Standard training metrics like loss fail to explain the emergence of complex capabilities in large language models. We take a spectral approach to investigate the geometry of learned representations across pretraining and post-training, measuring effective rank (RankMe) and eigenspectrum decay (α-ReQ). With OLMo (1B-7B) and Pythia (160M-12B) models, we uncover a consistent non-monotonic sequence of three geometric phases during autoregressive pretraining. The initial "warmup" phase exhibits rapid representational collapse. This is followed by an "entropy-seeking" phase, where the manifold's dimensionality expands substantially, coinciding with peak n-gram memorization. Subsequently, a "compression-seeking" phase imposes anisotropic consolidation, selectively preserving variance along dominant eigendirections while contracting others, a transition marked with significant improvement in downstream task performance. We show these phases can emerge from a fundamental interplay of cross-entropy optimization under skewed token frequencies and representational bottlenecks (d ≪ |V|). Post-training further transforms geometry: SFT and DPO drive "entropy-seeking" dynamics to integrate specific instructional or preferential data, improving in-distribution performance while degrading out-of-distribution robustness. Conversely, RLVR induces "compression-seeking", enhancing reward alignment but reducing generation diversity.

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.

Transformer-based language models spread FLOPs uniformly across input sequences. In this work we demonstrate that transformers can instead learn to dynamically allocate FLOPs (or compute) to specific positions in a sequence, optimising the allocation along the sequence for different layers across the model depth. Our method enforces a total compute budget by capping the number of tokens ($k$) that can participate in the self-attention and MLP computations at a given layer. The tokens to be processed are determined by the network using a top-$k$ routing mechanism. Since $k$ is defined a priori, this simple procedure uses a static computation graph with known tensor sizes, unlike other conditional computation techniques. Nevertheless, since the identities of the $k$ tokens are fluid, this method can expend FLOPs non-uniformly across the time and model depth dimensions. Thus, compute expenditure is entirely predictable in sum total, but dynamic and context-sensitive at the token-level. Not only do models trained in this way learn to dynamically allocate compute, they do so efficiently. These models match baseline performance for equivalent FLOPS and wall-clock times to train, but require a fraction of the FLOPs per forward pass, and can be upwards of 50\% faster to step during post-training sampling.

Creating agents that can interact naturally with humans is a common goal in artificial intelligence (AI) research. However, evaluating these interactions is challenging: collecting online human-agent interactions is slow and expensive, yet faster proxy metrics often do not correlate well with interactive evaluation. In this paper, we assess the merits of these existing evaluation metrics and present a novel approach to evaluation called the Standardised Test Suite (STS). The STS uses behavioural scenarios mined from real human interaction data. Agents see replayed scenario context, receive an instruction, and are then given control to complete the interaction offline. These agent continuations are recorded and sent to human annotators to mark as success or failure, and agents are ranked according to the proportion of continuations in which they succeed. The resulting STS is fast, controlled, interpretable, and representative of naturalistic interactions. Altogether, the STS consolidates much of what is desirable across many of our standard evaluation metrics, allowing us to accelerate research progress towards producing agents that can interact naturally with humans. A video may be found at https://youtu.be/YR1TngGORGQ.

Inferring the abstract relational and causal structure of the world is a major challenge for reinforcement-learning (RL) agents. For humans, language-particularly in the form of explanations-plays a considerable role in overcoming this challenge. Here, we show that language can play a similar role for deep RL agents in complex environments. While agents typically struggle to acquire relational and causal knowledge, augmenting their experience by training them to predict language descriptions and explanations can overcome these limitations. We show that language can help agents learn challenging relational tasks, and examine which aspects of language contribute to its benefits. We then show that explanations can help agents to infer not only relational but also causal structure. Language can shape the way that agents to generalize out-of-distribution from ambiguous, causally-confounded training, and explanations even allow agents to learn to perform experimental interventions to identify causal relationships. Our results suggest that language description and explanation may be powerful tools for improving agent learning and generalization.

It would be useful for machines to use computers as humans do so that they can aid us in everyday tasks. This is a setting in which there is also the potential to leverage large-scale expert demonstrations and human judgements of interactive behaviour, which are two ingredients that have driven much recent success in AI. Here we investigate the setting of computer control using keyboard and mouse, with goals specified via natural language. Instead of focusing on hand-designed curricula and specialized action spaces, we focus on developing a scalable method centered on reinforcement learning combined with behavioural priors informed by actual human-computer interactions. We achieve state-of-the-art and human-level mean performance across all tasks within the MiniWob++ benchmark, a challenging suite of computer control problems, and find strong evidence of cross-task transfer. These results demonstrate the usefulness of a unified human-agent interface when training machines to use computers. Altogether our results suggest a formula for achieving competency beyond MiniWob++ and towards controlling computers, in general, as a human would.

An important goal in artificial intelligence is to create agents that can both interact naturally with humans and learn from their feedback. Here we demonstrate how to use reinforcement learning from human feedback (RLHF) to improve upon simulated, embodied agents trained to a base level of competency with imitation learning. First, we collected data of humans interacting with agents in a simulated 3D world. We then asked annotators to record moments where they believed that agents either progressed toward or regressed from their human-instructed goal. Using this annotation data we leveraged a novel method - which we call "Inter-temporal Bradley-Terry" (IBT) modelling - to build a reward model that captures human judgments. Agents trained to optimise rewards delivered from IBT reward models improved with respect to all of our metrics, including subsequent human judgment during live interactions with agents. Altogether our results demonstrate how one can successfully leverage human judgments to improve agent behaviour, allowing us to use reinforcement learning in complex, embodied domains without programmatic reward functions. Videos of agent behaviour may be found at https://youtu.be/v_Z9F2_eKk4.

Large transformer-based models are able to perform in-context few-shot learning, without being explicitly trained for it. This observation raises the question: what aspects of the training regime lead to this emergent behavior? Here, we show that this behavior is driven by the distributions of the training data itself. In-context learning emerges when the training data exhibits particular distributional properties such as burstiness (items appear in clusters rather than being uniformly distributed over time) and having large numbers of rarely occurring classes. In-context learning also emerges more strongly when item meanings or interpretations are dynamic rather than fixed. These properties are exemplified by natural language, but are also inherent to naturalistic data in a wide range of other domains. They also depart significantly from the uniform, i.i.d. training distributions typically used for standard supervised learning. In our initial experiments, we found that in-context learning traded off against more conventional weight-based learning, and models were unable to achieve both simultaneously. However, our later experiments uncovered that the two modes of learning could co-exist in a single model when it was trained on data following a skewed Zipfian distribution -- another common property of naturalistic data, including language. In further experiments, we found that naturalistic data distributions were only able to elicit in-context learning in transformers, and not in recurrent models. In sum, our findings indicate how the transformer architecture works together with particular properties of the training data to drive the intriguing emergent in-context learning behaviour of large language models, and how future work might encourage both in-context and in-weights learning in domains beyond language.

Human language learners are exposed to a trickle of informative, context-sensitive language, but a flood of raw sensory data. Through both social language use and internal processes of rehearsal and practice, language learners are able to build high-level, semantic representations that explain their perceptions. Here, we take inspiration from such processes of "inner speech" in humans (Vygotsky, 1934) to better understand the role of intra-agent speech in embodied behavior. First, we formally pose intra-agent speech as a semi-supervised problem and develop two algorithms that enable visually grounded captioning with little labeled language data. We then experimentally compute scaling curves over different amounts of labeled data and compare the data efficiency against a supervised learning baseline. Finally, we incorporate intra-agent speech into an embodied, mobile manipulator agent operating in a 3D virtual world, and show that with as few as 150 additional image captions, intra-agent speech endows the agent with the ability to manipulate and answer questions about a new object without any related task-directed experience (zero-shot). Taken together, our experiments suggest that modelling intra-agent speech is effective in enabling embodied agents to learn new tasks efficiently and without direct interaction experience.

New neurons are continuously generated in the subgranular zone of the dentate gyrus throughout adulthood. These new neurons gradually integrate into hippocampal circuits, forming new naive synapses. Viewed from this perspective, these new neurons may represent a significant source of “wiring” noise in hippocampal networks. In machine learning, such noise injection is commonly used as a regularization technique. Regularization techniques help prevent overfitting training data and allow models to generalize learning to new, unseen data. Using a computational modeling approach, here we ask whether a neurogenesis-like process similarly acts as a regularizer, facilitating generalization in a category learning task. In a convolutional neural network (CNN) trained on the CIFAR-10 object recognition dataset, we modeled neurogenesis as a replacement/turnover mechanism, where weights for a randomly chosen small subset of hidden layer neurons were reinitialized to new values as the model learned to categorize 10 different classes of objects. We found that neurogenesis enhanced generalization on unseen test data compared to networks with no neurogenesis. Moreover, neurogenic networks either outperformed or performed similarly to networks with conventional noise injection (i.e., dropout, weight decay, and neural noise). These results suggest that neurogenesis can enhance generalization in hippocampal learning through noise injection, expanding on the roles that neurogenesis may have in cognition.

New neurons are continuously generated in the subgranular zone of the dentate gyrus throughout adulthood. These new neurons gradually integrate into hippocampal circuits, forming new naïve synapses. Viewed from this perspective, these new neurons may represent a significant source of ‘wiring’ noise in hippocampal networks. In machine learning, such noise injection is commonly used as a regularization technique. Regularization techniques help prevent overfitting training data, and allow models to generalize learning to new, unseen data. Using a computational modeling approach, here we ask whether a neurogenesis-like process similarly acts as a regularizer, facilitating generalization in a category learning task. In a convolutional neural network (CNN) trained on the CIFAR-10 object recognition dataset, we modeled neurogenesis as a replacement/turnover mechanism, where weights for a randomly chosen small subset of neurons in a chosen hidden layer were re-initialized to new values as the model learned to categorize 10 different classes of objects. We found that neurogenesis enhanced generalization on unseen test data compared to networks with no neurogenesis. Moreover, neurogenic networks either outperformed or performed similarly to networks with conventional noise injection (i.e., dropout, weight decay, and neural noise). These results suggest that neurogenesis can enhance generalization in hippocampal learning through noise-injection, expanding on the roles that neurogenesis may have in cognition. Author Summary In deep neural networks, various forms of noise injection are used as regularization techniques to prevent overfitting and promote generalization on unseen test data. Here, we were interested in whether adult neurogenesis– the lifelong production of new neurons in the hippocampus– might similarly function as a regularizer in the brain. We explored this question computationally, assessing whether implementing a neurogenesis-like process in a hidden layer within a convolutional neural network trained in a category learning task would prevent overfitting and promote generalization. We found that neurogenesis regularization was as least as effective as, or more effective than, conventional regularizers (i.e., dropout, weight decay and neural noise) in improving model performance. These results suggest that optimal levels of hippocampal neurogenesis may improve memory-guided decision making by preventing overfitting, thereby promoting the formation of more generalized memories that can be applied in a broader range of circumstances. We outline how these predictions may be evaluated behaviorally in rodents with altered hippocampal neurogenesis.

Artificial intelligence (AI) shares many generalizability challenges with psychology. But the fields publish differently. AI publishes fast, through rapid preprint sharing and conference publications. Psychology publishes more slowly, but creates integrative reviews and meta-analyses. We discuss the complementary advantages of each strategy, and suggest that incorporating both types of strategies could lead to more generalizable research in both fields.

Neural networks have achieved success in a wide array of perceptual tasks but often fail at tasks involving both perception and higher-level reasoning. On these more challenging tasks, bespoke approaches (such as modular symbolic components, independent dynamics models or semantic parsers) targeted towards that specific type of task have typically performed better. The downside to these targeted approaches, however, is that they can be more brittle than general-purpose neural networks, requiring significant modification or even redesign according to the particular task at hand. Here, we propose a more general neural-network-based approach to dynamic visual reasoning problems that obtains state-of-the-art performance on three different domains, in each case outperforming bespoke modular approaches tailored specifically to the task. Our method relies on learned object-centric representations, self-attention and self-supervised dynamics learning, and all three elements together are required for strong performance to emerge. The success of this combination suggests that there may be no need to trade off flexibility for performance on problems involving spatio-temporal or causal-style reasoning. With the right soft biases and learning objectives in a neural network we may be able to attain the best of both worlds.

We propose the challenge of rapid task-solving in novel environments (RTS), wherein an agent must solve a series of tasks as rapidly as possible in an unfamiliar environment. An effective RTS agent must balance between exploring the unfamiliar environment and solving its current task, all while building a model of the new environment over which it can plan when faced with later tasks. While modern deep RL agents exhibit some of these abilities in isolation, none are suitable for the full RTS challenge. To enable progress toward RTS, we introduce two challenge domains: (1) a minimal RTS challenge called the Memory&Planning Game and (2) One-Shot StreetLearn Navigation, which introduces scale and complexity from real-world data. We demonstrate that state-of-the-art deep RL agents fail at RTS in both domains, and that this failure is due to an inability to plan over gathered knowledge. We develop Episodic Planning Networks (EPNs) and show that deep-RL agents with EPNs excel at RTS, outperforming the nearest baseline by factors of 2-3 and learning to navigate held-out StreetLearn maps within a single episode. We show that EPNs learn to execute a value iteration-like planning algorithm and that they generalize to situations beyond their training experience. algorithm and that they generalize to situations beyond their training experience.

Since the earliest days of reinforcement learning, the workhorse method for assigning credit to actions over time has been temporal-difference (TD) learning, which propagates credit backward timestep-by-timestep. This approach suffers when delays between actions and rewards are long and when intervening unrelated events contribute variance to long-term returns. We propose state-associative (SA) learning, where the agent learns associations between states and arbitrarily distant future rewards, then propagates credit directly between the two. In this work, we use SA-learning to model the contribution of past states to the current reward. With this model we can predict each state's contribution to the far future, a quantity we call "synthetic returns". TD-learning can then be applied to select actions that maximize these synthetic returns (SRs). We demonstrate the effectiveness of augmenting agents with SRs across a range of tasks on which TD-learning alone fails. We show that the learned SRs are interpretable: they spike for states that occur after critical actions are taken. Finally, we show that our IMPALA-based SR agent solves Atari Skiing -- a game with a lengthy reward delay that posed a major hurdle to deep-RL agents -- 25 times faster than the published state-of-the-art.

A common vision from science fiction is that robots will one day inhabit our physical spaces, sense the world as we do, assist our physical labours, and communicate with us through natural language. Here we study how to design artificial agents that can interact naturally with humans using the simplification of a virtual environment. We show that imitation learning of human-human interactions in a simulated world, in conjunction with self-supervised learning, is sufficient to produce a multimodal interactive agent, which we call MIA, that successfully interacts with non-adversarial humans 75% of the time. We further identify architectural and algorithmic techniques that improve performance, such as hierarchical action selection. Altogether, our results demonstrate that imitation of multi-modal, real-time human behaviour may provide a straightforward and surprisingly effective means of imbuing agents with a rich behavioural prior from which agents might then be fine-tuned for specific purposes, thus laying a foundation for training capable agents for interactive robots or digital assistants. A video of MIA's behaviour may be found at https://youtu.be/ZFgRhviF7mY

The ability to use symbols is the pinnacle of human intelligence, but has yet to be fully replicated in machines. Here we argue that the path towards symbolically fluent artificial intelligence (AI) begins with a reinterpretation of what symbols are, how they come to exist, and how a system behaves when it uses them. We begin by offering an interpretation of symbols as entities whose meaning is established by convention. But crucially, something is a symbol only for those who demonstrably and actively participate in this convention. We then outline how this interpretation thematically unifies the behavioural traits humans exhibit when they use symbols. This motivates our proposal that the field place a greater emphasis on symbolic behaviour rather than particular computational mechanisms inspired by more restrictive interpretations of symbols. Finally, we suggest that AI research explore social and cultural engagement as a tool to develop the cognitive machinery necessary for symbolic behaviour to emerge. This approach will allow for AI to interpret something as symbolic on its own rather than simply manipulate things that are only symbols to human onlookers, and thus will ultimately lead to AI with more human-like symbolic fluency.

Transformer-based language models have proved capable of rudimentary symbolic reasoning, underlining the effectiveness of applying self-attention computations to sets of discrete entities. In this work, we apply this lesson to videos of physical interaction between objects. We show that self-attention-based models operating on discrete, learned, object-centric representations perform well on spatio-temporal reasoning tasks which were expressly designed to trouble traditional neural network models and to require higher-level cognitive processes such as causal reasoning and understanding of intuitive physics and narrative structure. We achieve state of the art results on two datasets, CLEVRER and CATER, significantly outperforming leading hybrid neuro-symbolic models. Moreover, we find that techniques from language modelling, such as BERT-style semi-supervised predictive losses, allow our model to surpass neuro-symbolic approaches while using 40% less labelled data. Our results corroborate the idea that neural networks can reason about the causal, dynamic structure of visual data and attain understanding of intuitive physics, which counters the popular claim that they are only effective at perceptual pattern-recognition and not reasoning per se.

Commentary on "The Generalizability Crisis" (Yarkoni, 2020). Artificial intelligence (AI) shares many generalizability challenges with psychology. But the fields publish differently. AI publishes fast, through rapid preprint sharing and conference publications. Psychology publishes more slowly, but creates integrative reviews and meta-analyses. We discuss the complementary advantages of each strategy, and suggest that incorporating both types of strategies could lead to more generalizable research in both fields.