Stephanie C. Y. Chan

Maty Bohacek, Nino Scherrer, Nicholas Dufour et al. · stanford

The evaluation of large language models relies heavily on standardized benchmarks. These benchmarks provide useful aggregated metrics, but can obscure (i) particular sub-areas where the models are weak ("model gaps") and (ii) imbalanced coverage in the benchmarks themselves ("benchmark gaps"). To automatically uncover both types of gaps, we propose a simple new method using concept activations from sparse autoencoders, to identify fine-grained gaps on a per-concept basis. The method also benefits from grounding evaluation in the model's internal representations, as well as easy comparison across benchmarks. We applied the method to five popular open-source models and more than a dozen benchmarks, as illustrative examples. As validation of the approach, we found that our automatic, unsupervised method was able to recover model gaps that have been previously documented in the literature (e.g. relating to sycophancy), in addition to identifying novel model gaps. We were also able to automatically uncover benchmark gaps: core concepts that should fall within the scope of a given benchmark. Our "competency gaps" method can be used to complement existing benchmarks, by providing a concept-level decomposition of model behavior, and by helping benchmark developers iterate upon benchmark design. Code is available at https://competency-gaps.github.io.

Andrew K. Lampinen, Ishita Dasgupta, Stephanie C. Y. Chan et al. · deepmind, stanford

Language Models (LMs) can perform new tasks by adapting to a few in-context examples. For humans, explanations that connect examples to task principles can improve learning. We therefore investigate whether explanations of few-shot examples can help LMs. We annotate questions from 40 challenging tasks with answer explanations, and various matched control explanations. We evaluate how different types of explanations, instructions, and controls affect zero- and few-shot performance. We analyze these results using statistical multilevel modeling techniques that account for the nested dependencies among conditions, tasks, prompts, and models. We find that explanations can improve performance -- even without tuning. Furthermore, explanations hand-tuned for performance on a small validation set offer substantially larger benefits, and building a prompt by selecting examples and explanations together substantially improves performance over selecting examples alone. Finally, even untuned explanations outperform carefully matched controls, suggesting that the benefits are due to the link between an example and its explanation, rather than lower-level features. However, only large models benefit. In summary, explanations can support the in-context learning of large LMs on challenging tasks.

Ishita Dasgupta, Andrew K. Lampinen, Stephanie C. Y. Chan et al. · deepmind, stanford

Reasoning is a key ability for an intelligent system. Large language models (LMs) achieve above-chance performance on abstract reasoning tasks, but exhibit many imperfections. However, human abstract reasoning is also imperfect. For example, human reasoning is affected by our real-world knowledge and beliefs, and shows notable "content effects"; humans reason more reliably when the semantic content of a problem supports the correct logical inferences. These content-entangled reasoning patterns play a central role in debates about the fundamental nature of human intelligence. Here, we investigate whether language models $\unicode{x2014}$ whose prior expectations capture some aspects of human knowledge $\unicode{x2014}$ similarly mix content into their answers to logical problems. We explored this question across three logical reasoning tasks: natural language inference, judging the logical validity of syllogisms, and the Wason selection task. We evaluate state of the art large language models, as well as humans, and find that the language models reflect many of the same patterns observed in humans across these tasks $\unicode{x2014}$ like humans, models answer more accurately when the semantic content of a task supports the logical inferences. These parallels are reflected both in answer patterns, and in lower-level features like the relationship between model answer distributions and human response times. Our findings have implications for understanding both these cognitive effects in humans, and the factors that contribute to language model performance.

Stephanie C. Y. Chan, Adam Santoro, Andrew K. Lampinen et al. · deepmind, stanford

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.

Allison C. Tam, Neil C. Rabinowitz, Andrew K. Lampinen et al. · deepmind, stanford

Effective exploration is a challenge in reinforcement learning (RL). Novelty-based exploration methods can suffer in high-dimensional state spaces, such as continuous partially-observable 3D environments. We address this challenge by defining novelty using semantically meaningful state abstractions, which can be found in learned representations shaped by natural language. In particular, we evaluate vision-language representations, pretrained on natural image captioning datasets. We show that these pretrained representations drive meaningful, task-relevant exploration and improve performance on 3D simulated environments. We also characterize why and how language provides useful abstractions for exploration by considering the impacts of using representations from a pretrained model, a language oracle, and several ablations. We demonstrate the benefits of our approach in two very different task domains -- one that stresses the identification and manipulation of everyday objects, and one that requires navigational exploration in an expansive world. Our results suggest that using language-shaped representations could improve exploration for various algorithms and agents in challenging environments.

Stephanie C. Y. Chan, Ishita Dasgupta, Junkyung Kim et al. · deepmind, stanford

Transformer models can use two fundamentally different kinds of information: information stored in weights during training, and information provided ``in-context'' at inference time. In this work, we show that transformers exhibit different inductive biases in how they represent and generalize from the information in these two sources. In particular, we characterize whether they generalize via parsimonious rules (rule-based generalization) or via direct comparison with observed examples (exemplar-based generalization). This is of important practical consequence, as it informs whether to encode information in weights or in context, depending on how we want models to use that information. In transformers trained on controlled stimuli, we find that generalization from weights is more rule-based whereas generalization from context is largely exemplar-based. In contrast, we find that in transformers pre-trained on natural language, in-context learning is significantly rule-based, with larger models showing more rule-basedness. We hypothesise that rule-based generalization from in-context information might be an emergent consequence of large-scale training on language, which has sparse rule-like structure. Using controlled stimuli, we verify that transformers pretrained on data containing sparse rule-like structure exhibit more rule-based generalization.

Stephanie C. Y. Chan, Andrew K. Lampinen, Pierre H. Richemond et al. · deepmind, stanford

As humans and animals learn in the natural world, they encounter distributions of entities, situations and events that are far from uniform. Typically, a relatively small set of experiences are encountered frequently, while many important experiences occur only rarely. The highly-skewed, heavy-tailed nature of reality poses particular learning challenges that humans and animals have met by evolving specialised memory systems. By contrast, most popular RL environments and benchmarks involve approximately uniform variation of properties, objects, situations or tasks. How will RL algorithms perform in worlds (like ours) where the distribution of environment features is far less uniform? To explore this question, we develop three complementary RL environments where the agent's experience varies according to a Zipfian (discrete power law) distribution. On these benchmarks, we find that standard Deep RL architectures and algorithms acquire useful knowledge of common situations and tasks, but fail to adequately learn about rarer ones. To understand this failure better, we explore how different aspects of current approaches may be adjusted to help improve performance on rare events, and show that the RL objective function, the agent's memory system and self-supervised learning objectives can all influence an agent's ability to learn from uncommon experiences. Together, these results show that learning robustly from skewed experience is a critical challenge for applying Deep RL methods beyond simulations or laboratories, and our Zipfian environments provide a basis for measuring future progress towards this goal.

Thilo Hagendorff, Ishita Dasgupta, Marcel Binz et al. · deepmind, stanford

Large language models (LLMs) show increasingly advanced emergent capabilities and are being incorporated across various societal domains. Understanding their behavior and reasoning abilities therefore holds significant importance. We argue that a fruitful direction for research is engaging LLMs in behavioral experiments inspired by psychology that have traditionally been aimed at understanding human cognition and behavior. In this article, we highlight and summarize theoretical perspectives, experimental paradigms, and computational analysis techniques that this approach brings to the table. It paves the way for a "machine psychology" for generative artificial intelligence (AI) that goes beyond performance benchmarks and focuses instead on computational insights that move us toward a better understanding and discovery of emergent abilities and behavioral patterns in LLMs. We review existing work taking this approach, synthesize best practices, and highlight promising future directions. We also highlight the important caveats of applying methodologies designed for understanding humans to machines. We posit that leveraging tools from experimental psychology to study AI will become increasingly valuable as models evolve to be more powerful, opaque, multi-modal, and integrated into complex real-world settings.

Aaditya K. Singh, Stephanie C. Y. Chan, Ted Moskovitz et al.

Transformer neural networks can exhibit a surprising capacity for in-context learning (ICL) despite not being explicitly trained for it. Prior work has provided a deeper understanding of how ICL emerges in transformers, e.g. through the lens of mechanistic interpretability, Bayesian inference, or by examining the distributional properties of training data. However, in each of these cases, ICL is treated largely as a persistent phenomenon; namely, once ICL emerges, it is assumed to persist asymptotically. Here, we show that the emergence of ICL during transformer training is, in fact, often transient. We train transformers on synthetic data designed so that both ICL and in-weights learning (IWL) strategies can lead to correct predictions. We find that ICL first emerges, then disappears and gives way to IWL, all while the training loss decreases, indicating an asymptotic preference for IWL. The transient nature of ICL is observed in transformers across a range of model sizes and datasets, raising the question of how much to "overtrain" transformers when seeking compact, cheaper-to-run models. We find that L2 regularization may offer a path to more persistent ICL that removes the need for early stopping based on ICL-style validation tasks. Finally, we present initial evidence that ICL transience may be caused by competition between ICL and IWL circuits.

Stephanie C. Y. Chan, Samuel Fishman, John Canny et al.

Lack of reliability is a well-known issue for reinforcement learning (RL) algorithms. This problem has gained increasing attention in recent years, and efforts to improve it have grown substantially. To aid RL researchers and production users with the evaluation and improvement of reliability, we propose a set of metrics that quantitatively measure different aspects of reliability. In this work, we focus on variability and risk, both during training and after learning (on a fixed policy). We designed these metrics to be general-purpose, and we also designed complementary statistical tests to enable rigorous comparisons on these metrics. In this paper, we first describe the desired properties of the metrics and their design, the aspects of reliability that they measure, and their applicability to different scenarios. We then describe the statistical tests and make additional practical recommendations for reporting results. The metrics and accompanying statistical tools have been made available as an open-source library at https://github.com/google-research/rl-reliability-metrics. We apply our metrics to a set of common RL algorithms and environments, compare them, and analyze the results.

SIMA Team, Maria Abi Raad, Arun Ahuja et al. · deepmind, stanford

Building embodied AI systems that can follow arbitrary language instructions in any 3D environment is a key challenge for creating general AI. Accomplishing this goal requires learning to ground language in perception and embodied actions, in order to accomplish complex tasks. The Scalable, Instructable, Multiworld Agent (SIMA) project tackles this by training agents to follow free-form instructions across a diverse range of virtual 3D environments, including curated research environments as well as open-ended, commercial video games. Our goal is to develop an instructable agent that can accomplish anything a human can do in any simulated 3D environment. Our approach focuses on language-driven generality while imposing minimal assumptions. Our agents interact with environments in real-time using a generic, human-like interface: the inputs are image observations and language instructions and the outputs are keyboard-and-mouse actions. This general approach is challenging, but it allows agents to ground language across many visually complex and semantically rich environments while also allowing us to readily run agents in new environments. In this paper we describe our motivation and goal, the initial progress we have made, and promising preliminary results on several diverse research environments and a variety of commercial video games.

Aaditya K. Singh, Ted Moskovitz, Felix Hill et al.

In-context learning is a powerful emergent ability in transformer models. Prior work in mechanistic interpretability has identified a circuit element that may be critical for in-context learning -- the induction head (IH), which performs a match-and-copy operation. During training of large transformers on natural language data, IHs emerge around the same time as a notable phase change in the loss. Despite the robust evidence for IHs and this interesting coincidence with the phase change, relatively little is known about the diversity and emergence dynamics of IHs. Why is there more than one IH, and how are they dependent on each other? Why do IHs appear all of a sudden, and what are the subcircuits that enable them to emerge? We answer these questions by studying IH emergence dynamics in a controlled setting by training on synthetic data. In doing so, we develop and share a novel optogenetics-inspired causal framework for modifying activations throughout training. Using this framework, we delineate the diverse and additive nature of IHs. By clamping subsets of activations throughout training, we then identify three underlying subcircuits that interact to drive IH formation, yielding the phase change. Furthermore, these subcircuits shed light on data-dependent properties of formation, such as phase change timing, already showing the promise of this more in-depth understanding of subcircuits that need to "go right" for an induction head.

Andrew Kyle Lampinen, Stephanie C. Y. Chan, Aaditya K. Singh et al. · deepmind, stanford

The ability of language models to learn a task from a few examples in context has generated substantial interest. Here, we provide a perspective that situates this type of supervised few-shot learning within a much broader spectrum of meta-learned in-context learning. Indeed, we suggest that any distribution of sequences in which context non-trivially decreases loss on subsequent predictions can be interpreted as eliciting a kind of in-context learning. We suggest that this perspective helps to unify the broad set of in-context abilities that language models exhibit -- such as adapting to tasks from instructions or role play, or extrapolating time series. This perspective also sheds light on potential roots of in-context learning in lower-level processing of linguistic dependencies (e.g. coreference or parallel structures). Finally, taking this perspective highlights the importance of generalization, which we suggest can be studied along several dimensions: not only the ability to learn something novel, but also flexibility in learning from different presentations, and in applying what is learned. We discuss broader connections to past literature in meta-learning and goal-conditioned agents, and other perspectives on learning and adaptation. We close by suggesting that research on in-context learning should consider this broader spectrum of in-context capabilities and types of generalization.

Andrew K. Lampinen, Arslan Chaudhry, Stephanie C. Y. Chan et al. · deepmind, stanford

Large language models exhibit exciting capabilities, yet can show surprisingly narrow generalization from finetuning. E.g. they can fail to generalize to simple reversals of relations they are trained on, or fail to make simple logical deductions based on trained information. These failures to generalize factual information from fine-tuning can significantly hinder the reasoning capabilities of these models. On the other hand, language models' in-context learning (ICL) shows different inductive biases and deductive reasoning capabilities. Here, we explore these differences in generalization and deductive reasoning between in-context- and fine-tuning-based learning. To do so, we constructed several novel datasets to evaluate and improve models' abilities to make generalizations over factual information from novel data. These datasets are designed to create clean tests of generalization, by isolating the knowledge in the dataset from that in pretraining. We expose pretrained large models to controlled subsets of the information in these datasets -- either through ICL or fine-tuning -- and evaluate their performance on test sets that require various types of generalization. We find overall that in data-matched settings, ICL can generalize several types of inferences more flexibly than fine-tuning (though we also find some qualifications of prior findings, such as cases when fine-tuning can generalize to reversals embedded in a larger structure of knowledge). We build on these findings to propose a method to enable improved generalization from fine-tuning: adding in-context reasoning traces to finetuning data. We show that this method improves generalization across various splits of our datasets and other benchmarks. Our results have implications for understanding the generalization afforded by different modes of learning in language models, and practically improving their performance.

Aaditya K. Singh, Ted Moskovitz, Sara Dragutinovic et al.

In-context learning (ICL) is a powerful ability that emerges in transformer models, enabling them to learn from context without weight updates. Recent work has established emergent ICL as a transient phenomenon that can sometimes disappear after long training times. In this work, we sought a mechanistic understanding of these transient dynamics. Firstly, we find that, after the disappearance of ICL, the asymptotic strategy is a remarkable hybrid between in-weights and in-context learning, which we term "context-constrained in-weights learning" (CIWL). CIWL is in competition with ICL, and eventually replaces it as the dominant strategy of the model (thus leading to ICL transience). However, we also find that the two competing strategies actually share sub-circuits, which gives rise to cooperative dynamics as well. For example, in our setup, ICL is unable to emerge quickly on its own, and can only be enabled through the simultaneous slow development of asymptotic CIWL. CIWL thus both cooperates and competes with ICL, a phenomenon we term "strategy coopetition." We propose a minimal mathematical model that reproduces these key dynamics and interactions. Informed by this model, we were able to identify a setup where ICL is truly emergent and persistent.

Andrew Kyle Lampinen, Stephanie C. Y. Chan, Yuxuan Li et al. · deepmind, stanford

A common approach in neuroscience is to study neural representations as a means to understand a system -- increasingly, by relating the neural representations to the internal representations learned by computational models. However, a recent work in machine learning (Lampinen, 2024) shows that learned feature representations may be biased to over-represent certain features, and represent others more weakly and less-consistently. For example, simple (linear) features may be more strongly and more consistently represented than complex (highly nonlinear) features. These biases could pose challenges for achieving full understanding of a system through representational analysis. In this perspective, we illustrate these challenges -- showing how feature representation biases can lead to strongly biased inferences from common analyses like PCA, regression, and RSA. We also present homomorphic encryption as a simple case study of the potential for strong dissociation between patterns of representation and computation. We discuss the implications of these results for representational comparisons between systems, and for neuroscience more generally.

Nicolas Zucchet, Francesco d'Angelo, Andrew K. Lampinen et al. · deepmind, stanford

Emergence is a fascinating property of large language models and neural networks more broadly: as models scale and train for longer, they sometimes develop new abilities in sudden ways. Despite initial studies, we still lack a comprehensive understanding of how and when these abilities emerge. To address this gap, we study the emergence over training of sparse attention, a critical and frequently observed attention pattern in Transformers. By combining theoretical analysis of a toy model with empirical observations on small Transformers trained on a linear regression variant, we uncover the mechanics driving sparse attention emergence and reveal that emergence timing follows power laws based on task structure, architecture, and optimizer choice. We additionally find that repetition can greatly speed up emergence. Finally, we confirm these results on a well-studied in-context associative recall task. Our findings provide a simple, theoretically grounded framework for understanding how data distributions and model design influence the learning dynamics behind one form of emergence.

Yuxuan Li, Declan Campbell, Stephanie C. Y. Chan et al. · deepmind, stanford

Many of language models' impressive capabilities originate from their in-context learning: based on instructions or examples, they can infer and perform new tasks without weight updates. In this work, we investigate when representations for new tasks are formed in language models, and how these representations change over the course of context. We focus on ''transferrable'' task representations -- vector representations that can restore task contexts in another instance of the model, even without the full prompt. We show that these representations evolve in non-monotonic and sporadic ways, and are distinct from a more inert representation of high-level task categories that persists throughout the context. Specifically, when more examples are provided in the context, transferrable task representations successfully condense evidence. This allows better transfer of task contexts and aligns well with the performance improvement. However, this evidence accrual process exhibits strong locality along the sequence dimension, coming online only at certain tokens -- despite task identity being reliably decodable throughout the context. Moreover, these local but transferrable task representations tend to capture minimal ''task scopes'', such as a semantically-independent subtask. For longer and composite tasks, models rely on more temporally-distributed representations. This two-fold locality (temporal and semantic) underscores a kind of just-in-time computational process that language models use to perform new tasks on the fly.

Alexander Y. Ku, Thomas L. Griffiths, Stephanie C. Y. Chan

Transformer models learn in two distinct modes: in-weights learning (IWL), encoding knowledge into model weights, and in-context learning (ICL), adapting flexibly to context without weight modification. To better understand the interplay between these learning modes, we draw inspiration from evolutionary biology's analogous adaptive strategies: genetic encoding (akin to IWL, adapting over generations and fixed within an individual's lifetime) and phenotypic plasticity (akin to ICL, enabling flexible behavioral responses to environmental cues). In evolutionary biology, environmental predictability dictates the balance between these strategies: stability favors genetic encoding, while reliable predictive cues promote phenotypic plasticity. We experimentally operationalize these dimensions of predictability and systematically investigate their influence on the ICL/IWL balance in Transformers. Using regression and classification tasks, we show that high environmental stability decisively favors IWL, as predicted, with a sharp transition at maximal stability. Conversely, high cue reliability enhances ICL efficacy, particularly when stability is low. Furthermore, learning dynamics reveal task-contingent temporal evolution: while a canonical ICL-to-IWL shift occurs in some settings (e.g., classification with many classes), we demonstrate that scenarios with easier IWL (e.g., fewer classes) or slower ICL acquisition (e.g., regression) can exhibit an initial IWL phase later yielding to ICL dominance. These findings support a relative-cost hypothesis for explaining these learning mode transitions, establishing predictability as a critical factor governing adaptive strategies in Transformers, and offering novel insights for understanding ICL and guiding training methodologies.

Andrew Kyle Lampinen, Stephanie C. Y. Chan, Katherine Hermann

Representation learning, and interpreting learned representations, are key areas of focus in machine learning and neuroscience. Both fields generally use representations as a means to understand or improve a system's computations. In this work, however, we explore surprising dissociations between representation and computation that may pose challenges for such efforts. We create datasets in which we attempt to match the computational role that different features play, while manipulating other properties of the features or the data. We train various deep learning architectures to compute these multiple abstract features about their inputs. We find that their learned feature representations are systematically biased towards representing some features more strongly than others, depending upon extraneous properties such as feature complexity, the order in which features are learned, and the distribution of features over the inputs. For example, features that are simpler to compute or learned first tend to be represented more strongly and densely than features that are more complex or learned later, even if all features are learned equally well. We also explore how these biases are affected by architectures, optimizers, and training regimes (e.g., in transformers, features decoded earlier in the output sequence also tend to be represented more strongly). Our results help to characterize the inductive biases of gradient-based representation learning. We then illustrate the downstream effects of these biases on various commonly-used methods for analyzing or intervening on representations. These results highlight a key challenge for interpretability $-$ or for comparing the representations of models and brains $-$ disentangling extraneous biases from the computationally important aspects of a system's internal representations.

Andrew K. Lampinen, Nicholas A. Roy, Ishita Dasgupta et al.

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.

Andrew Kyle Lampinen, Stephanie C. Y. Chan, Andrea Banino et al.

Reinforcement learning agents often forget details of the past, especially after delays or distractor tasks. Agents with common memory architectures struggle to recall and integrate across multiple timesteps of a past event, or even to recall the details of a single timestep that is followed by distractor tasks. To address these limitations, we propose a Hierarchical Chunk Attention Memory (HCAM), which helps agents to remember the past in detail. HCAM stores memories by dividing the past into chunks, and recalls by first performing high-level attention over coarse summaries of the chunks, and then performing detailed attention within only the most relevant chunks. An agent with HCAM can therefore "mentally time-travel" -- remember past events in detail without attending to all intervening events. We show that agents with HCAM substantially outperform agents with other memory architectures at tasks requiring long-term recall, retention, or reasoning over memory. These include recalling where an object is hidden in a 3D environment, rapidly learning to navigate efficiently in a new neighborhood, and rapidly learning and retaining new object names. Agents with HCAM can extrapolate to task sequences much longer than they were trained on, and can even generalize zero-shot from a meta-learning setting to maintaining knowledge across episodes. HCAM improves agent sample efficiency, generalization, and generality (by solving tasks that previously required specialized architectures). Our work is a step towards agents that can learn, interact, and adapt in complex and temporally-extended environments.