Can Machines Learn Without Humans? The Rise of Self-Supervised Learning

Teaching a child to recognize a cat takes very little effort.

You point at one. You say the word. Maybe you do it two or three times. And the child, drawing on an extraordinarily rich web of prior experience, context, and intuition, understands. Not just that specific cat. Cats in general. Cats from unusual angles. Cats in poor lighting. Cartoon cats. Even cats partially hidden behind furniture.

Teaching a traditional machine learning system to recognize a cat takes something entirely different: thousands, sometimes hundreds of thousands of labelled images, each one painstakingly tagged by a human annotator. Cat. Not cat. Cat. Not cat. The system learns, but only because an army of people spent enormous time and money generating the teaching signal it needed.

This difference, between how biological intelligence learns and how artificial intelligence has traditionally learned, is one of the most fundamental limitations in the history of machine learning. And it is a limitation that a powerful new paradigm is quietly, decisively dismantling.

Self-supervised learning, the ability of AI systems to learn from vast quantities of unlabeled data, without human annotation, by discovering structure in the data itself, is one of the most significant developments in machine learning of the past decade. It is the engine powering the large language models that have transformed public awareness of AI. It is enabling breakthroughs in computer vision, biology, robotics, and scientific research. And it represents a fundamental shift in the relationship between human effort and machine capability.

This article explains what self-supervised learning is, why it matters, where it is producing remarkable results, and what it means for the future of artificial intelligence, explained clearly, for anyone.

The Problem With Supervised Learning

To understand why self-supervised learning is significant, you first need to understand what it is replacing, and why that approach has fundamental limitations.

Traditional machine learning is predominantly supervised learning: a paradigm in which an AI system learns by being shown examples that have been labelled by humans. For image recognition, the system sees images labelled with their contents. For spam detection, it sees emails labelled as spam or not spam. For language translation, it sees sentence pairs labelled with their equivalents in another language.

The system learns by comparing its predictions to the human-provided labels, measuring its errors, and adjusting its internal parameters to reduce those errors over many thousands or millions of examples. With enough labelled data, it can develop remarkable accuracy.

The critical dependency is on the labelled data, and producing labelled data at scale is expensive, slow, and limited.

The Labelling Bottleneck

Consider what labelling requires. For medical imaging, training AI to detect tumors in radiology scans, each image must be labelled by a qualified radiologist. Expert time is scarce and expensive. Training a high-quality medical imaging AI might require tens of thousands of expert-labelled images, a dataset that represents years of radiologist time and millions of dollars of cost.

For language understanding, training AI to comprehend the nuances of legal documents, scientific papers, or financial contracts, labelling requires domain experts who understand the material deeply enough to create accurate annotations. The cost of expert labelling in specialized domains is often prohibitive.

For any new domain, any new language, any new task, the labelling bottleneck reappears. Supervised learning is powerful within the domains where labelled data is abundant. Outside those domains, it struggles.

And this limitation is not just a practical inconvenience. It reflects something fundamentally different from how biological intelligence works. Humans do not learn through labelled examples. We learn through immersion, through experiencing the world, finding patterns, building internal models of reality, and then applying those models to new situations. A child who has never been told what makes one thing heavier than another can nevertheless lift two objects and infer their relative weights. We learn from the structure of our experience, not from the annotations attached to it.

Self-supervised learning is an attempt to give AI systems something closer to this kind of learning, the ability to extract knowledge from the structure of raw, unlabeled data itself.

What Is Self-Supervised Learning?

Self-supervised learning is a machine learning paradigm in which a system generates its own supervisory signal from the structure of unlabeled data, without any human annotation.

The key insight is that real-world data has intrinsic structure. Language has grammar, coherence, and statistical regularities. Images have spatial relationships, object persistence, and visual consistency across viewpoints. Video has temporal continuity. Audio has acoustic patterns and rhythmic structure. Biological sequences have evolutionary relationships encoded in their variation.

A system that can be trained to predict these structural properties, to predict the next word in a sentence, to reconstruct a masked region of an image, to predict the next frame of a video, is simultaneously learning the underlying regularities of the data domain. The prediction task is a pretext, a way of generating a training signal, but the knowledge the system develops to perform the task is real and transferable.

The elegance of this approach is that the supervisory signal is effectively free. You do not need human annotators to tell the system what is in an image. You mask part of the image and ask the system to reconstruct it. The image itself provides the teaching signal; the correct reconstruction is the original unmasked image. The human role in the learning process is reduced to selecting and curating the raw data, not annotating it.

Mechanics: How Self-Supervised Learning Actually Works

Self-supervised learning is not a single technique, it is a family of approaches, unified by the principle of generating supervisory signal from data structure. Understanding the main variants clarifies both how technology works and why it is so powerful.

Masked Prediction: The Language Model Approach

The approach that has had the most visible impact is masked prediction in language, and it is the foundation of the large language models that have transformed public awareness of AI.

The training task is deceptively simple: given a sequence of text with some words removed (masked), predict the missing words. The model sees "The cat sat on the [MASK] and fell asleep" and must predict that the missing word is likely "mat" or "sofa" or another contextually appropriate term.

To predict the missing word accurately, the model must develop a deep understanding of language, grammar, context, meaning, and the relationships between concepts. It cannot simply memories the training data. It must learn the underlying structure of language well enough to generalize novel sentences it has never seen.

This approach, used in Google's BERT model and subsequently refined in countless variants, trains on text that exists in the world: books, websites, scientific papers, code repositories, and any other text that can be collected (Devlin et al., 2019). The internet contains hundreds of trillions of words. The labelling cost of training a masked language model on this data is essentially zero.

Autoregressive language models, like the GPT series from OpenAI and Claude from Anthropic, use a slightly different self-supervised task: predict the next word given all previous words. Rather than masking tokens at random, these models learn by trying to predict each word in a sequence from the words that came before it. Trained at scale on vast corpora of text, these models develop a remarkable range of capabilities, writing, reasoning, coding, question answering, that characterize modern large language models.

The same text that exists in the world provides all the training signal needed. No human annotation required.

Contrastive Learning: Learning by Comparison

Contrastive learning is a self-supervised approach that teaches a model to recognize similarity and difference without being told what categories objects belong to.

The approach works by creating pairs of related and unrelated examples and training the model to represent related examples similarly (close together in its internal representation space) and unrelated examples differently (far apart).

For images, related examples are typically different views of the same image, the same photograph randomly cropped, color-adjusted, rotated, or otherwise augmented. Unrelated examples are images from different photographs. The model learns to produce similar internal representations for the different views of the same image, and different representations for genuinely different images.

To do this well, the model cannot simply memories the transformations applied to create the pairs. It must learn the underlying visual content, what makes two views of the same image the same image, which requires developing representations of objects, scenes, and visual relationships. These representations turn out to be remarkably useful for downstream tasks like object recognition, even though the model was never told what any object is.

SimCLR, MoCo, and CLIP, the foundational contrastive learning models, demonstrated that models trained this way develop visual representations comparable to or exceeding supervised models trained on the same downstream tasks, using far less labelled data for fine-tuning (Chen et al., 2020).

CLIP, developed by OpenAI, took this further by training on image-text pairs, photographs paired with their captions, from the internet, teaching a model to align visual and textual representations without requiring any task-specific annotations (Radford et al., 2021). CLIP's ability to understand images in terms of natural language descriptions, learned entirely from self-supervised training on web data, enabled a range of downstream applications that had previously required extensive supervised training.

Masked Autoencoders: Learning to Reconstruct

The masked prediction approach from language has been successfully applied to images through masked autoencoders (MAE), an approach developed by researchers at Meta AI (He et al., 2022).

A masked autoencoder takes an image, masks a large proportion of it, typically 75%, and trains the model to reconstruct the missing regions from the visible portions. To do this convincingly, the model must develop a rich understanding of the visual structure of images, the spatial relationships between objects, the textures and patterns that characterize different surfaces, and the way visual scenes are organized.

The results were striking models trained with MAE on unlabeled image data developed visual representations competitive with the best supervised models, using a fraction of the labelled data for downstream fine-tuning. The approach has since been applied to video, point clouds (3D data), and multimodal data.

Self-Supervised Learning in Biology: AlphaFold and Beyond

One of the most consequential applications of self-supervised learning is in molecular biology, specifically, in learning the structure and function of proteins from the evolutionary record encoded in protein sequences.

Proteins are chains of amino acids whose biological function is determined by their three-dimensional structure, a structure that has been extraordinarily difficult and expensive to determine experimentally. The protein structure prediction problem, predicting a protein's 3D structure from its amino acid sequence, was one of the defining unsolved problems of structural biology for fifty years.

DeepMind's AlphaFold2, which essentially solved this problem, used a form of self-supervised learning on evolutionary sequence data, learning the relationships between amino acid sequences and protein structures by processing the enormous database of known protein sequences and the patterns of variation within protein families (Jumper et al., 2021). The supervisory signal came from the structure of the evolutionary data itself, not from human annotation.

More recently, protein language models, self-supervised models trained to predict masked amino acids in protein sequences, directly analogous to BERT in natural language, have demonstrated that the statistical structure of protein sequences encodes information about protein structure and function that the model can learn without any explicit structural labels (Lin et al., 2022). ESM-2, a protein language model developed by Meta AI, learned representations of protein sequences that enable structure prediction and functional annotation with remarkable accuracy, trained entirely through self-supervised learning on protein sequence databases.

The implications for drug discovery, synthetic biology, and medicine are profound. The ability to learn the language of proteins, as self-supervised models have learned the language of human text, opens pathways to understanding and engineering biological molecules at a scale and speed that was previously impossible.

Why Scale Changes Everything

One of the most important properties of self-supervised learning is that it benefits dramatically from scale, in ways that supervised learning, limited by the cost of labelled data, simply cannot.

In supervised learning, you can only use as much data as you can afford to label. In self-supervised learning, you can use essentially as much data as it exists, and in many domains, particularly language and images, the amount of data that exists is extraordinary.

The internet contains an estimated 100 trillion words of text (across all languages, websites, books, and documents). Supervised learning can only use the small fraction that has been labelled. Self-supervised learning can use all of it.

This scaling insight is captured in what has become known in the field as the scaling laws for language models, empirical relationships between model size, training data volume, compute budget, and model capability (Kaplan et al., 2020). Within certain ranges, larger models trained on more data with more compute produce better models in predictable ways, and the ceiling on how much unlabeled data is available is far, far above the point where current models are trained.

The practical implication is that self-supervised learning has enabled a qualitative shift in what machine learning can achieve, not because the algorithms are fundamentally different from what existed before, but because removing the labelling bottleneck allows the algorithms to operate at a scale that reveals capabilities that do not emerge at smaller scales.

This phenomenon, sometimes called emergence, is one of the most surprising and important properties of large-scale self-supervised models. As models grow larger and are trained in more data, new capabilities appear that were not present in smaller models and that were not explicitly trained for. Large language models develop the ability to reason, to translate between languages they were not explicitly trained on, to write code, to solve mathematical problems, capabilities that emerged from the self-supervised training objective of next-word prediction, applied at sufficient scale (Wei et al., 2022).

Foundation Models: The Paradigm Shift

The convergence of self-supervised learning with large-scale training has produced a new paradigm in machine learning: foundation models.

A foundation model is a large model trained at scale on broad, unlabeled data through self-supervised learning, which can then be adapted to a wide range of downstream tasks through fine-tuning on relatively small amounts of task-specific data (Bommasani et al., 2021).

The name reflects the structural role these models play. Rather than training a separate model for each specific task, one model for medical image analysis, another for legal document review, another for financial prediction, a single large foundation model provides a rich, general-purpose representation of the relevant domain (language, images, proteins, etc.) that can be adapted to many specific tasks.

This represents a fundamental change in the economics of machine learning. The expensive part, training the large self-supervised model, happens once, at scale, using unlabeled data. The cheap part, adapting it to specific tasks, requires only small amounts of task-specific labelled data and limited compute. The marginal cost of deploying machine learning in a new domain drops dramatically.

Foundation models now exist across multiple domains:

Language: GPT-4, Claude, Gemini, and Llama are all language foundation models, trained self-supervised on vast text corpora, adaptable to an enormous range of language tasks.

Vision: CLIP, DINO, and SAM (Segment Anything Model) from Meta AI are vision foundation models, trained self-supervised on image and image-text data, adaptable to visual recognition, segmentation, and understanding tasks.

Biology: ESM-2 and AlphaFold are biology foundation models, trained self-supervised protein sequence and structure data, adaptable to protein function prediction, drug target identification, and protein design.

Multimodal: GPT-4V, Gemini, and Claude 3 are multimodal foundation models, combining vision and language understanding, trained self-supervised on text-image pairs, adaptable to tasks requiring understanding across both modalities.

The foundation model paradigm, enabled by self-supervised learning, has become the dominant approach for building capable AI systems in 2026.

Self-Supervised Learning in Robotics: Teaching Robots to Learn from Experience

One of the most exciting frontiers for self-supervised learning is robotics, where the bottleneck of labelled data is particularly severe because collecting robot experience in the real world is slow, expensive, and potentially dangerous.

Traditional approaches to robot learning, programming explicit behaviors, or training on labelled datasets of robot actions and outcomes, scale poorly to the enormous diversity of real-world environments and tasks. Robots trained in one environment often fail catastrophically in slightly different environments, because they have learned rules rather than underlying principles.

Self-supervised learning offers a different approach: train robots to learn from their own experience by predicting the consequences of their actions, reconstructing observations from partial information, or developing internal models of their physical environment, all without human labels.

Google's RT-2 model, a large vision-language-action model trained through a combination of self-supervised learning and robotic experience, demonstrated that robots trained at scale on diverse web data and robotic data can develop remarkably general manipulation capabilities that transfer to novel objects and novel instructions (Brohan et al., 2023).

Meta AI's work on self-supervised learning for robot locomotion, teaching robots to walk, navigate, and recover from perturbations through self-supervised interaction with their physical environment, has demonstrated that robots can develop robust locomotion skills through experience-based self-supervised learning rather than hand-engineered controllers.

Vision is a robot that learns about the physical world the way children do, through active exploration, through interacting with objects, through building internal models of how the world responds to actions, rather than through labelled demonstrations. Self-supervised learning is the key technical ingredient making this vision increasingly plausible.

What Self-Supervised Learning Cannot Do, Yet

Honest assessment of this technology requires acknowledging where it falls short.

Grounding in physical reality remains limited. Large self-supervised language models learn from text, a representation of human experience, not experience itself. They can discuss the properties of physical objects, but their "understanding" is derived from text descriptions rather than from sensory experience. This produces characteristic failure modes: confident statements about physical phenomena that contradict basic physics, because the model has learned the patterns of how people describe physical phenomena rather than the phenomena themselves.

Self-supervised learning can encode the biases in its training data. The statistical structure that self-supervised models learn from unlabeled data includes the biases, errors, and distortions present in that data. A language model trained on internet text learns the statistical regularities of internet text, including its gender biases, racial biases, and the underrepresentation of perspectives that are underrepresented online.

Emergent capabilities are not fully predictable or controllable. The emergence of capabilities in large self-supervised models, capabilities that were not present in smaller models and were not explicitly trained for, is both what makes these models powerful and what makes them difficult to control and understand. The same process that produces impressive reasoning capabilities can produce unexpected and unwanted behaviors.

Data curation matters enormously. Self-supervised learning is not truly unsupervised, significant human effort goes into selecting, filtering, and curating the training data. The quality of a self-supervised model is strongly dependent on the quality and composition of its training data. The human role has shifted from annotating data to curating it, which is a different but still substantial form of human involvement.

What the Future Holds

The trajectory of self-supervised learning research points toward several significant near-term developments.

World models. The most ambitious near-term direction for self-supervised learning is the development of world models, AI systems that learn internal models of the causal structure of reality, enabling them to predict the consequences of actions and plan toward goals in ways that current systems cannot. Yann LeCun at Meta AI has argued that world models, trained self-supervised on sensory experience, are the key missing ingredient for human-level AI (LeCun, 2022). The research programmed is ambitious and the technical challenges are severe, but the direction is increasingly well-defined.

Multimodal foundation models. The boundaries between modalities, language, vision, audio, video, and physical interaction, are dissolving in self-supervised foundation models that are trained across all of them simultaneously. The resulting models develop richer representations that capture relationships between modalities that single-modality models miss.

Continual learning. Current self-supervised models are trained once on a fixed dataset and then deployed. Future systems will learn continuously from ongoing experience, updating their representations as the world changes, as they encounter new situations, and as new data becomes available.

Science as a domain. The application of self-supervised learning to scientific data, protein sequences, molecular structures, physical simulations, astronomical observations, is one of the most consequential research directions of the coming decade. The ability to learn the "language" of different scientific domains from the statistical structure of scientific data could accelerate discovery in ways that are difficult to overestimate.

What This Means for Everyone

Self-supervised learning is not a niche technical development that concerns only researchers. Its implications touch anyone who uses AI-powered technology, which, in 2026, is most people.

The AI tools you use are powered by self-supervised learning. Every time you interact with a large language model, whether for writing assistance, research, coding, translation, or conversation, you are using a system whose fundamental capabilities were developed through self-supervised learning on vast amounts of unlabeled text.

The cost of AI is falling because labelling is no longer the bottleneck. Self-supervised learning has fundamentally changed the economics of AI development, enabling capabilities that would have required prohibitively expensive labelling efforts under the supervised paradigm. This democratization of AI capability has broad implications for who can build AI systems and what they can build them to do.

The gap between human and machine learning is narrowing. The move from supervised learning, which required human annotation for every piece of knowledge transferred to the system, toward self-supervised learning, which extracts knowledge from the structure of data itself, represents a meaningful step toward learning paradigms that more closely resemble how biological intelligence works.

The data that exists in the world has become a resource. Self-supervised learning has made the vast quantities of unlabeled data that exist, text, images, video, biological sequences, sensor readings, into training material for AI systems. This has implications for privacy, for intellectual property, for the representation of different communities and languages in AI systems, and for the power dynamics of AI development.

The Bottom Line

The question "can machines learn without humans?" has a precise and important answer in 2026: not entirely, but far more than was possible before.

Self-supervised learning has not eliminated the human role in machine learning. It has shifted that role, from the laborious work of annotating millions of data points to the higher-level work of selecting and curating training data, designing learning objectives, and guiding the development of systems that learn from the structure of the world itself.

In doing so, it has removed the most significant bottleneck in AI development, the need for labelled data, and enabled a qualitative leap in AI capability that has fundamentally changed what AI systems can do and where they can be deployed.

The large language models, the vision-language systems, the protein structure predictors, and the robotic learning systems that represent the frontier of AI in 2026 are all, in different ways, products of this paradigm shift. Their remarkable capabilities, emerging from training objectives that ask systems to predict, reconstruct, and make sense of unlabeled data, represent what happens when machine learning is freed from the labelling bottleneck and given access to the vast, structured, information-rich data that the world naturally produces.

Children learn about the world by experiencing it, by finding patterns, building models, and testing predictions against reality. Self-supervised learning is, at last, giving machines something closer to that kind of learning. What that means for the long arc of artificial intelligence is a question whose answer is still being written, one unlabeled data point at a time.

Cover Image by Freepik [www.freepik.com]

References

Bommasani, R., Hudson, D.A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M.S., Bohg, J., Bosselut, A., Brunskill, E., Brynjolfsson, E. and Liang, P. (2021) 'On the opportunities and risks of foundation models', arXiv preprint arXiv:2108.07258. Available at: https://arxiv.org/abs/2108.07258 (Accessed: 20 June 2026).

Brohan, A., Brown, N., Carbajal, J., Chebotar, Y., Chen, X., Choromanski, K., Ding, T., Driess, D., Dubey, A., Finn, C., Florence, P., Fu, C., Arenas, M.G., Gopalakrishnan, K., Han, K., Hausman, K., Herzog, A., Hsu, J., Ichter, B., Irpan, A., Jang, E., Joshi, N., Julian, R., Kalashnikov, D., Kuang, Y., Leal, I., Lee, L., Li, S., Liang, J., Leong, W.L., Levine, S., Lu, Y., Malla, H. and Zeng, A. (2023) 'RT-2: vision-language-action models transfer web knowledge to robotic control', arXiv preprint arXiv:2307.15818. Available at: https://arxiv.org/abs/2307.15818 (Accessed: 20 June 2026).

Chen, T., Kornblith, S., Norouzi, M. and Hinton, G. (2020) 'A simple framework for contrastive learning of visual representations', in Proceedings of the 37th International Conference on Machine Learning (ICML 2020), pp. 1597–1607. Available at: https://arxiv.org/abs/2002.05709 (Accessed: 19 June 2026).

Devlin, J., Chang, M.W., Lee, K. and Toutanova, K. (2019) 'BERT: pre-training of deep bidirectional transformers for language understanding', in Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1, pp. 4171–4186. doi:10.18653/v1/N19-1423.

He, K., Chen, X., Xie, S., Li, Y., Dollár, P. and Girshick, R. (2022) 'Masked autoencoders are scalable vision learners', in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR 2022), pp. 16000–16009. Available at: https://arxiv.org/abs/2111.06377 (Accessed: 19 June 2026).

Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., Bridgland, A., Meyer, C., Kohl, S.A.A., Ballard, A.J., Cowie, A., Romera-Paredes, B., Nikolov, S., Jain, R., Adler, J., Back, T., Petersen, S., Reiman, D., Clancy, E., Zielinski, M., Steinegger, M., Pacholska, M., Berghammer, T., Bodenstein, S., Silver, D., Vinyals, O., Senior, A.W., Kavukcuoglu, K., Kohli, P. and Hassabis, D. (2021) 'Highly accurate protein structure prediction with AlphaFold', Nature, 596(7873), pp. 583–589. doi:10.1038/s41586-021-03819-2.

Kaplan, J., McCandlish, S., Henighan, T., Brown, T.B., Chess, B., Child, R., Gray, S., Radford, A., Wu, J. and Amodei, D. (2020) 'Scaling laws for neural language models', arXiv preprint arXiv:2001.08361. Available at: https://arxiv.org/abs/2001.08361 (Accessed: 19 June 2026).

LeCun, Y. (2022) A path towards autonomous machine intelligence. New York: Meta AI. Available at: https://openreview.net/pdf?id=BZ5a1r-kVsf (Accessed: 21 June 2026).

Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y., dos Santos Costa, A., Fazel-Zarandi, M., Sercu, T., Candido, S. and Rives, A. (2022) 'Evolutionary-scale prediction of atomic level protein structure with a language model', bioRxiv preprint. doi:10.1101/2022.07.20.500902.

Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G. and Sutskever, I. (2021) 'Learning transferable visual models from natural language supervision', in Proceedings of the 38th International Conference on Machine Learning (ICML 2021), pp. 8748–8763. Available at: https://arxiv.org/abs/2103.00020 (Accessed: 19 June 2026).

Rives, A., Meier, J., Sercu, T., Goyal, S., Lin, Z., Liu, J., Guo, D., Ott, M., Zitnick, C.L., Ma, J. and Fergus, R. (2021) 'Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences', Proceedings of the National Academy of Sciences, 118(15), e2016239118. doi:10.1073/pnas.2016239118.

Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł. and Polosukhin, I. (2017) 'Attention is all you need', in Advances in Neural Information Processing Systems, 30. Available at: https://arxiv.org/abs/1706.03762 (Accessed: 18 June 2026).

Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., Chi, E.H., Hashimoto, T., Vinyals, O., Liang, P., Dean, J. and Fedus, W. (2022) 'Emergent abilities of large language models', Transactions on Machine Learning Research. Available at: https://arxiv.org/abs/2206.07682 (Accessed: 20 June 2026).

Xie, S., Santurkar, S., Ma, T. and Liang, P. (2023) 'Data selection for language models via importance resampling', in Advances in Neural Information Processing Systems, 36. Available at: https://arxiv.org/abs/2302.03169 (Accessed: 20 June 2026).