Optimizing ML Training with Metagradient Descent

A major challenge in training large-scale machine learning models is configuring the training process to maximize model performance, i.e., finding the best training setup from a vast design space. In this work, we unlock a gradient-based approach to this problem. We first introduce an algorithm for efficiently calculating metagradients -- gradients through model training -- at scale. We then introduce a "smooth model training" framework that enables effective optimization using metagradients. With metagradient descent (MGD), we greatly improve on existing dataset selection methods, outperform accuracy-degrading data poisoning attacks by an order of magnitude, and automatically find competitive learning rate schedules.

Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation

Mitigating reward hacking--where AI systems misbehave due to flaws or misspecifications in their learning objectives--remains a key challenge in constructing capable and aligned models. We show that we can monitor a frontier reasoning model, such as OpenAI o3-mini, for reward hacking in agentic coding environments by using another LLM that observes the model's chain-of-thought (CoT) reasoning. CoT monitoring can be far more effective than monitoring agent actions and outputs alone, and we further found that a LLM weaker than o3-mini, namely GPT-4o, can effectively monitor a stronger model. Because CoT monitors can be effective at detecting exploits, it is natural to ask whether those exploits can be suppressed by incorporating a CoT monitor directly into the agent's training objective. While we show that integrating CoT monitors into the reinforcement learning reward can indeed produce more capable and more aligned agents in the low optimization regime, we find that with too much optimization, agents learn obfuscated reward hacking, hiding their intent within the CoT while still exhibiting a significant rate of reward hacking. Because it is difficult to tell when CoTs have become obfuscated, it may be necessary to pay a monitorability tax by not applying strong optimization pressures directly to the chain-of-thought, ensuring that CoTs remain monitorable and useful for detecting misaligned behavior.

Do Large Language Model Benchmarks Test Reliability?

When deploying large language models (LLMs), it is important to ensure that these models are not only capable, but also reliable. Many benchmarks have been created to track LLMs' growing capabilities, however there has been no similar focus on measuring their reliability. To understand the potential ramifications of this gap, we investigate how well current benchmarks quantify model reliability. We find that pervasive label errors can compromise these evaluations, obscuring lingering model failures and hiding unreliable behavior. Motivated by this gap in the evaluation of reliability, we then propose the concept of so-called platinum benchmarks, i.e., benchmarks carefully curated to minimize label errors and ambiguity. As a first attempt at constructing such benchmarks, we revise examples from fifteen existing popular benchmarks. We evaluate a wide range of models on these platinum benchmarks and find that, indeed, frontier LLMs still exhibit failures on simple tasks such as elementary-level math word problems. Analyzing these failures further reveals previously unidentified patterns of problems on which frontier models consistently struggle. We provide code at https://github.com/MadryLab/platinum-benchmarks

OpenAI o1 System Card

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.

Attribute-to-Delete: Machine Unlearning via Datamodel Matching

Machine unlearning -- efficiently removing the effect of a small "forget set" of training data on a pre-trained machine learning model -- has recently attracted significant research interest. Despite this interest, however, recent work shows that existing machine unlearning techniques do not hold up to thorough evaluation in non-convex settings. In this work, we introduce a new machine unlearning technique that exhibits strong empirical performance even in such challenging settings. Our starting point is the perspective that the goal of unlearning is to produce a model whose outputs are statistically indistinguishable from those of a model re-trained on all but the forget set. This perspective naturally suggests a reduction from the unlearning problem to that of data attribution, where the goal is to predict the effect of changing the training set on a model's outputs. Thus motivated, we propose the following meta-algorithm, which we call Datamodel Matching (DMM): given a trained model, we (a) use data attribution to predict the output of the model if it were re-trained on all but the forget set points; then (b) fine-tune the pre-trained model to match these predicted outputs. In a simple convex setting, we show how this approach provably outperforms a variety of iterative unlearning algorithms. Empirically, we use a combination of existing evaluations and a new metric based on the KL-divergence to show that even in non-convex settings, DMM achieves strong unlearning performance relative to existing algorithms. An added benefit of DMM is that it is a meta-algorithm, in the sense that future advances in data attribution translate directly into better unlearning algorithms, pointing to a clear direction for future progress in unlearning.

ContextCite: Attributing Model Generation to Context

How do language models use information provided as context when generating a response? Can we infer whether a particular generated statement is actually grounded in the context, a misinterpretation, or fabricated? To help answer these questions, we introduce the problem of context attribution: pinpointing the parts of the context (if any) that led a model to generate a particular statement. We then present ContextCite, a simple and scalable method for context attribution that can be applied on top of any existing language model. Finally, we showcase the utility of ContextCite through three applications: (1) helping verify generated statements (2) improving response quality by pruning the context and (3) detecting poisoning attacks. We provide code for ContextCite at https://github.com/MadryLab/context-cite.

Data Debiasing with Datamodels (D3M): Improving Subgroup Robustness via Data Selection

Machine learning models can fail on subgroups that are underrepresented during training. While techniques such as dataset balancing can improve performance on underperforming groups, they require access to training group annotations and can end up removing large portions of the dataset. In this paper, we introduce Data Debiasing with Datamodels (D3M), a debiasing approach which isolates and removes specific training examples that drive the model's failures on minority groups. Our approach enables us to efficiently train debiased classifiers while removing only a small number of examples, and does not require training group annotations or additional hyperparameter tuning.

Measuring Strategization in Recommendation: Users Adapt Their Behavior to Shape Future Content

Most modern recommendation algorithms are data-driven: they generate personalized recommendations by observing users' past behaviors. A common assumption in recommendation is that how a user interacts with a piece of content (e.g., whether they choose to "like" it) is a reflection of the content, but not of the algorithm that generated it. Although this assumption is convenient, it fails to capture user strategization: that users may attempt to shape their future recommendations by adapting their behavior to the recommendation algorithm. In this work, we test for user strategization by conducting a lab experiment and survey. To capture strategization, we adopt a model in which strategic users select their engagement behavior based not only on the content, but also on how their behavior affects downstream recommendations. Using a custom music player that we built, we study how users respond to different information about their recommendation algorithm as well as to different incentives about how their actions affect downstream outcomes. We find strong evidence of strategization across outcome metrics, including participants' dwell time and use of "likes." For example, participants who are told that the algorithm mainly pays attention to "likes" and "dislikes" use those functions 1.9x more than participants told that the algorithm mainly pays attention to dwell time. A close analysis of participant behavior (e.g., in response to our incentive conditions) rules out experimenter demand as the main driver of these trends. Further, in our post-experiment survey, nearly half of participants self-report strategizing "in the wild," with some stating that they ignore content they actually like to avoid over-recommendation of that content in the future. Together, our findings suggest that user strategization is common and that platforms cannot ignore the effect of their algorithms on user behavior.

Decomposing and Editing Predictions by Modeling Model Computation

How does the internal computation of a machine learning model transform inputs into predictions? In this paper, we introduce a task called component modeling that aims to address this question. The goal of component modeling is to decompose an ML model's prediction in terms of its components -- simple functions (e.g., convolution filters, attention heads) that are the "building blocks" of model computation. We focus on a special case of this task, component attribution, where the goal is to estimate the counterfactual impact of individual components on a given prediction. We then present COAR, a scalable algorithm for estimating component attributions; we demonstrate its effectiveness across models, datasets, and modalities. Finally, we show that component attributions estimated with COAR directly enable model editing across five tasks, namely: fixing model errors, ``forgetting'' specific classes, boosting subpopulation robustness, localizing backdoor attacks, and improving robustness to typographic attacks. We provide code for COAR at https://github.com/MadryLab/modelcomponents .

Ask Your Distribution Shift if Pre-Training is Right for You

Pre-training is a widely used approach to develop models that are robust to distribution shifts. However, in practice, its effectiveness varies: fine-tuning a pre-trained model improves robustness significantly in some cases but not at all in others (compared to training from scratch). In this work, we seek to characterize the failure modes that pre-training can and cannot address. In particular, we focus on two possible failure modes of models under distribution shift: poor extrapolation (e.g., they cannot generalize to a different domain) and biases in the training data (e.g., they rely on spurious features). Our study suggests that, as a rule of thumb, pre-training can help mitigate poor extrapolation but not dataset biases. After providing theoretical motivation and empirical evidence for this finding, we explore two of its implications for developing robust models: (1) pre-training and interventions designed to prevent exploiting biases have complementary robustness benefits, and (2) fine-tuning on a (very) small, non-diverse but de-biased dataset can result in significantly more robust models than fine-tuning on a large and diverse but biased dataset. Code is available at https://github.com/MadryLab/pretraining-distribution-shift-robustness.