Taylor Webb

Taylor Webb, Keith J. Holyoak, Hongjing Lu

The recent advent of large language models has reinvigorated debate over whether human cognitive capacities might emerge in such generic models given sufficient training data. Of particular interest is the ability of these models to reason about novel problems zero-shot, without any direct training. In human cognition, this capacity is closely tied to an ability to reason by analogy. Here, we performed a direct comparison between human reasoners and a large language model (the text-davinci-003 variant of GPT-3) on a range of analogical tasks, including a non-visual matrix reasoning task based on the rule structure of Raven's Standard Progressive Matrices. We found that GPT-3 displayed a surprisingly strong capacity for abstract pattern induction, matching or even surpassing human capabilities in most settings; preliminary tests of GPT-4 indicated even better performance. Our results indicate that large language models such as GPT-3 have acquired an emergent ability to find zero-shot solutions to a broad range of analogy problems.

Shanka Subhra Mondal, Taylor Webb, Jonathan D. Cohen

A core component of human intelligence is the ability to identify abstract patterns inherent in complex, high-dimensional perceptual data, as exemplified by visual reasoning tasks such as Raven's Progressive Matrices (RPM). Motivated by the goal of designing AI systems with this capacity, recent work has focused on evaluating whether neural networks can learn to solve RPM-like problems. Previous work has generally found that strong performance on these problems requires the incorporation of inductive biases that are specific to the RPM problem format, raising the question of whether such models might be more broadly useful. Here, we investigated the extent to which a general-purpose mechanism for processing visual scenes in terms of objects might help promote abstract visual reasoning. We found that a simple model, consisting only of an object-centric encoder and a transformer reasoning module, achieved state-of-the-art results on both of two challenging RPM-like benchmarks (PGM and I-RAVEN), as well as a novel benchmark with greater visual complexity (CLEVR-Matrices). These results suggest that an inductive bias for object-centric processing may be a key component of abstract visual reasoning, obviating the need for problem-specific inductive biases.

Awni Altabaa, Taylor Webb, Jonathan Cohen et al.

An extension of Transformers is proposed that enables explicit relational reasoning through a novel module called the Abstractor. At the core of the Abstractor is a variant of attention called relational cross-attention. The approach is motivated by an architectural inductive bias for relational learning that disentangles relational information from object-level features. This enables explicit relational reasoning, supporting abstraction and generalization from limited data. The Abstractor is first evaluated on simple discriminative relational tasks and compared to existing relational architectures. Next, the Abstractor is evaluated on purely relational sequence-to-sequence tasks, where dramatic improvements are seen in sample efficiency compared to standard Transformers. Finally, Abstractors are evaluated on a collection of tasks based on mathematical problem solving, where consistent improvements in performance and sample efficiency are observed.

Taylor Webb, Shanka Subhra Mondal, Ida Momennejad

Large language models (LLMs) demonstrate impressive performance on a wide variety of tasks, but they often struggle with tasks that require multi-step reasoning or goal-directed planning. Both cognitive neuroscience and reinforcement learning (RL) have proposed a number of interacting functional components that together implement search and evaluation in multi-step decision making. These components include conflict monitoring, state prediction, state evaluation, task decomposition, and orchestration. To improve planning with LLMs, we propose an agentic architecture, the Modular Agentic Planner (MAP), in which planning is accomplished via the recurrent interaction of the specialized modules mentioned above, each implemented using an LLM. MAP improves planning through the interaction of specialized modules that break down a larger problem into multiple brief automated calls to the LLM. We evaluate MAP on three challenging planning tasks -- graph traversal, Tower of Hanoi, and the PlanBench benchmark -- as well as an NLP task requiring multi-step reasoning (strategyQA). We find that MAP yields significant improvements over both standard LLM methods (zero-shot prompting, in-context learning) and competitive baselines (chain-of-thought, multi-agent debate, and tree-of-thought), can be effectively combined with smaller and more cost-efficient LLMs (Llama3-70B), and displays superior transfer across tasks. These results suggest the benefit of a modular and multi-agent approach to planning with LLMs.

Yukang Yang, Declan Campbell, Kaixuan Huang et al.

Many recent studies have found evidence for emergent reasoning capabilities in large language models (LLMs), but debate persists concerning the robustness of these capabilities, and the extent to which they depend on structured reasoning mechanisms. To shed light on these issues, we study the internal mechanisms that support abstract reasoning in LLMs. We identify an emergent symbolic architecture that implements abstract reasoning via a series of three computations. In early layers, symbol abstraction heads convert input tokens to abstract variables based on the relations between those tokens. In intermediate layers, symbolic induction heads perform sequence induction over these abstract variables. Finally, in later layers, retrieval heads predict the next token by retrieving the value associated with the predicted abstract variable. These results point toward a resolution of the longstanding debate between symbolic and neural network approaches, suggesting that emergent reasoning in neural networks depends on the emergence of symbolic mechanisms.

Taylor Webb, Keith J. Holyoak, Hongjing Lu

We recently reported evidence that large language models are capable of solving a wide range of text-based analogy problems in a zero-shot manner, indicating the presence of an emergent capacity for analogical reasoning. Two recent commentaries have challenged these results, citing evidence from so-called `counterfactual' tasks in which the standard sequence of the alphabet is arbitrarily permuted so as to decrease similarity with materials that may have been present in the language model's training data. Here, we reply to these critiques, clarifying some misunderstandings about the test materials used in our original work, and presenting evidence that language models are also capable of generalizing to these new counterfactual task variants.

Oliver Eberle, Thomas McGee, Hamza Giaffar et al.

What algorithms do LLMs actually learn and use to solve problems? Studies addressing this question are sparse, as research priorities are focused on improving performance through scale, leaving a theoretical and empirical gap in understanding emergent algorithms. This position paper proposes AlgEval: a framework for systematic research into the algorithms that LLMs learn and use. AlgEval aims to uncover algorithmic primitives, reflected in latent representations, attention, and inference-time compute, and their algorithmic composition to solve task-specific problems. We highlight potential methodological paths and a case study toward this goal, focusing on emergent search algorithms. Our case study illustrates both the formation of top-down hypotheses about candidate algorithms, and bottom-up tests of these hypotheses via circuit-level analysis of attention patterns and hidden states. The rigorous, systematic evaluation of how LLMs actually solve tasks provides an alternative to resource-intensive scaling, reorienting the field toward a principled understanding of underlying computations. Such algorithmic explanations offer a pathway to human-understandable interpretability, enabling comprehension of the model's internal reasoning performance measures. This can in turn lead to more sample-efficient methods for training and improving performance, as well as novel architectures for end-to-end and multi-agent systems.

Rim Assouel, Declan Campbell, Taylor Webb

To accurately process a visual scene, observers must bind features together to represent individual objects. This capacity is necessary, for instance, to distinguish an image containing a red square and a blue circle from an image containing a blue square and a red circle. Recent work has found that language models solve this 'binding problem' via a set of symbol-like, content-independent indices, but it is unclear whether similar mechanisms are employed by vision language models (VLMs). This question is especially relevant, given the persistent failures of VLMs on tasks that require binding. Here, we identify a set of emergent symbolic mechanisms that support binding in VLMs via a content-independent, spatial indexing scheme. Moreover, we find that binding errors can be traced directly to failures in these mechanisms. Taken together, these results shed light on the mechanisms that support symbol-like processing in VLMs, and suggest possible avenues for addressing the persistent binding failures exhibited by these models.

Konstantin Demin, Taylor Webb, Eric Elmoznino et al.

We spell out a definition of sentience that may be useful for designing and building it in machines. We propose that for sentience to be meaningful for AI, it must be fleshed out in functional, computational terms, in enough detail to allow for implementation. Yet, this notion of sentience must also reflect something essentially 'subjective', beyond just having the general capacity to encode perceptual content. For this specific functional notion of sentience to occur, we propose that certain sensory signals need to be both assertoric (persistent) and qualitative. To illustrate the definition in more concrete terms, we sketch out some ways for potential implementation, given current technology. Understanding what it takes for artificial agents to be functionally sentient can also help us avoid creating them inadvertently, or at least, realize that we have created them in a timely manner.

Andrew Jun Lee, Taylor Webb, Trevor Bihl et al.

The ability to learn new visual concepts from limited examples is a hallmark of human cognition. While traditional category learning models represent each example as an unstructured feature vector, compositional concept learning is thought to depend on (1) structured representations of examples (e.g., directed graphs consisting of objects and their relations) and (2) the identification of shared relational structure across examples through analogical mapping. Here, we introduce Probabilistic Schema Induction (PSI), a prototype model that employs deep learning to perform analogical mapping over structured representations of only a handful of examples, forming a compositional concept called a schema. In doing so, PSI relies on a novel conception of similarity that weighs object-level similarity and relational similarity, as well as a mechanism for amplifying relations relevant to classification, analogous to selective attention parameters in traditional models. We show that PSI produces human-like learning performance and outperforms two controls: a prototype model that uses unstructured feature vectors extracted from a deep learning model, and a variant of PSI with weaker structured representations. Notably, we find that PSI's human-like performance is driven by an adaptive strategy that increases relational similarity over object-level similarity and upweights the contribution of relations that distinguish classes. These findings suggest that structured representations and analogical mapping are critical to modeling rapid human-like learning of compositional visual concepts, and demonstrate how deep learning can be leveraged to create psychological models.

Shanka Subhra Mondal, Steven Frankland, Taylor Webb et al.

Deep neural networks have made tremendous gains in emulating human-like intelligence, and have been used increasingly as ways of understanding how the brain may solve the complex computational problems on which this relies. However, these still fall short of, and therefore fail to provide insight into how the brain supports strong forms of generalization of which humans are capable. One such case is out-of-distribution (OOD) generalization-successful performance on test examples that lie outside the distribution of the training set. Here, we identify properties of processing in the brain that may contribute to this ability. We describe a two-part algorithm that draws on specific features of neural computation to achieve OOD generalization, and provide a proof of concept by evaluating performance on two challenging cognitive tasks. First we draw on the fact that the mammalian brain represents metric spaces using grid cell code (e.g., in the entorhinal cortex): abstract representations of relational structure, organized in recurring motifs that cover the representational space. Second, we propose an attentional mechanism that operates over the grid cell code using Determinantal Point Process (DPP), that we call DPP attention (DPP-A) -- a transformation that ensures maximum sparseness in the coverage of that space. We show that a loss function that combines standard task-optimized error with DPP-A can exploit the recurring motifs in the grid cell code, and can be integrated with common architectures to achieve strong OOD generalization performance on analogy and arithmetic tasks. This provides both an interpretation of how the grid cell code in the mammalian brain may contribute to generalization performance, and at the same time a potential means for improving such capabilities in artificial neural networks.

Zack Dulberg, Taylor Webb, Jonathan Cohen

Learning to count is an important example of the broader human capacity for systematic generalization, and the development of counting is often characterized by an inflection point when children rapidly acquire proficiency with the procedures that support this ability. We aimed to model this process by training a reinforcement learning agent to select N items from a binary vector when instructed (known as the give-$N$ task). We found that a memory-augmented modular network architecture based on the recently proposed Emergent Symbol Binding Network (ESBN) exhibited an inflection during learning that resembled human development. This model was also capable of systematic extrapolation outside the range of its training set - for example, trained only to select between 1 and 10 items, it could succeed at selecting 11 to 15 items as long as it could make use of an arbitrary count sequence of at least that length. The close parallels to child development and the capacity for extrapolation suggest that our model could shed light on the emergence of systematicity in humans.