Tomer Ullman

Colin Conwell, Tomer Ullman · mit

Relations are basic building blocks of human cognition. Classic and recent work suggests that many relations are early developing, and quickly perceived. Machine models that aspire to human-level perception and reasoning should reflect the ability to recognize and reason generatively about relations. We report a systematic empirical examination of a recent text-guided image generation model (DALL-E 2), using a set of 15 basic physical and social relations studied or proposed in the literature, and judgements from human participants (N = 169). Overall, we find that only ~22% of images matched basic relation prompts. Based on a quantitative examination of people's judgments, we suggest that current image generation models do not yet have a grasp of even basic relations involving simple objects and agents. We examine reasons for model successes and failures, and suggest possible improvements based on computations observed in biological intelligence.

Tomer Ullman

Intuitive psychology is a pillar of common-sense reasoning. The replication of this reasoning in machine intelligence is an important stepping-stone on the way to human-like artificial intelligence. Several recent tasks and benchmarks for examining this reasoning in Large-Large Models have focused in particular on belief attribution in Theory-of-Mind tasks. These tasks have shown both successes and failures. We consider in particular a recent purported success case, and show that small variations that maintain the principles of ToM turn the results on their head. We argue that in general, the zero-hypothesis for model evaluation in intuitive psychology should be skeptical, and that outlying failure cases should outweigh average success rates. We also consider what possible future successes on Theory-of-Mind tasks by more powerful LLMs would mean for ToM tasks with people.

Natalie Shapira, Chris Wendler, Avery Yen et al.

We report an exploratory red-teaming study of autonomous language-model-powered agents deployed in a live laboratory environment with persistent memory, email accounts, Discord access, file systems, and shell execution. Over a two-week period, twenty AI researchers interacted with the agents under benign and adversarial conditions. Focusing on failures emerging from the integration of language models with autonomy, tool use, and multi-party communication, we document eleven representative case studies. Observed behaviors include unauthorized compliance with non-owners, disclosure of sensitive information, execution of destructive system-level actions, denial-of-service conditions, uncontrolled resource consumption, identity spoofing vulnerabilities, cross-agent propagation of unsafe practices, and partial system takeover. In several cases, agents reported task completion while the underlying system state contradicted those reports. We also report on some of the failed attempts. Our findings establish the existence of security-, privacy-, and governance-relevant vulnerabilities in realistic deployment settings. These behaviors raise unresolved questions regarding accountability, delegated authority, and responsibility for downstream harms, and warrant urgent attention from legal scholars, policymakers, and researchers across disciplines. This report serves as an initial empirical contribution to that broader conversation.

Felix A. Sosa, Tomer Ullman

Humans can generate reasonable answers to novel queries (Schulz, 2012): if I asked you what kind of food you want to eat for lunch, you would respond with a food, not a time. The thought that one would respond "After 4pm" to "What would you like to eat" is either a joke or a mistake, and seriously entertaining it as a lunch option would likely never happen in the first place. While understanding how people come up with new ideas, thoughts, explanations, and hypotheses that obey the basic constraints of a novel search space is of central importance to cognitive science, there is no agreed-on formal model for this kind of reasoning. We propose that a core component of any such reasoning system is a type theory: a formal imposition of structure on the kinds of computations an agent can perform, and how they're performed. We motivate this proposal with three empirical observations: adaptive constraints on learning and inference (i.e. generating reasonable hypotheses), how people draw distinctions between improbability and impossibility, and people's ability to reason about things at varying levels of abstraction.

Eric Bigelow, Daniel Wurgaft, YingQiao Wang et al.

Large language models (LLMs) can be controlled at inference time through prompts (in-context learning) and internal activations (activation steering). Different accounts have been proposed to explain these methods, yet their common goal of controlling model behavior raises the question of whether these seemingly disparate methodologies can be seen as specific instances of a broader framework. Motivated by this, we develop a unifying, predictive account of LLM control from a Bayesian perspective. Specifically, we posit that both context- and activation-based interventions impact model behavior by altering its belief in latent concepts: steering operates by changing concept priors, while in-context learning leads to an accumulation of evidence. This results in a closed-form Bayesian model that is highly predictive of LLM behavior across context- and activation-based interventions in a set of domains inspired by prior work on many-shot in-context learning. This model helps us explain prior empirical phenomena - e.g., sigmoidal learning curves as in-context evidence accumulates - while predicting novel ones - e.g., additivity of both interventions in log-belief space, which results in distinct phases such that sudden and dramatic behavioral shifts can be induced by slightly changing intervention controls. Taken together, this work offers a unified account of prompt-based and activation-based control of LLM behavior, and a methodology for empirically predicting the effects of these interventions.

Colin Conwell, Rupert Tawiah-Quashie, Tomer Ullman

Despite remarkable progress in multi-modal AI research, there is a salient domain in which modern AI continues to lag considerably behind even human children: the reliable deployment of logical operators. Here, we examine three forms of logical operators: relations, negations, and discrete numbers. We asked human respondents (N=178 in total) to evaluate images generated by a state-of-the-art image-generating AI (DALL-E 3) prompted with these `logical probes', and find that none reliably produce human agreement scores greater than 50\%. The negation probes and numbers (beyond 3) fail most frequently. In a 4th experiment, we assess a `grounded diffusion' pipeline that leverages targeted prompt engineering and structured intermediate representations for greater compositional control, but find its performance is judged even worse than that of DALL-E 3 across prompts. To provide further clarity on potential sources of success and failure in these text-to-image systems, we supplement our 4 core experiments with multiple auxiliary analyses and schematic diagrams, directly quantifying, for example, the relationship between the N-gram frequency of relational prompts and the average match to generated images; the success rates for 3 different prompt modification strategies in the rendering of negation prompts; and the scalar variability / ratio dependence (`approximate numeracy') of prompts involving integers. We conclude by discussing the limitations inherent to `grounded' multimodal learning systems whose grounding relies heavily on vector-based semantics (e.g. DALL-E 3), or under-specified syntactical constraints (e.g. `grounded diffusion'), and propose minimal modifications (inspired by development, based in imagery) that could help to bridge the lingering compositional gap between scale and structure. All data and code is available at https://github.com/ColinConwell/T2I-Probology

Sonia K. Murthy, Rosie Zhao, Jennifer Hu et al. · deepmind

Value trade-offs are an integral part of human decision-making and language use, however, current tools for interpreting such dynamic and multi-faceted notions of values in LLMs are limited. In cognitive science, so-called "cognitive models" provide formal accounts of such trade-offs in humans, by modeling the weighting of a speaker's competing utility functions in choosing an action or utterance. Here we use a leading cognitive model of polite speech to systematically evaluate value trade-offs in two encompassing model settings: degrees of reasoning "effort" in frontier black-box models, and RL post-training dynamics of open-source models. Our results highlight patterns of higher informational utility than social utility in reasoning models' default behavior, and demonstrate that these patterns shift in predictable ways when models are prompted to prioritize certain goals over others. Our findings from LLMs' training dynamics suggest large shifts in utility values early on in training with persistent effects of the choice of base model and pretraining data, compared to feedback dataset or alignment method. Our framework offers a flexible tool for probing value trade-offs across diverse model types, providing insights for generating hypotheses about other social behaviors such as sycophancy and for shaping training regimes that better control trade-offs between values during model development.

YingQiao Wang, Eric Bigelow, Boyi Li et al.

We propose a novel cognitively-inspired method to improve and interpret physical simulation in vision-language models. Our ``Chain of Time" method involves generating a series of intermediate images during a simulation, and it is motivated by in-context reasoning in machine learning, as well as mental simulation in humans. Chain of Time is used at inference time, and requires no additional fine-tuning. We apply the Chain-of-Time method to synthetic and real-world domains, including 2-D graphics simulations and natural 3-D videos. These domains test a variety of particular physical properties, including velocity, acceleration, fluid dynamics, and conservation of momentum. We found that using Chain-of-Time simulation substantially improves the performance of a state-of-the-art image generation model. Beyond examining performance, we also analyzed the specific states of the world simulated by an image model at each time step, which sheds light on the dynamics underlying these simulations. This analysis reveals insights that are hidden from traditional evaluations of physical reasoning, including cases where an image generation model is able to simulate physical properties that unfold over time, such as velocity, gravity, and collisions. Our analysis also highlights particular cases where the image generation model struggles to infer particular physical parameters from input images, despite being capable of simulating relevant physical processes.

Sonia K. Murthy, Tomer Ullman, Jennifer Hu

Researchers in social science and psychology have recently proposed using large language models (LLMs) as replacements for humans in behavioral research. In addition to arguments about whether LLMs accurately capture population-level patterns, this has raised questions about whether LLMs capture human-like conceptual diversity. Separately, it is debated whether post-training alignment (RLHF or RLAIF) affects models' internal diversity. Inspired by human studies, we use a new way of measuring the conceptual diversity of synthetically-generated LLM "populations" by relating the internal variability of simulated individuals to the population-level variability. We use this approach to evaluate non-aligned and aligned LLMs on two domains with rich human behavioral data. While no model reaches human-like diversity, aligned models generally display less diversity than their instruction fine-tuned counterparts. Our findings highlight potential trade-offs between increasing models' value alignment and decreasing the diversity of their conceptual representations.

Eric Bigelow, Ari Holtzman, Hidenori Tanaka et al.

Estimating uncertainty in Large Language Models (LLMs) is important for properly evaluating LLMs, and ensuring safety for users. However, prior approaches to uncertainty estimation focus on the final answer in generated text, ignoring intermediate steps that might dramatically impact the outcome. We hypothesize that there exist key forking tokens, such that re-sampling the system at those specific tokens, but not others, leads to very different outcomes. To test this empirically, we develop a novel approach to representing uncertainty dynamics across individual tokens of text generation, and applying statistical models to test our hypothesis. Our approach is highly flexible: it can be applied to any dataset and any LLM, without fine tuning or accessing model weights. We use our method to analyze LLM responses on 7 different tasks across 4 domains, spanning a wide range of typical use cases. We find many examples of forking tokens, including surprising ones such as punctuation marks, suggesting that LLMs are often just a single token away from saying something very different.

Jennifer Hu, Felix Sosa, Tomer Ullman

The question of whether large language models (LLMs) possess Theory of Mind (ToM) -- often defined as the ability to reason about others' mental states -- has sparked significant scientific and public interest. However, the evidence as to whether LLMs possess ToM is mixed, and the recent growth in evaluations has not resulted in a convergence. Here, we take inspiration from cognitive science to re-evaluate the state of ToM evaluation in LLMs. We argue that a major reason for the disagreement on whether LLMs have ToM is a lack of clarity on whether models should be expected to match human behaviors, or the computations underlying those behaviors. We also highlight ways in which current evaluations may be deviating from "pure" measurements of ToM abilities, which also contributes to the confusion. We conclude by discussing several directions for future research, including the relationship between ToM and pragmatic communication, which could advance our understanding of artificial systems as well as human cognition.

Tomer Ullman

Illusions are entertaining, but they are also a useful diagnostic tool in cognitive science, philosophy, and neuroscience. A typical illusion shows a gap between how something "really is" and how something "appears to be", and this gap helps us understand the mental processing that lead to how something appears to be. Illusions are also useful for investigating artificial systems, and much research has examined whether computational models of perceptions fall prey to the same illusions as people. Here, I invert the standard use of perceptual illusions to examine basic processing errors in current vision language models. I present these models with illusory-illusions, neighbors of common illusions that should not elicit processing errors. These include such things as perfectly reasonable ducks, crooked lines that truly are crooked, circles that seem to have different sizes because they are, in fact, of different sizes, and so on. I show that many current vision language systems mistakenly see these illusion-illusions as illusions. I suggest that such failures are part of broader failures already discussed in the literature.

Jennifer Hu, Felix Sosa, Tomer Ullman

Some things are impossible, but some things may be even more impossible than impossible. Levitating a feather using one's mind is impossible in our world, but fits into our intuitive theories of possible worlds, whereas levitating a feather using the number five cannot be conceived in any possible world ("inconceivable"). While prior work has examined the distinction between improbable and impossible events, there has been little empirical research on inconceivability. Here, we investigate whether people maintain a distinction between impossibility and inconceivability, and how such distinctions might be made. We find that people can readily distinguish the impossible from the inconceivable, using categorization studies similar to those used to investigate the differences between impossible and improbable (Experiment 1). However, this distinction is not explained by people's subjective ratings of event likelihood, which are near zero and indistinguishable between impossible and inconceivable event descriptions (Experiment 2). Finally, we ask whether the probabilities assigned to event descriptions by statistical language models (LMs) can be used to separate modal categories, and whether these probabilities align with people's ratings (Experiment 3). We find high-level similarities between people and LMs: both distinguish among impossible and inconceivable event descriptions, and LM-derived string probabilities predict people's ratings of event likelihood across modal categories. Our findings suggest that fine-grained knowledge about exceedingly rare events (i.e., the impossible and inconceivable) may be learned via statistical learning over linguistic forms, yet leave open the question of whether people represent the distinction between impossible and inconceivable as a difference not of degree, but of kind.

Bo Zhao, Maya Okawa, Eric J. Bigelow et al.

As large language models (LLMs) increasingly power conversational agents, understanding how they model users' emotional states is critical for ethical deployment. Inspired by emotion wheels -- a psychological framework that argues emotions organize hierarchically -- we analyze probabilistic dependencies between emotional states in model outputs. We find that LLMs naturally form hierarchical emotion trees that align with human psychological models, and larger models develop more complex hierarchies. We also uncover systematic biases in emotion recognition across socioeconomic personas, with compounding misclassifications for intersectional, underrepresented groups. Human studies reveal striking parallels, suggesting that LLMs internalize aspects of social perception. Beyond highlighting emergent emotional reasoning in LLMs, our results hint at the potential of using cognitively-grounded theories for developing better model evaluations.

Chuanyang Jin, Yutong Wu, Jing Cao et al.

Theory of Mind (ToM), the ability to understand people's mental states, is an essential ingredient for developing machines with human-level social intelligence. Recent machine learning models, particularly large language models, seem to show some aspects of ToM understanding. However, existing ToM benchmarks use unimodal datasets - either video or text. Human ToM, on the other hand, is more than video or text understanding. People can flexibly reason about another person's mind based on conceptual representations (e.g., goals, beliefs, plans) extracted from any available data. To address this, we introduce a multimodal Theory of Mind question answering (MMToM-QA) benchmark. MMToM-QA comprehensively evaluates machine ToM both on multimodal data and on different kinds of unimodal data about a person's activity in a household environment. To engineer multimodal ToM capacity, we propose a novel method, BIP-ALM (Bayesian Inverse Planning Accelerated by Language Models). BIP-ALM extracts unified representations from multimodal data and utilizes language models for scalable Bayesian inverse planning. We conducted a systematic comparison of human performance, BIP-ALM, and state-of-the-art models, including GPT-4. The experiments demonstrate that large language models and large multimodal models still lack robust ToM capacity. BIP-ALM, on the other hand, shows promising results, by leveraging the power of both model-based mental inference and language models.

Michael B. Chang, Tomer Ullman, Antonio Torralba et al.

We present the Neural Physics Engine (NPE), a framework for learning simulators of intuitive physics that naturally generalize across variable object count and different scene configurations. We propose a factorization of a physical scene into composable object-based representations and a neural network architecture whose compositional structure factorizes object dynamics into pairwise interactions. Like a symbolic physics engine, the NPE is endowed with generic notions of objects and their interactions; realized as a neural network, it can be trained via stochastic gradient descent to adapt to specific object properties and dynamics of different worlds. We evaluate the efficacy of our approach on simple rigid body dynamics in two-dimensional worlds. By comparing to less structured architectures, we show that the NPE's compositional representation of the structure in physical interactions improves its ability to predict movement, generalize across variable object count and different scene configurations, and infer latent properties of objects such as mass.