely to yield how-actually explanations.

Raphaël Millière, Cameron Buckner

In this paper, the second of two companion pieces, we explore novel philosophical questions raised by recent progress in large language models (LLMs) that go beyond the classical debates covered in the first part. We focus particularly on issues related to interpretability, examining evidence from causal intervention methods about the nature of LLMs' internal representations and computations. We also discuss the implications of multimodal and modular extensions of LLMs, recent debates about whether such systems may meet minimal criteria for consciousness, and concerns about secrecy and reproducibility in LLM research. Finally, we discuss whether LLM-like systems may be relevant to modeling aspects of human cognition, if their architectural characteristics and learning scenario are adequately constrained.