How can we build AI systems that can learn any set of individual human values both quickly and safely, avoiding causing harm or violating societal standards for acceptable behavior during the learning process? We explore the effects of representational alignment between humans and AI agents on learning human values.

How can we build AI systems that can learn any set of individual human values both quickly and safely, avoiding causing harm or violating societal standards for acceptable behavior during the learning process? We explore the effects of representational alignment between humans and AI agents on learning human values. Making AI systems learn human-like representations of the world has many known benefits, including improving generalization, robustness to domain shifts, and few-shot learning performance. We demonstrate that this kind of representational alignment can also support safely learning and exploring human values in the context of personalization. We begin with a theoretical prediction, show that it applies to learning human morality judgments, then show that our results generalize to ten different aspects of human values -- including ethics, honesty, and fairness -- training AI agents on each set of values in a multi-armed bandit setting, where rewards reflect human value judgments over the chosen action. Using a set of textual action descriptions, we collect value judgments from humans, as well as similarity judgments from both humans and multiple language models, and demonstrate that representational alignment enables both safe exploration and improved generalization when learning human values.

Should we care whether AI systems have representations of the world that are similar to those of humans?

The superposition hypothesis states that single neurons may participate in representing multiple features in order for the neural network to represent more features than it has neurons. In neuroscience and AI, representational alignment metrics measure the extent to which different deep neural networks (DNNs) or brains represent similar information. In this work, we explore a critical question: does superposition interact with alignment metrics in any undesirable way? We hypothesize that models which represent the same features in different superposition arrangements, i.e., their neurons have different linear combinations of the features, will interfere with predictive mapping metrics (semi-matching, soft-matching, linear regression), producing lower alignment than expected. We develop a theory for how permutation metrics are dependent on superposition arrangements. This is tested by training sparse autoencoders (SAEs) to disentangle superposition in toy models, where alignment scores are shown to typically increase when a model's base neurons are replaced with its sparse overcomplete latent codes. We find similar increases for DNN-DNN and DNN-brain linear regression alignment in the visual domain. Our results suggest that superposition disentanglement is necessary for mapping metrics to uncover the true representational alignment between neural networks.

Spatial functional organization is a hallmark of biological brains: neurons are arranged topographically according to their response properties, at multiple scales. In contrast, representations within most machine learning models lack spatial biases, instead manifesting as disorganized vector spaces that are difficult to visualize and interpret. Here, we propose a novel form of self-attention that turns Transformers into "Topoformers" with topographic organization. We introduce spatial querying - where keys and queries are arranged on 2D grids, and local pools of queries are associated with a given key - and spatial reweighting, where we convert the standard fully connected layer of self-attention into a locally connected layer. We first demonstrate the feasibility of our approach by training a 1-layer Topoformer on a sentiment classification task. Training with spatial querying encourages topographic organization in the queries and keys, and spatial reweighting separately encourages topographic organization in the values and self-attention outputs. We then apply the Topoformer motifs at scale, training a BERT architecture with a masked language modeling objective. We find that the topographic variant performs on par with a non-topographic control model on NLP benchmarks, yet produces interpretable topographic organization as evaluated via eight linguistic test suites. Finally, analyzing an fMRI dataset of human brain responses to a large set of naturalistic sentences, we demonstrate alignment between low-dimensional topographic variability in the Topoformer model and human brain language network. Scaling up Topoformers further holds promise for greater interpretability in NLP research, and for more accurate models of the organization of linguistic information in the human brain.

The ability to translate diverse patterns of inputs into structured patterns of behavior has been thought to rest on both humans' and machines' ability to learn robust representations of relevant concepts. The rapid advancement of transformer-based large language models (LLMs) has led to a diversity of computational ingredients -- architectures, fine tuning methods, and training datasets among others -- but it remains unclear which of these ingredients are most crucial for building models that develop human-like representations. Further, most current LLM benchmarks are not suited to measuring representational alignment between humans and models, making benchmark scores unreliable for assessing if current LLMs are making progress towards becoming useful cognitive models. We address these limitations by first evaluating a set of over 70 models that widely vary in their computational ingredients on a triplet similarity task, a method well established in the cognitive sciences for measuring human conceptual representations, using concepts from the THINGS database. Comparing human and model representations, we find that models that undergo instruction-finetuning and which have larger dimensionality of attention heads are among the most human aligned, while multimodal pretraining and parameter size have limited bearing on alignment. Correlations between alignment scores and scores on existing benchmarks reveal that while some benchmarks (e.g., MMLU) are better suited than others (e.g., MUSR) for capturing representational alignment, no existing benchmark is capable of fully accounting for the variance of alignment scores, demonstrating their insufficiency in capturing human-AI alignment. Taken together, our findings help highlight the computational ingredients most essential for advancing LLMs towards models of human conceptual representation and address a key benchmarking gap in LLM evaluation.

We test the Platonic Representation Hypothesis (PRH) in astronomy by measuring representational convergence across a range of foundation models trained on different data types. Using spectroscopic and imaging observations from JWST, HSC, Legacy Survey, and DESI, we compare representations from vision transformers, self-supervised models, and astronomy-specific architectures via mutual $k$-nearest neighbour analysis. We observe consistent scaling: representational alignment generally increases with model capacity across our tested architectures, supporting convergence toward a shared representation of galaxy astrophysics. Our results suggest that astronomical foundation models can use pre-trained general-purpose architectures, allowing us to capitalise on the broader machine learning community's already-spent computational investment.

Humans can effortlessly describe what they see, yet establishing a shared representational format between vision and language remains a significant challenge. Emerging evidence suggests that human brain representations in both vision and language are well predicted by semantic feature spaces obtained from large language models (LLMs). This raises the possibility that sensory systems converge in their inherent ability to transform their inputs onto shared, embedding-like representational space. However, it remains unclear how such a space manifests in human behaviour. To investigate this, sixty-three participants performed behavioural similarity judgements separately on 100 natural scene images and 100 corresponding sentence captions from the Natural Scenes Dataset. We found that visual and linguistic similarity judgements not only converge at the behavioural level but also predict a remarkably similar network of fMRI brain responses evoked by viewing the natural scene images. Furthermore, computational models trained to map images onto LLM-embeddings outperformed both category-trained and AlexNet controls in explaining the behavioural similarity structure. These findings demonstrate that human visual and linguistic similarity judgements are grounded in a shared, modality-agnostic representational structure that mirrors how the visual system encodes experience. The convergence between sensory and artificial systems suggests a common capacity of how conceptual representations are formed-not as arbitrary products of first order, modality-specific input, but as structured representations that reflect the stable, relational properties of the external world.