Language models (LMs) can be directed to perform user- and context-dependent tasks by using labeled examples or natural language prompts. But selecting examples or writing prompts can be challenging---especially in tasks that require users to precisely articulate nebulous preferences or reason about complex edge cases. For such tasks, we introduce **Generative Active Task Elicitation (GATE)**, a method for using *LMs themselves* to guide the task specification process. GATE is a learning framework in which models elicit and infer human preferences through free-form, language-based interaction with users. We identify prototypical challenges that users face when specifying preferences, and design three preference modeling tasks to study these challenges: content recommendation, moral reasoning, and email validation. In preregistered experiments, we show that LMs that learn to perform these tasks using GATE (by interactively querying users with open-ended questions) obtain preference specifications that are more informative than user-written prompts or examples. GATE matches existing task specification methods in the moral reasoning task, and significantly outperforms them in the content recommendation and email validation tasks. Users additionally report that interactive task elicitation requires less effort than prompting or example labeling and surfaces considerations that they did not anticipate on their own. Our findings suggest that LM-driven elicitation can be a powerful tool for aligning models to complex human preferences and values.

Understanding neural networks is challenging in part because of the dense, continuous nature of their hidden states. We explore whether we can train neural networks to have hidden states that are sparse, discrete, and more interpretable by quantizing their continuous features into what we call codebook features. Codebook features are produced by finetuning neural networks with vector quantization bottlenecks at each layer, producing a network whose hidden features are the sum of a small number of discrete vector codes chosen from a larger codebook. Surprisingly, we find that neural networks can operate under this extreme bottleneck with only modest degradation in performance. In addition, we can control a model’s behavior by finding codes that activate on a desired behavior, then activating those same codes during generation. We first validate codebook features on a finite state machine dataset with far more hidden states than neurons. In this setting, our approach overcomes the superposition problem by assigning states to distinct codes, and we find that we can make the neural network behave as if it is in a different state by activating the code for that state. We then train Transformer language models with up to 410M parameters on two natural language datasets. We identify codes in these models representing diverse, disentangled concepts (ranging from negative emotions to months of the year) and find that we can guide the model to generate different topics and pronoun genders by activating these codes during inference. Overall, codebook features appear to be a promising unit of analysis and control for neural networks and interpretability. Our codebase and models are open-sourced at this URL.

We investigate a family of simple long-context attacks on large language models: prompting with hundreds of demonstrations of undesirable behavior. This attack is newly feasible with the larger context windows recently deployed by language model providers like Google DeepMind, OpenAI and Anthropic. We find that in diverse, realistic circumstances, the effectiveness of this attack follows a power law, up to hundreds of shots. We demonstrate the success of this attack on the most widely used state-of-the-art closed-weight models, and across various tasks. Our results suggest very long contexts present a rich new attack surface for LLMs.

When we transfer a pretrained language model to a new language, there are many axes of variation that change at once.To disentangle the impact of different factors like syntactic similarity and vocabulary similarity, we propose a set of controlled transfer studies: we systematically transform the language of the GLUE benchmark, altering one axis of crosslingual variation at a time, and then measure the resulting drops in a pretrained model's downstream performance.We find that models can largely recover from syntactic-style shifts, but cannot recover from vocabulary misalignment and embedding matrix re-initialization, even with continued pretraining on 15 million tokens.Moreover, good-quality tokenizers in the transfer language do not make vocabulary alignment easier.Our experiments provide insights into the factors of cross-lingual transfer that researchers should most focus on when designing language transfer scenarios.

Analysis of compressible turbulent flows is essential for applications related to propulsion, energy generation, and the environment. Here, we present BLASTNet 2.0, a 2.2 TB network-of-datasets containing 744 full-domain samples from 34 high-fidelity direct numerical simulations, which addresses the current limited availability of 3D high-fidelity reacting and non-reacting compressible turbulent flow simulation data. With this data, we benchmark a total of 49 variations of five deep learning approaches for 3D super-resolution - which can be applied for improving scientific imaging, simulations, turbulence models, as well as in computer vision applications. We perform neural scaling analysis on these models to examine the performance of different machine learning (ML) approaches, including two scientific ML techniques. We demonstrate that (i) predictive performance can scale with model size and cost, (ii) architecture matters significantly, especially for smaller models, and (iii) the benefits of physics-based losses can persist with increasing model size. The outcomes of this benchmark study are anticipated to offer insights that can aid the design of 3D super-resolution models, especially for turbulence models, while this data is expected to foster ML methods for a broad range of flow physics applications. This data is publicly available with download links and browsing tools consolidated at https://blastnet.github.io.

What role do augmentations play in contrastive learning? Recent work suggests that good augmentations are label-preserving with respect to a specific downstream task. We complicate this picture by showing that label-destroying augmentations can be useful in the foundation model setting, where the goal is to learn diverse, general-purpose representations for multiple downstream tasks. We perform contrastive learning experiments on a range of image and audio datasets with multiple downstream tasks (e.g. for digits superimposed on photographs, predicting the class of one vs. the other). We find that Viewmaker Networks, a recently proposed model for learning augmentations for contrastive learning, produce label-destroying augmentations that stochastically destroy features needed for different downstream tasks. These augmentations are interpretable (e.g. altering shapes, digits, or letters added to images) and surprisingly often result in better performance compared to expert-designed augmentations, despite not preserving label information. To support our empirical results, we theoretically analyze a simple contrastive learning setting with a linear model. In this setting, label-destroying augmentations are crucial for preventing one set of features from suppressing the learning of features useful for another downstream task. Our results highlight the need for analyzing the interaction between multiple downstream tasks when trying to explain the success of foundation models.

Universal self-supervised (SSL) algorithms hold enormous promise for making machine learning accessible to high-impact domains such as protein biology, manufacturing, and genomics. We present DABS 2.0: a set of improved datasets and algorithms for advancing research on universal SSL. We extend the recently-introduced DABS benchmark with the addition of five real-world science and engineering domains: protein biology, bacterial genomics, multispectral satellite imagery, semiconductor wafers, and particle physics, bringing the total number of domains in the benchmark to twelve. We also propose a new universal SSL algorithm, Capri, and a generalized version of masked autoencoding, and apply both on all twelve domains---the most wide-ranging exploration of SSL yet. We find that multiple algorithms show gains across domains, outperforming previous baselines. In addition, we demonstrate the usefulness of DABS for scientific study of SSL by investigating the optimal corruption rate for each algorithm, showing that the best setting varies based on the domain. Code will be released at http://github.com/alextamkin/dabs}{http://github.com/alextamkin/dabs

Models can fail in unpredictable ways during deployment due to task ambiguity, when multiple behaviors are consistent with the provided training data. An example is an object classifier trained on red squares and blue circles: when encountering blue squares, the intended behavior is undefined. We investigate whether pretrained models are better active learners, capable of disambiguating between the possible tasks a user may be trying to specify. Intriguingly, we find that better active learning is an emergent property of the pretraining process: pretrained models require up to 5 times fewer labels when using uncertainty-based active learning, while non-pretrained models see no or even negative benefit. We find these gains come from an ability to select examples with attributes that disambiguate the intended behavior, such as rare product categories or atypical backgrounds. These attributes are far more linearly separable in pretrained model's representation spaces vs non-pretrained models, suggesting a possible mechanism for this behavior.

While maximizing expected return is the goal in most reinforcement learning approaches, risk-sensitive objectives such as conditional value at risk (CVaR) are more suitable for many high-stakes applications. However, relatively little is known about how to explore to quickly learn policies with good CVaR. In this paper, we present the first algorithm for sample-efficient learning of CVaR-optimal policies in Markov decision processes based on the optimism in the face of uncertainty principle. This method relies on a novel optimistic version of the distributional Bellman operator that moves probability mass from the lower to the upper tail of the return distribution. We prove asymptotic convergence and optimism of this operator for the tabular policy evaluation case. We further demonstrate that our algorithm finds CVaR-optimal policies substantially faster than existing baselines in several simulated environments with discrete and continuous state spaces.

Language exhibits structure at a wide range of scales, from subwords to words, sentences, paragraphs, and documents. We propose building models that isolate scale-specific information in deep representations, and develop methods for encouraging models during training to learn more about particular scales of interest. Our method for creating scale-specific neurons in deep NLP models constrains how the activation of a neuron can change across the tokens of an input by interpreting those activations as a digital signal and filtering out parts of its frequency spectrum. This technique enables us to extract scale-specific information from BERT representations: by filtering out different frequencies we can produce new representations that perform well on part of speech tagging (word-level), dialog speech acts classification (utterance-level), or topic classification (document-level), while performing poorly on the other tasks. We also present a prism layer for use during training, which constrains different neurons of a BERT model to different parts of the frequency spectrum. Our proposed BERT + Prism model is better able to predict masked tokens using long-range context, and produces individual multiscale representations that perform with comparable or improved performance across all three tasks. Our methods are general and readily applicable to other domains besides language, such as images, audio, and video.

Drones are becoming ubiquitous and offer support to people in various tasks, such as photography, in increasingly interactive social contexts. We introduce drone.io, a projected body-centric graphical user interface for human-drone interaction. Using two simple gestures, users can interact with a drone in a natural manner. drone.io is the first human-drone graphical user interface embedded on a drone to provide both input and output capabilities. This paper describes the design process of drone.io. We present a proof of concept, drone-based implementation, as well as a fully functional prototype for a drone tour-guide scenario. We report drone.io's evaluation in three user studies (N=27) and show that people were able to use the interface with little prior training. We contribute to the field of human-robot interaction and the growing field of human-drone interaction.