Frontier Language Models are not Robust to Adversarial Arithmetic, or "What do I need to say so you agree 2+2=5?

We introduce and study the problem of adversarial arithmetic, which provides a simple yet challenging testbed for language model alignment. This problem is comprised of arithmetic questions posed in natural language, with an arbitrary adversarial string inserted before the question is complete. Even in the simple setting of 1-digit addition problems, it is easy to find adversarial prompts that make all tested models (including PaLM2, GPT4, Claude2) misbehave, and even to steer models to a particular wrong answer. We additionally provide a simple algorithm for finding successful attacks by querying those same models, which we name "prompt inversion rejection sampling" (PIRS). We finally show that models can be partially hardened against these attacks via reinforcement learning and via agentic constitutional loops. However, we were not able to make a language model fully robust against adversarial arithmetic attacks.

Improving Large Language Model Fine-tuning for Solving Math Problems

Despite their success in many natural language tasks, solving math problems remains a significant challenge for large language models (LLMs). A large gap exists between LLMs' pass-at-one and pass-at-N performance in solving math problems, suggesting LLMs might be close to finding correct solutions, motivating our exploration of fine-tuning methods to unlock LLMs' performance. Using the challenging MATH dataset, we investigate three fine-tuning strategies: (1) solution fine-tuning, where we fine-tune to generate a detailed solution for a given math problem; (2) solution-cluster re-ranking, where the LLM is fine-tuned as a solution verifier/evaluator to choose among generated candidate solution clusters; (3) multi-task sequential fine-tuning, which integrates both solution generation and evaluation tasks together efficiently to enhance the LLM performance. With these methods, we present a thorough empirical study on a series of PaLM 2 models and find: (1) The quality and style of the step-by-step solutions used for fine-tuning can make a significant impact on the model performance; (2) While solution re-ranking and majority voting are both effective for improving the model performance when used separately, they can also be used together for an even greater performance boost; (3) Multi-task fine-tuning that sequentially separates the solution generation and evaluation tasks can offer improved performance compared with the solution fine-tuning baseline. Guided by these insights, we design a fine-tuning recipe that yields approximately 58.8% accuracy on the MATH dataset with fine-tuned PaLM 2-L models, an 11.2% accuracy improvement over the few-shot performance of pre-trained PaLM 2-L model with majority voting.

Transformer-Based Learned Optimization

In this paper, we propose a new approach to learned optimization. As common in the literature, we represent the computation of the update step of the optimizer with a neural network. The parameters of the optimizer are then learned on a set of training optimization tasks, in order to perform minimisation efficiently. Our main innovation is to propose a new neural network architecture for the learned optimizer inspired by the classic BFGS algorithm. As in BFGS, we estimate a preconditioning matrix as a sum of rank-one updates but use a transformer-based neural network to predict these updates jointly with the step length and direction. In contrast to several recent learned optimization approaches, our formulation allows for conditioning across different dimensions of the parameter space of the target problem while remaining applicable to optimization tasks of variable dimensionality without retraining. We demonstrate the advantages of our approach on a benchmark composed of objective functions traditionally used for evaluation of optimization algorithms, as well as on the real world-task of physics-based reconstruction of articulated 3D human motion.

VeLO: Training Versatile Learned Optimizers by Scaling Up

While deep learning models have replaced hand-designed features across many domains, these models are still trained with hand-designed optimizers. In this work, we leverage the same scaling approach behind the success of deep learning to learn versatile optimizers. We train an optimizer for deep learning which is itself a small neural network that ingests gradients and outputs parameter updates. Meta-trained with approximately four thousand TPU-months of compute on a wide variety of optimization tasks, our optimizer not only exhibits compelling performance, but optimizes in interesting and unexpected ways. It requires no hyperparameter tuning, instead automatically adapting to the specifics of the problem being optimized. We open source our learned optimizer, meta-training code, the associated train and test data, and an extensive optimizer benchmark suite with baselines at velo-code.github.io.

Allowing Safe Contact in Robotic Goal-Reaching: Planning and Tracking in Operational and Null Spaces

In recent years, impressive results have been achieved in robotic manipulation. While many efforts focus on generating collision-free reference signals, few allow safe contact between the robot bodies and the environment. However, in human's daily manipulation, contact between arms and obstacles is prevalent and even necessary. This paper investigates the benefit of allowing safe contact during robotic manipulation and advocates generating and tracking compliance reference signals in both operational and null spaces. In addition, to optimize the collision-allowed trajectories, we present a hybrid solver that integrates sampling- and gradient-based approaches. We evaluate the proposed method on a goal-reaching task in five simulated and real-world environments with different collisional conditions. We show that allowing safe contact improves goal-reaching efficiency and provides feasible solutions in highly collisional scenarios where collision-free constraints cannot be enforced. Moreover, we demonstrate that planning in null space, in addition to operational space, improves trajectory safety.

Practical tradeoffs between memory, compute, and performance in learned optimizers

Optimization plays a costly and crucial role in developing machine learning systems. In learned optimizers, the few hyperparameters of commonly used hand-designed optimizers, e.g. Adam or SGD, are replaced with flexible parametric functions. The parameters of these functions are then optimized so that the resulting learned optimizer minimizes a target loss on a chosen class of models. Learned optimizers can both reduce the number of required training steps and improve the final test loss. However, they can be expensive to train, and once trained can be expensive to use due to computational and memory overhead for the optimizer itself. In this work, we identify and quantify the design features governing the memory, compute, and performance trade-offs for many learned and hand-designed optimizers. We further leverage our analysis to construct a learned optimizer that is both faster and more memory efficient than previous work.

Gradients are Not All You Need

Differentiable programming techniques are widely used in the community and are responsible for the machine learning renaissance of the past several decades. While these methods are powerful, they have limits. In this short report, we discuss a common chaos based failure mode which appears in a variety of differentiable circumstances, ranging from recurrent neural networks and numerical physics simulation to training learned optimizers. We trace this failure to the spectrum of the Jacobian of the system under study, and provide criteria for when a practitioner might expect this failure to spoil their differentiation based optimization algorithms.

Brax -- A Differentiable Physics Engine for Large Scale Rigid Body Simulation

We present Brax, an open source library for rigid body simulation with a focus on performance and parallelism on accelerators, written in JAX. We present results on a suite of tasks inspired by the existing reinforcement learning literature, but remade in our engine. Additionally, we provide reimplementations of PPO, SAC, ES, and direct policy optimization in JAX that compile alongside our environments, allowing the learning algorithm and the environment processing to occur on the same device, and to scale seamlessly on accelerators. Finally, we include notebooks that facilitate training of performant policies on common OpenAI Gym MuJoCo-like tasks in minutes.

Training Learned Optimizers with Randomly Initialized Learned Optimizers

Learned optimizers are increasingly effective, with performance exceeding that of hand designed optimizers such as Adam~\citep{kingma2014adam} on specific tasks \citep{metz2019understanding}. Despite the potential gains available, in current work the meta-training (or `outer-training') of the learned optimizer is performed by a hand-designed optimizer, or by an optimizer trained by a hand-designed optimizer \citep{metz2020tasks}. We show that a population of randomly initialized learned optimizers can be used to train themselves from scratch in an online fashion, without resorting to a hand designed optimizer in any part of the process. A form of population based training is used to orchestrate this self-training. Although the randomly initialized optimizers initially make slow progress, as they improve they experience a positive feedback loop, and become rapidly more effective at training themselves. We believe feedback loops of this type, where an optimizer improves itself, will be important and powerful in the future of machine learning. These methods not only provide a path towards increased performance, but more importantly relieve research and engineering effort.

Tasks, stability, architecture, and compute: Training more effective learned optimizers, and using them to train themselves

Much as replacing hand-designed features with learned functions has revolutionized how we solve perceptual tasks, we believe learned algorithms will transform how we train models. In this work we focus on general-purpose learned optimizers capable of training a wide variety of problems with no user-specified hyperparameters. We introduce a new, neural network parameterized, hierarchical optimizer with access to additional features such as validation loss to enable automatic regularization. Most learned optimizers have been trained on only a single task, or a small number of tasks. We train our optimizers on thousands of tasks, making use of orders of magnitude more compute, resulting in optimizers that generalize better to unseen tasks. The learned optimizers not only perform well, but learn behaviors that are distinct from existing first order optimizers. For instance, they generate update steps that have implicit regularization and adapt as the problem hyperparameters (e.g. batch size) or architecture (e.g. neural network width) change. Finally, these learned optimizers show evidence of being useful for out of distribution tasks such as training themselves from scratch.