Matt Jones

Aaron Jaech, Adam Kalai, Adam Lerer et al. · openai

The o1 model series is trained with large-scale reinforcement learning to reason using chain of thought. These advanced reasoning capabilities provide new avenues for improving the safety and robustness of our models. In particular, our models can reason about our safety policies in context when responding to potentially unsafe prompts, through deliberative alignment. This leads to state-of-the-art performance on certain benchmarks for risks such as generating illicit advice, choosing stereotyped responses, and succumbing to known jailbreaks. Training models to incorporate a chain of thought before answering has the potential to unlock substantial benefits, while also increasing potential risks that stem from heightened intelligence. Our results underscore the need for building robust alignment methods, extensively stress-testing their efficacy, and maintaining meticulous risk management protocols. This report outlines the safety work carried out for the OpenAI o1 and OpenAI o1-mini models, including safety evaluations, external red teaming, and Preparedness Framework evaluations.

Frieda Born, Tom Neuhäuser, Lukas Muttenthaler et al. · deepmind, stanford

Modern machine learning models typically represent inputs as fixed points in a high-dimensional embedding space. While this approach has been proven powerful for a wide range of downstream tasks, it fundamentally differs from the way humans process information. Because humans are constantly adapting to their environment, they represent objects and their relationships in a highly context-sensitive manner. To address this gap, we propose a method for context-sensitive similarity computation from neural network embeddings, applied to modeling a triplet odd-one-out task with an anchor image serving as simultaneous context. Modeling context enables us to achieve up to a 15% improvement in odd-one-out accuracy over a context-insensitive model. We find that this improvement is consistent across both original and "human-aligned" vision foundation models.

Alexandre Galashov, Matt Jones, Rosemary Ke et al. · deepmind

We explore a class of supervised training objectives that allow a language model to dynamically and autonomously scale the number of compute steps used for each input token. For any token, the model can request additional compute steps by emitting a <don't know> output. If the model is granted a delay, a specialized <pause> token is inserted at the next input step, providing the model with additional compute resources to generate an output. The model can request multiple pauses. To train the model to use <don't know> outputs judiciously and to calibrate its uncertainty, we frame the selection of each output token as a sequential-decision problem with a time cost. We refer to the class of methods as $\textit{Catch Your Breath}$ losses and we study three methods in this class: CYB-AP frames the model's task as anytime prediction, where an output may be required at any step and accuracy is discounted over time; CYB-VA is a variational approach that aims to maximize prediction accuracy subject to a specified distribution over stopping times; and CYB-DP imposes a penalty based on a computational budget. Through fine-tuning experiments, we identify the best performing loss variant. The CYB model needs only one third as much training data as the baseline (no pause) model needs to achieve the same performance, and half as much data as a model with pauses and a cross-entropy loss. We find that the CYB model requests additional steps when doing so improves accuracy, and the model adapts its processing time to token-level complexity and context. For example, it often pauses after plural nouns like $\textit{patients}$ and $\textit{challenges}$ but never pauses after the first token of contracted words like $\textit{wasn}$ and $\textit{didn}$, and it shows high variability for ambiguous tokens like $\textit{won}$, which could function as either a verb or part of a contraction.

Gerardo Duran-Martin, Matias Altamirano, Alexander Y. Shestopaloff et al.

We derive a novel, provably robust, and closed-form Bayesian update rule for online filtering in state-space models in the presence of outliers and misspecified measurement models. Our method combines generalised Bayesian inference with filtering methods such as the extended and ensemble Kalman filter. We use the former to show robustness and the latter to ensure computational efficiency in the case of nonlinear models. Our method matches or outperforms other robust filtering methods (such as those based on variational Bayes) at a much lower computational cost. We show this empirically on a range of filtering problems with outlier measurements, such as object tracking, state estimation in high-dimensional chaotic systems, and online learning of neural networks.

Yanlai Yang, Matt Jones, Michael C. Mozer et al.

We explore the training dynamics of neural networks in a structured non-IID setting where documents are presented cyclically in a fixed, repeated sequence. Typically, networks suffer from catastrophic interference when training on a sequence of documents; however, we discover a curious and remarkable property of LLMs finetuned sequentially in this setting: they exhibit anticipatory behavior, recovering from the forgetting on documents before encountering them again. This behavior occurs even though the documents are never presented in context together. The behavior emerges and becomes more robust as the architecture scales up its number of parameters. Through comprehensive experiments and visualizations, we demonstrate a new mechanism by which over-parametrized neural networks can recover from catastrophic interference and uncover new insights into training over-parameterized networks in cyclically structured environments.

Manfred K. Warmuth, Wojciech Kotłowski, Matt Jones et al.

It is well known that the class of rotation invariant algorithms are suboptimal even for learning sparse linear problems when the number of examples is below the "dimension" of the problem. This class includes any gradient descent trained neural net with a fully-connected input layer (initialized with a rotationally symmetric distribution). The simplest sparse problem is learning a single feature out of $d$ features. In that case the classification error or regression loss grows with $1-k/n$ where $k$ is the number of examples seen. These lower bounds become vacuous when the number of examples $k$ reaches the dimension $d$. We show that when noise is added to this sparse linear problem, rotation invariant algorithms are still suboptimal after seeing $d$ or more examples. We prove this via a lower bound for the Bayes optimal algorithm on a rotationally symmetrized problem. We then prove much lower upper bounds on the same problem for simple non-rotation invariant algorithms. Finally we analyze the gradient flow trajectories of many standard optimization algorithms in some simple cases and show how they veer toward or away from the sparse targets. We believe that our trajectory categorization will be useful in designing algorithms that can exploit sparse targets and our method for proving lower bounds will be crucial for analyzing other families of algorithms that admit different classes of invariances.

Peter G. Chang, Gerardo Durán-Martín, Alexander Y Shestopaloff et al.

We propose an efficient online approximate Bayesian inference algorithm for estimating the parameters of a nonlinear function from a potentially non-stationary data stream. The method is based on the extended Kalman filter (EKF), but uses a novel low-rank plus diagonal decomposition of the posterior precision matrix, which gives a cost per step which is linear in the number of model parameters. In contrast to methods based on stochastic variational inference, our method is fully deterministic, and does not require step-size tuning. We show experimentally that this results in much faster (more sample efficient) learning, which results in more rapid adaptation to changing distributions, and faster accumulation of reward when used as part of a contextual bandit algorithm.

Sebastian Bobadilla-Suarez, Matt Jones, Bradley C. Love

Induction benefits from useful priors. Penalized regression approaches, like ridge regression, shrink weights toward zero but zero association is usually not a sensible prior. Inspired by simple and robust decision heuristics humans use, we constructed non-zero priors for penalized regression models that provide robust and interpretable solutions across several tasks. Our approach enables estimates from a constrained model to serve as a prior for a more general model, yielding a principled way to interpolate between models of differing complexity. We successfully applied this approach to a number of decision and classification problems, as well as analyzing simulated brain imaging data. Models with robust priors had excellent worst-case performance. Solutions followed from the form of the heuristic that was used to derive the prior. These new algorithms can serve applications in data analysis and machine learning, as well as help in understanding how people transition from novice to expert performance.

James M. Foster, Matt Jones

Research in analogical reasoning suggests that higher-order cognitive functions such as abstract reasoning, far transfer, and creativity are founded on recognizing structural similarities among relational systems. Here we integrate theories of analogy with the computational framework of reinforcement learning (RL). We propose a psychology theory that is a computational synergy between analogy and RL, in which analogical comparison provides the RL learning algorithm with a measure of relational similarity, and RL provides feedback signals that can drive analogical learning. Simulation results support the power of this approach.