Oluwasanmi Koyejo

Yu Sun, Xinhao Li, Karan Dalal et al.

Self-attention performs well in long context but has quadratic complexity. Existing RNN layers have linear complexity, but their performance in long context is limited by the expressive power of their hidden states. We present a practical framework for instantiating sequence modeling layers with linear complexity and expressive hidden states. The key idea is to make the hidden state a machine learning model itself, and the update rule a step of self-supervised learning. Since the hidden state is updated by training even on test sequences, our layers are called Test-Time Training (TTT) layers. We consider two instantiations: TTT-Linear and TTT-MLP, whose hidden state is a linear model and a two-layer MLP respectively. We evaluate our instantiations at the scale of 125M to 1.3B parameters, comparing with a strong Transformer and Mamba, a modern RNN. Similar to Transformer, TTT-Linear and TTT-MLP can keep reducing perplexity by conditioning on more tokens, while Mamba cannot after 16k context. TTT-MLP still faces challenges in memory I/O, but shows larger potential in long context, pointing to a promising direction for future research.

Safinah Ali, Sohini Upadhyay, Gaurush Hiranandani et al. · harvard

Metric Elicitation (ME) is a framework for eliciting classification metrics that better align with implicit user preferences based on the task and context. The existing ME strategy so far is based on the assumption that users can most easily provide preference feedback over classifier statistics such as confusion matrices. This work examines ME, by providing a first ever implementation of the ME strategy. Specifically, we create a web-based ME interface and conduct a user study that elicits users' preferred metrics in a binary classification setting. We discuss the study findings and present guidelines for future research in this direction.

Peiye Zhuang, Liqian Ma, Oluwasanmi Koyejo et al.

Recent work on 3D-aware image synthesis has achieved compelling results using advances in neural rendering. However, 3D-aware synthesis of face dynamics hasn't received much attention. Here, we study how to explicitly control generative model synthesis of face dynamics exhibiting non-rigid motion (e.g., facial expression change), while simultaneously ensuring 3D-awareness. For this we propose a Controllable Radiance Field (CoRF): 1) Motion control is achieved by embedding motion features within the layered latent motion space of a style-based generator; 2) To ensure consistency of background, motion features and subject-specific attributes such as lighting, texture, shapes, albedo, and identity, a face parsing net, a head regressor and an identity encoder are incorporated. On head image/video data we show that CoRFs are 3D-aware while enabling editing of identity, viewing directions, and motion.

Rylan Schaeffer, Mikail Khona, Zachary Robertson et al.

Double descent is a surprising phenomenon in machine learning, in which as the number of model parameters grows relative to the number of data, test error drops as models grow ever larger into the highly overparameterized (data undersampled) regime. This drop in test error flies against classical learning theory on overfitting and has arguably underpinned the success of large models in machine learning. This non-monotonic behavior of test loss depends on the number of data, the dimensionality of the data and the number of model parameters. Here, we briefly describe double descent, then provide an explanation of why double descent occurs in an informal and approachable manner, requiring only familiarity with linear algebra and introductory probability. We provide visual intuition using polynomial regression, then mathematically analyze double descent with ordinary linear regression and identify three interpretable factors that, when simultaneously all present, together create double descent. We demonstrate that double descent occurs on real data when using ordinary linear regression, then demonstrate that double descent does not occur when any of the three factors are ablated. We use this understanding to shed light on recent observations in nonlinear models concerning superposition and double descent. Code is publicly available.

Yu Sun, Xinhao Li, Karan Dalal et al.

We reformulate the problem of supervised learning as learning to learn with two nested loops (i.e. learning problems). The inner loop learns on each individual instance with self-supervision before final prediction. The outer loop learns the self-supervised task used by the inner loop, such that its final prediction improves. Our inner loop turns out to be equivalent to linear attention when the inner-loop learner is only a linear model, and to self-attention when it is a kernel estimator. For practical comparison with linear or self-attention layers, we replace each of them in a transformer with an inner loop, so our outer loop is equivalent to training the architecture. When each inner-loop learner is a neural network, our approach vastly outperforms transformers with linear attention on ImageNet from 224 x 224 raw pixels in both accuracy and FLOPs, while (regular) transformers cannot run.

Ibrahim Alabdulmohsin, Jessica Schrouff, Oluwasanmi Koyejo

We propose a novel reduction-to-binary (R2B) approach that enforces demographic parity for multiclass classification with non-binary sensitive attributes via a reduction to a sequence of binary debiasing tasks. We prove that R2B satisfies optimality and bias guarantees and demonstrate empirically that it can lead to an improvement over two baselines: (1) treating multiclass problems as multi-label by debiasing labels independently and (2) transforming the features instead of the labels. Surprisingly, we also demonstrate that independent label debiasing yields competitive results in most (but not all) settings. We validate these conclusions on synthetic and real-world datasets from social science, computer vision, and healthcare.

Enyi Jiang, Yibo Jacky Zhang, Sanmi Koyejo

Federated Domain Adaptation (FDA) describes the federated learning (FL) setting where source clients and a server work collaboratively to improve the performance of a target client where limited data is available. The domain shift between the source and target domains, coupled with limited data of the target client, makes FDA a challenging problem, e.g., common techniques such as federated averaging and fine-tuning fail due to domain shift and data scarcity. To theoretically understand the problem, we introduce new metrics that characterize the FDA setting and a theoretical framework with novel theorems for analyzing the performance of server aggregation rules. Further, we propose a novel lightweight aggregation rule, Federated Gradient Projection ($\texttt{FedGP}$), which significantly improves the target performance with domain shift and data scarcity. Moreover, our theory suggests an $\textit{auto-weighting scheme}$ that finds the optimal combinations of the source and target gradients. This scheme improves both $\texttt{FedGP}$ and a simpler heuristic aggregation rule. Extensive experiments verify the theoretical insights and illustrate the effectiveness of the proposed methods in practice.

Katherine Tsai, Boxin Zhao, Sanmi Koyejo et al.

Joint multimodal functional data acquisition, where functional data from multiple modes are measured simultaneously from the same subject, has emerged as an exciting modern approach enabled by recent engineering breakthroughs in the neurological and biological sciences. One prominent motivation to acquire such data is to enable new discoveries of the underlying connectivity by combining multimodal signals. Despite the scientific interest, there remains a gap in principled statistical methods for estimating the graph underlying multimodal functional data. To this end, we propose a new integrative framework that models the data generation process and identifies operators mapping from the observation space to the latent space. We then develop an estimator that simultaneously estimates the transformation operators and the latent graph. This estimator is based on the partial correlation operator, which we rigorously extend from the multivariate to the functional setting. Our procedure is provably efficient, with the estimator converging to a stationary point with quantifiable statistical error. Furthermore, we show recovery of the latent graph under mild conditions. Our work is applied to analyze simultaneously acquired multimodal brain imaging data where the graph indicates functional connectivity of the brain. We present simulation and empirical results that support the benefits of joint estimation.

Olawale Salaudeen, Oluwasanmi Koyejo

We propose a Target Conditioned Representation Independence (TCRI) objective for domain generalization. TCRI addresses the limitations of existing domain generalization methods due to incomplete constraints. Specifically, TCRI implements regularizers motivated by conditional independence constraints that are sufficient to strictly learn complete sets of invariant mechanisms, which we show are necessary and sufficient for domain generalization. Empirically, we show that TCRI is effective on both synthetic and real-world data. TCRI is competitive with baselines in average accuracy while outperforming them in worst-domain accuracy, indicating desired cross-domain stability.

Maohao Shen, Bowen Jiang, Jacky Yibo Zhang et al.

Active learning enables efficient model training by leveraging interactions between machine learning agents and human annotators. We study and propose a novel framework that formulates batch active learning from the sparse approximation's perspective. Our active learning method aims to find an informative subset from the unlabeled data pool such that the corresponding training loss function approximates its full data pool counterpart. We realize the framework as sparsity-constrained discontinuous optimization problems, which explicitly balance uncertainty and representation for large-scale applications and could be solved by greedy or proximal iterative hard thresholding algorithms. The proposed method can adapt to various settings, including both Bayesian and non-Bayesian neural networks. Numerical experiments show that our work achieves competitive performance across different settings with lower computational complexity.

Peiye Zhuang, Bliss Chapman, Ran Li et al.

In the experimental sciences, statistical power analyses are often used before data collection to determine the required sample size. However, traditional power analyses can be costly when data are difficult or expensive to collect. We propose synthetic power analyses; a framework for estimating statistical power at various sample sizes, and empirically explore the performance of synthetic power analysis for sample size selection in cognitive neuroscience experiments. To this end, brain imaging data is synthesized using an implicit generative model conditioned on observed cognitive processes. Further, we propose a simple procedure to modify the statistical tests which result in conservative statistics. Our empirical results suggest that synthetic power analysis could be a low-cost alternative to pilot data collection when the proposed experiments share cognitive processes with previously conducted experiments.

Pablo Robles-Granda, Katherine Tsai, Oluwasanmi Koyejo

Probabilistic generative models of graphs are important tools that enable representation and sampling. Many recent works have created probabilistic models of graphs that are capable of representing not only entity interactions but also their attributes. However, given a generative model of random attributed graph(s), the general conditions that establish goodness of fit are not clear a-priori. In this paper, we define goodness of fit in terms of the mean square contingency coefficient for random binary networks. For this statistic, we outline a procedure for assessing the quality of the structure of a learned attributed graph by ensuring that the discrepancy of the mean square contingency coefficient (constant, or random) is minimal with high probability. We apply these criteria to verify the representation capability of a probabilistic generative model for various popular types of graph models.

Zachary Robertson, Oluwasanmi Koyejo

Feedback Alignment (FA) methods are biologically inspired local learning rules for training neural networks with reduced communication between layers. While FA has potential applications in distributed and privacy-aware ML, limitations in multi-class classification and lack of theoretical understanding of the alignment mechanism have constrained its impact. This study introduces a unified framework elucidating the operational principles behind alignment in FA. Our key contributions include: (1) a novel conservation law linking changes in synaptic weights to implicit regularization that maintains alignment with the gradient, with support from experiments, (2) sufficient conditions for convergence based on the concept of alignment dominance, and (3) empirical analysis showing better alignment can enhance FA performance on complex multi-class tasks. Overall, these theoretical and practical advancements improve interpretability of bio-plausible learning rules and provide groundwork for developing enhanced FA algorithms.

Zachary Robertson, Oluwasanmi Koyejo

The growing demand for data and AI-generated digital goods, such as personalized written content and artwork, necessitates effective pricing and feedback mechanisms that account for uncertain utility and costly production. Motivated by these developments, this study presents a novel mechanism design addressing a general repeated-auction setting where the utility derived from a sold good is revealed post-sale. The mechanism's novelty lies in using pairwise comparisons for eliciting information from the bidder, arguably easier for humans than assigning a numerical value. Our mechanism chooses allocations using an epsilon-greedy strategy and relies on pairwise comparisons between realized utility from allocated goods and an arbitrary value, avoiding the learning-to-bid problem explored in previous work. We prove this mechanism to be asymptotically truthful, individually rational, and welfare and revenue maximizing. The mechanism's relevance is broad, applying to any setting with made-to-order goods of variable quality. Experimental results on multi-label toxicity annotation data, an example of negative utilities, highlight how our proposed mechanism could enhance social welfare in data auctions. Overall, our focus on human factors contributes to the development of more human-aware and efficient mechanism design.

Shivalika Singh, Yiyang Nan, Alex Wang et al.

Measuring progress is fundamental to the advancement of any scientific field. As benchmarks play an increasingly central role, they also grow more susceptible to distortion. Chatbot Arena has emerged as the go-to leaderboard for ranking the most capable AI systems. Yet, in this work we identify systematic issues that have resulted in a distorted playing field. We find that undisclosed private testing practices benefit a handful of providers who are able to test multiple variants before public release and retract scores if desired. We establish that the ability of these providers to choose the best score leads to biased Arena scores due to selective disclosure of performance results. At an extreme, we identify 27 private LLM variants tested by Meta in the lead-up to the Llama-4 release. We also establish that proprietary closed models are sampled at higher rates (number of battles) and have fewer models removed from the arena than open-weight and open-source alternatives. Both these policies lead to large data access asymmetries over time. Providers like Google and OpenAI have received an estimated 19.2% and 20.4% of all data on the arena, respectively. In contrast, a combined 83 open-weight models have only received an estimated 29.7% of the total data. We show that access to Chatbot Arena data yields substantial benefits; even limited additional data can result in relative performance gains of up to 112% on the arena distribution, based on our conservative estimates. Together, these dynamics result in overfitting to Arena-specific dynamics rather than general model quality. The Arena builds on the substantial efforts of both the organizers and an open community that maintains this valuable evaluation platform. We offer actionable recommendations to reform the Chatbot Arena's evaluation framework and promote fairer, more transparent benchmarking for the field

Anjiang Wei, Yuheng Wu, Yingjia Wan et al. · stanford

We introduce SATBench, a benchmark for evaluating the logical reasoning capabilities of large language models (LLMs) through logical puzzles derived from Boolean satisfiability (SAT) problems. Unlike prior work that focuses on inference rule-based reasoning, which often involves deducing conclusions from a set of premises, our approach leverages the search-based nature of SAT problems, where the objective is to find a solution that fulfills a specified set of logical constraints. Each instance in SATBench is generated from a SAT formula, then translated into a puzzle using LLMs. The generation process is fully automated and allows for adjustable difficulty by varying the number of clauses. All 2100 puzzles are validated through both LLM-based and solver-based consistency checks, with human validation on a subset. Experimental results show that even the strongest model, o4-mini, achieves only 65.0% accuracy on hard UNSAT problems, close to the random baseline of 50%. Our error analysis reveals systematic failures such as satisfiability bias, context inconsistency, and condition omission, highlighting limitations of current LLMs in search-based logical reasoning. Our code and data are publicly available at https://github.com/Anjiang-Wei/SATBench

Asad Aali, Vasiliki Bikia, Maya Varma et al. · stanford

With the growing use of language models (LMs) in clinical environments, there is an immediate need to evaluate the accuracy and safety of LM-generated medical text. Currently, such evaluation relies solely on manual physician review. However, detecting errors in LM-generated text is challenging because 1) manual review is costly and 2) expert-composed reference outputs are often unavailable in real-world settings. While the "LM-as-judge" paradigm (a LM evaluating another LM) offers scalable evaluation, even frontier LMs can miss subtle but clinically significant errors. To address these challenges, we propose MedVAL, a novel, self-supervised, data-efficient distillation method that leverages synthetic data to train evaluator LMs to assess whether LM-generated medical outputs are factually consistent with inputs, without requiring physician labels or reference outputs. To evaluate LM performance, we introduce MedVAL-Bench, a dataset of 840 physician-annotated outputs across 6 diverse medical tasks capturing real-world challenges. Across 10 state-of-the-art LMs spanning open-source and proprietary models, MedVAL distillation significantly improves (p < 0.001) alignment with physicians across seen and unseen tasks, increasing average F1 scores from 66% to 83%. Despite strong baseline performance, MedVAL improves the best-performing proprietary LM (GPT-4o) by 8% without training on physician-labeled data, demonstrating a performance statistically non-inferior to a single human expert (p < 0.001). To support a scalable, risk-aware pathway towards clinical integration, we open-source: 1) Codebase (https://github.com/StanfordMIMI/MedVAL), 2) MedVAL-Bench (https://huggingface.co/datasets/stanfordmimi/MedVAL-Bench), 3) MedVAL-4B (https://huggingface.co/stanfordmimi/MedVAL-4B). Our benchmark provides evidence of LMs approaching expert-level ability in validating AI-generated medical text.

Karan Dalal, Daniel Koceja, Gashon Hussein et al.

Transformers today still struggle to generate one-minute videos because self-attention layers are inefficient for long context. Alternatives such as Mamba layers struggle with complex multi-scene stories because their hidden states are less expressive. We experiment with Test-Time Training (TTT) layers, whose hidden states themselves can be neural networks, therefore more expressive. Adding TTT layers into a pre-trained Transformer enables it to generate one-minute videos from text storyboards. For proof of concept, we curate a dataset based on Tom and Jerry cartoons. Compared to baselines such as Mamba~2, Gated DeltaNet, and sliding-window attention layers, TTT layers generate much more coherent videos that tell complex stories, leading by 34 Elo points in a human evaluation of 100 videos per method. Although promising, results still contain artifacts, likely due to the limited capability of the pre-trained 5B model. The efficiency of our implementation can also be improved. We have only experimented with one-minute videos due to resource constraints, but the approach can be extended to longer videos and more complex stories. Sample videos, code and annotations are available at: https://test-time-training.github.io/video-dit

Rylan Schaeffer, Joshua Kazdan, John Hughes et al.

Recent research across mathematical problem solving, proof assistant programming and multimodal jailbreaking documents a striking finding: when (multimodal) language model tackle a suite of tasks with multiple attempts per task -- succeeding if any attempt is correct -- then the negative log of the average success rate scales a power law in the number of attempts. In this work, we identify an apparent puzzle: a simple mathematical calculation predicts that on each problem, the failure rate should fall exponentially with the number of attempts. We confirm this prediction empirically, raising a question: from where does aggregate polynomial scaling emerge? We then answer this question by demonstrating per-problem exponential scaling can be made consistent with aggregate polynomial scaling if the distribution of single-attempt success probabilities is heavy tailed such that a small fraction of tasks with extremely low success probabilities collectively warp the aggregate success trend into a power law - even as each problem scales exponentially on its own. We further demonstrate that this distributional perspective explains previously observed deviations from power law scaling, and provides a simple method for forecasting the power law exponent with an order of magnitude lower relative error, or equivalently, ${\sim}2-4$ orders of magnitude less inference compute. Overall, our work contributes to a better understanding of how neural language model performance improves with scaling inference compute and the development of scaling-predictable evaluations of (multimodal) language models.

Olawale Salaudeen, Anka Reuel, Ahmed Ahmed et al.

While the capabilities and utility of AI systems have advanced, rigorous norms for evaluating these systems have lagged. Grand claims, such as models achieving general reasoning capabilities, are supported with model performance on narrow benchmarks, like performance on graduate-level exam questions, which provide a limited and potentially misleading assessment. We provide a structured approach for reasoning about the types of evaluative claims that can be made given the available evidence. For instance, our framework helps determine whether performance on a mathematical benchmark is an indication of the ability to solve problems on math tests or instead indicates a broader ability to reason. Our framework is well-suited for the contemporary paradigm in machine learning, where various stakeholders provide measurements and evaluations that downstream users use to validate their claims and decisions. At the same time, our framework also informs the construction of evaluations designed to speak to the validity of the relevant claims. By leveraging psychometrics' breakdown of validity, evaluations can prioritize the most critical facets for a given claim, improving empirical utility and decision-making efficacy. We illustrate our framework through detailed case studies of vision and language model evaluations, highlighting how explicitly considering validity strengthens the connection between evaluation evidence and the claims being made.

Andy Zhou, Kevin Wu, Francesco Pinto et al.

As large language models (LLMs) become increasingly capable, security and safety evaluation are crucial. While current red teaming approaches have made strides in assessing LLM vulnerabilities, they often rely heavily on human input and lack comprehensive coverage of emerging attack vectors. This paper introduces AutoRedTeamer, a novel framework for fully automated, end-to-end red teaming against LLMs. AutoRedTeamer combines a multi-agent architecture with a memory-guided attack selection mechanism to enable continuous discovery and integration of new attack vectors. The dual-agent framework consists of a red teaming agent that can operate from high-level risk categories alone to generate and execute test cases and a strategy proposer agent that autonomously discovers and implements new attacks by analyzing recent research. This modular design allows AutoRedTeamer to adapt to emerging threats while maintaining strong performance on existing attack vectors. We demonstrate AutoRedTeamer's effectiveness across diverse evaluation settings, achieving 20% higher attack success rates on HarmBench against Llama-3.1-70B while reducing computational costs by 46% compared to existing approaches. AutoRedTeamer also matches the diversity of human-curated benchmarks in generating test cases, providing a comprehensive, scalable, and continuously evolving framework for evaluating the security of AI systems.

Belinda Mo, Kyssen Yu, Joshua Kazdan et al.

Recent interest in building foundation models for KGs has highlighted a fundamental challenge: knowledge-graph data is relatively scarce. The best-known KGs are primarily human-labeled, created by pattern-matching, or extracted using early NLP techniques. While human-generated KGs are in short supply, automatically extracted KGs are of questionable quality. We present a solution to this data scarcity problem in the form of a text-to-KG generator (KGGen), a package that uses language models to create high-quality graphs from plaintext. Unlike other KG extractors, KGGen clusters related entities to reduce sparsity in extracted KGs. KGGen is available as a Python library (\texttt{pip install kg-gen}), making it accessible to everyone. Along with KGGen, we release the first benchmark, Measure of of Information in Nodes and Edges (MINE), that tests an extractor's ability to produce a useful KG from plain text. We benchmark our new tool against existing extractors and demonstrate far superior performance.

Hejie Cui, Alyssa Unell, Bowen Chen et al. · stanford

Large language models (LLMs) have emerged as promising tools for assisting in medical tasks, yet processing Electronic Health Records (EHRs) presents unique challenges due to their longitudinal nature. While LLMs' capabilities to perform medical tasks continue to improve, their ability to reason over temporal dependencies across multiple patient visits and time frames remains unexplored. We introduce TIMER (Temporal Instruction Modeling and Evaluation for Longitudinal Clinical Records), a framework that incorporate instruction-response pairs grounding to different parts of a patient's record as a critical dimension in both instruction evaluation and tuning for longitudinal clinical records. We develop TIMER-Bench, the first time-aware benchmark that evaluates temporal reasoning capabilities over longitudinal EHRs, as well as TIMER-Instruct, an instruction-tuning methodology for LLMs to learn reasoning over time. We demonstrate that models fine-tuned with TIMER-Instruct improve performance by 7.3% on human-generated benchmarks and 9.2% on TIMER-Bench, indicating that temporal instruction-tuning improves model performance for reasoning over EHR.

Katherine Tsai, Stephen R. Pfohl, Olawale Salaudeen et al.

We study the problem of domain adaptation under distribution shift, where the shift is due to a change in the distribution of an unobserved, latent variable that confounds both the covariates and the labels. In this setting, neither the covariate shift nor the label shift assumptions apply. Our approach to adaptation employs proximal causal learning, a technique for estimating causal effects in settings where proxies of unobserved confounders are available. We demonstrate that proxy variables allow for adaptation to distribution shift without explicitly recovering or modeling latent variables. We consider two settings, (i) Concept Bottleneck: an additional ''concept'' variable is observed that mediates the relationship between the covariates and labels; (ii) Multi-domain: training data from multiple source domains is available, where each source domain exhibits a different distribution over the latent confounder. We develop a two-stage kernel estimation approach to adapt to complex distribution shifts in both settings. In our experiments, we show that our approach outperforms other methods, notably those which explicitly recover the latent confounder.

Olawale Salaudeen, Sanmi Koyejo

Given a causal graph representing the data-generating process shared across different domains/distributions, enforcing sufficient graph-implied conditional independencies can identify domain-general (non-spurious) feature representations. For the standard input-output predictive setting, we categorize the set of graphs considered in the literature into two distinct groups: (i) those in which the empirical risk minimizer across training domains gives domain-general representations and (ii) those where it does not. For the latter case (ii), we propose a novel framework with regularizations, which we demonstrate are sufficient for identifying domain-general feature representations without a priori knowledge (or proxies) of the spurious features. Empirically, our proposed method is effective for both (semi) synthetic and real-world data, outperforming other state-of-the-art methods in average and worst-domain transfer accuracy.

Rylan Schaeffer, Nika Zahedi, Mikail Khona et al. · mit

Associative memory and probabilistic modeling are two fundamental topics in artificial intelligence. The first studies recurrent neural networks designed to denoise, complete and retrieve data, whereas the second studies learning and sampling from probability distributions. Based on the observation that associative memory's energy functions can be seen as probabilistic modeling's negative log likelihoods, we build a bridge between the two that enables useful flow of ideas in both directions. We showcase four examples: First, we propose new energy-based models that flexibly adapt their energy functions to new in-context datasets, an approach we term \textit{in-context learning of energy functions}. Second, we propose two new associative memory models: one that dynamically creates new memories as necessitated by the training data using Bayesian nonparametrics, and another that explicitly computes proportional memory assignments using the evidence lower bound. Third, using tools from associative memory, we analytically and numerically characterize the memory capacity of Gaussian kernel density estimators, a widespread tool in probababilistic modeling. Fourth, we study a widespread implementation choice in transformers -- normalization followed by self attention -- to show it performs clustering on the hypersphere. Altogether, this work urges further exchange of useful ideas between these two continents of artificial intelligence.

Olawale Salaudeen, Nicole Chiou, Shiny Weng et al.

Spurious correlations, unstable statistical shortcuts a model can exploit, are expected to degrade performance out-of-distribution (OOD). However, across many popular OOD generalization benchmarks, vanilla empirical risk minimization (ERM) often achieves the highest OOD accuracy. Moreover, gains in in-distribution accuracy generally improve OOD accuracy, a phenomenon termed accuracy on the line, which contradicts the expected harm of spurious correlations. We show that these observations are an artifact of misspecified OOD datasets that do not include shifts in spurious correlations that harm OOD generalization, the setting they are meant to evaluate. Consequently, current practice evaluates "robustness" without truly stressing the spurious signals we seek to eliminate; our work pinpoints when that happens and how to fix it. Contributions. (i) We derive necessary and sufficient conditions for a distribution shift to reveal a model's reliance on spurious features; when these conditions hold, "accuracy on the line" disappears. (ii) We audit leading OOD datasets and find that most still display accuracy on the line, suggesting they are misspecified for evaluating robustness to spurious correlations. (iii) We catalog the few well-specified datasets and summarize generalizable design principles, such as identifying datasets of natural interventions (e.g., a pandemic), to guide future well-specified benchmarks.

Xiangwen Wang, Yibo Jacky Zhang, Zhoujie Ding et al.

Compound AI systems, comprising multiple interacting components such as LLMs, foundation models, and external tools, have demonstrated remarkable improvements compared to single models in various tasks. To ensure their effective deployment in real-world applications, aligning these systems with human preferences is crucial. However, aligning the compound system via policy optimization, unlike the alignment of a single model, is challenging for two main reasons: (i) non-differentiable interactions between components make end-to-end gradient-based optimization method inapplicable, and (ii) system-level preferences cannot be directly transformed into component-level preferences. To address these challenges, we first formulate compound AI systems as Directed Acyclic Graphs (DAGs), explicitly modeling both component interactions and the associated data flows. Building on this formulation, we introduce $\textbf{SysDPO}$, a framework that extends Direct Preference Optimization (DPO) to enable joint system-level alignment. We propose two variants, SysDPO-Direct and SysDPO-Sampling, tailored for scenarios depending on whether we construct a system-specific preference dataset. We empirically demonstrate the effectiveness of our approach across two applications: the joint alignment of a language model and a diffusion model, and the joint alignment of an LLM collaboration system.

Rylan Schaeffer, Punit Singh Koura, Binh Tang et al.

The explosion of high-performing conversational language models (LMs) has spurred a shift from classic natural language processing (NLP) benchmarks to expensive, time-consuming and noisy human evaluations - yet the relationship between these two evaluation strategies remains hazy. In this paper, we conduct a large-scale study of four Chat Llama 2 models, comparing their performance on 160 standard NLP benchmarks (e.g., MMLU, ARC, BIG-Bench Hard) against extensive human preferences on more than 11k single-turn and 2k multi-turn dialogues from over 2k human annotators. Our findings are striking: most NLP benchmarks strongly correlate with human evaluations, suggesting that cheaper, automated metrics can serve as surprisingly reliable predictors of human preferences. Three human evaluations, such as adversarial dishonesty and safety, are anticorrelated with NLP benchmarks, while two are uncorrelated. Moreover, through overparameterized linear regressions, we show that NLP scores can accurately predict human evaluations across different model scales, offering a path to reduce costly human annotation without sacrificing rigor. Overall, our results affirm the continued value of classic benchmarks and illuminate how to harness them to anticipate real-world user satisfaction - pointing to how NLP benchmarks can be leveraged to meet evaluation needs of our new era of conversational AI.

Edward Chen, Natalie Dullerud, Thomas Niedermayr et al.

Countless science and engineering applications in multi-objective optimization (MOO) necessitate that decision-makers (DMs) select a Pareto-optimal solution which aligns with their preferences. Evaluating individual solutions is often expensive, necessitating cost-sensitive optimization techniques. Due to competing objectives, the space of trade-offs is also expansive -- thus, examining the full Pareto frontier may prove overwhelming to a DM. Such real-world settings generally have loosely-defined and context-specific desirable regions for each objective function that can aid in constraining the search over the Pareto frontier. We introduce a novel conceptual framework that operationalizes these priors using soft-hard functions, SHFs, which allow for the DM to intuitively impose soft and hard bounds on each objective -- which has been lacking in previous MOO frameworks. Leveraging a novel minimax formulation for Pareto frontier sampling, we propose a two-step process for obtaining a compact set of Pareto-optimal points which respect the user-defined soft and hard bounds: (1) densely sample the Pareto frontier using Bayesian optimization, and (2) sparsify the selected set to surface to the user, using robust submodular function optimization. We prove that (2) obtains the optimal compact Pareto-optimal set of points from (1). We further show that many practical problems fit within the SHF framework and provide extensive empirical validation on diverse domains, including brachytherapy, engineering design, and large language model personalization. Specifically, for brachytherapy, our approach returns a compact set of points with over 3% greater SHF-defined utility than the next best approach. Among the other diverse experiments, our approach consistently leads in utility, allowing the DM to reach >99% of their maximum possible desired utility within validation of 5 points.

Ahmed Ahmed, Kevin Klyman, Yi Zeng et al.

Companies that develop foundation models publish behavioral guidelines they pledge their models will follow, but it remains unclear if models actually do so. While providers such as OpenAI, Anthropic, and Google have published detailed specifications describing both desired safety constraints and qualitative traits for their models, there has been no systematic audit of adherence to these guidelines. We introduce an automated framework that audits models against their providers specifications by parsing behavioral statements, generating targeted prompts, and using models to judge adherence. Our central focus is on three way consistency between a provider specification, its model outputs, and its own models as judges; an extension of prior two way generator validator consistency. This establishes a necessary baseline: at minimum, a foundation model should consistently satisfy the developer behavioral specifications when judged by the developer evaluator models. We apply our framework to 16 models from six developers across more than 100 behavioral statements, finding systematic inconsistencies including compliance gaps of up to 20 percent across providers.

Syrine Belakaria, Joshua Kazdan, Charles Marx et al.

Reinforcement learning from human feedback (RLHF) has become a cornerstone of the training and alignment pipeline for large language models (LLMs). Recent advances, such as direct preference optimization (DPO), have simplified the preference learning step. However, collecting preference data remains a challenging and costly process, often requiring expert annotation. This cost can be mitigated by carefully selecting the data points presented for annotation. In this work, we propose an active learning approach to efficiently select prompt and preference pairs using a risk assessment strategy based on the Sharpe Ratio. To address the challenge of unknown preferences prior to annotation, our method evaluates the gradients of all potential preference annotations to assess their impact on model updates. These gradient-based evaluations enable risk assessment of data points regardless of the annotation outcome. By leveraging the DPO loss derivations, we derive a closed-form expression for computing these Sharpe ratios on a per-tuple basis, ensuring our approach remains both tractable and computationally efficient. We also introduce two variants of our method, each making different assumptions about prior information. Experimental results demonstrate that our method outperforms the baseline by up to 5% in win rates against the chosen completion with limited human preference data across several language models and real-world datasets.

Tai Dang, Long-Hung Pham, Sang T. Truong et al.

Despite decades of advancements in automated ligand screening, large-scale drug discovery remains resource-intensive and requires post-processing hit selection, a step where chemists manually select a few promising molecules based on their chemical intuition. This creates a major bottleneck in the virtual screening process for drug discovery, demanding experts to repeatedly balance complex trade-offs among drug properties across a vast pool of candidates. To improve the efficiency and reliability of this process, we propose a novel human-centered framework named CheapVS that allows chemists to guide the ligand selection process by providing preferences regarding the trade-offs between drug properties via pairwise comparison. Our framework combines preferential multi-objective Bayesian optimization with a docking model for measuring binding affinity to capture human chemical intuition for improving hit identification. Specifically, on a library of 100K chemical candidates targeting EGFR and DRD2, CheapVS outperforms state-of-the-art screening methods in identifying drugs within a limited computational budget. Notably, our method can recover up to 16/37 EGFR and 37/58 DRD2 known drugs while screening only 6% of the library, showcasing its potential to significantly advance drug discovery.

Pura Peetathawatchai, Wei-Ning Chen, Berivan Isik et al. · stanford

Personalizing large-scale diffusion models poses serious privacy risks, especially when adapting to small, sensitive datasets. A common approach is to fine-tune the model using differentially private stochastic gradient descent (DP-SGD), but this suffers from severe utility degradation due to the high noise needed for privacy, particularly in the small data regime. We propose an alternative that leverages Textual Inversion (TI), which learns an embedding vector for an image or set of images, to enable adaptation under differential privacy (DP) constraints. Our approach, Differentially Private Aggregation via Textual Inversion (DPAgg-TI), adds calibrated noise to the aggregation of per-image embeddings to ensure formal DP guarantees while preserving high output fidelity. We show that DPAgg-TI outperforms DP-SGD finetuning in both utility and robustness under the same privacy budget, achieving results closely matching the non-private baseline on style adaptation tasks using private artwork from a single artist and Paris 2024 Olympic pictograms. In contrast, DP-SGD fails to generate meaningful outputs in this setting.

Lovish Madaan, Aaditya K. Singh, Rylan Schaeffer et al.

Evaluation benchmarks are the cornerstone of measuring capabilities of large language models (LLMs), as well as driving progress in said capabilities. Originally designed to make claims about capabilities (or lack thereof) in fully pretrained models, evaluation benchmarks are now also extensively used to decide between various training choices. Despite this widespread usage, we rarely quantify the variance in our evaluation benchmarks, which dictates whether differences in performance are meaningful. Here, we define and measure a range of metrics geared towards measuring variance in evaluation benchmarks, including seed variance across initialisations, and monotonicity during training. By studying a large number of models -- both openly available and pretrained from scratch -- we provide empirical estimates for a variety of variance metrics, with considerations and recommendations for practitioners. We also evaluate the utility and tradeoffs of continuous versus discrete performance measures and explore options for better understanding and reducing this variance. We find that simple changes, such as framing choice tasks (like MMLU) as completion tasks, can often reduce variance for smaller scale ($\sim$7B) models, while more involved methods inspired from human testing literature (such as item analysis and item response theory) struggle to meaningfully reduce variance. Overall, our work provides insights into variance in evaluation benchmarks, suggests LM-specific techniques to reduce variance, and more generally encourages practitioners to carefully factor in variance when comparing models.

Xiaojun Xu, Jacky Yibo Zhang, Evelyn Ma et al.

Machine learning (ML) robustness and domain generalization are fundamentally correlated: they essentially concern data distribution shifts under adversarial and natural settings, respectively. On one hand, recent studies show that more robust (adversarially trained) models are more generalizable. On the other hand, there is a lack of theoretical understanding of their fundamental connections. In this paper, we explore the relationship between regularization and domain transferability considering different factors such as norm regularization and data augmentations (DA). We propose a general theoretical framework proving that factors involving the model function class regularization are sufficient conditions for relative domain transferability. Our analysis implies that ``robustness" is neither necessary nor sufficient for transferability; rather, regularization is a more fundamental perspective for understanding domain transferability. We then discuss popular DA protocols (including adversarial training) and show when they can be viewed as the function class regularization under certain conditions and therefore improve generalization. We conduct extensive experiments to verify our theoretical findings and show several counterexamples where robustness and generalization are negatively correlated on different datasets.

Jessica Schrouff, Natalie Harris, Oluwasanmi Koyejo et al.

Diagnosing and mitigating changes in model fairness under distribution shift is an important component of the safe deployment of machine learning in healthcare settings. Importantly, the success of any mitigation strategy strongly depends on the structure of the shift. Despite this, there has been little discussion of how to empirically assess the structure of a distribution shift that one is encountering in practice. In this work, we adopt a causal framing to motivate conditional independence tests as a key tool for characterizing distribution shifts. Using our approach in two medical applications, we show that this knowledge can help diagnose failures of fairness transfer, including cases where real-world shifts are more complex than is often assumed in the literature. Based on these results, we discuss potential remedies at each step of the machine learning pipeline.

Katherine Tsai, Oluwasanmi Koyejo, Mladen Kolar

Graphs from complex systems often share a partial underlying structure across domains while retaining individual features. Thus, identifying common structures can shed light on the underlying signal, for instance, when applied to scientific discoveries or clinical diagnoses. Furthermore, growing evidence shows that the shared structure across domains boosts the estimation power of graphs, particularly for high-dimensional data. However, building a joint estimator to extract the common structure may be more complicated than it seems, most often due to data heterogeneity across sources. This manuscript surveys recent work on statistical inference of joint Gaussian graphical models, identifying model structures that fit various data generation processes. Simulations under different data generation processes are implemented with detailed discussions on the choice of models.

Raj Kiriti Velicheti, Derek Xia, Oluwasanmi Koyejo

Federated learning systems that jointly preserve Byzantine robustness and privacy have remained an open problem. Robust aggregation, the standard defense for Byzantine attacks, generally requires server access to individual updates or nonlinear computation -- thus is incompatible with privacy-preserving methods such as secure aggregation via multiparty computation. To this end, we propose SHARE (Secure Hierarchical Robust Aggregation), a distributed learning framework designed to cryptographically preserve client update privacy and robustness to Byzantine adversaries simultaneously. The key idea is to incorporate secure averaging among randomly clustered clients before filtering malicious updates through robust aggregation. Experiments show that SHARE has similar robustness guarantees as existing techniques while enhancing privacy.

Gaurush Hiranandani, Jatin Mathur, Harikrishna Narasimhan et al.

We consider learning to optimize a classification metric defined by a black-box function of the confusion matrix. Such black-box learning settings are ubiquitous, for example, when the learner only has query access to the metric of interest, or in noisy-label and domain adaptation applications where the learner must evaluate the metric via performance evaluation using a small validation sample. Our approach is to adaptively learn example weights on the training dataset such that the resulting weighted objective best approximates the metric on the validation sample. We show how to model and estimate the example weights and use them to iteratively post-shift a pre-trained class probability estimator to construct a classifier. We also analyze the resulting procedure's statistical properties. Experiments on various label noise, domain shift, and fair classification setups confirm that our proposal compares favorably to the state-of-the-art baselines for each application.

Peiye Zhuang, Oluwasanmi Koyejo, Alexander G. Schwing

Controllable semantic image editing enables a user to change entire image attributes with a few clicks, e.g., gradually making a summer scene look like it was taken in winter. Classic approaches for this task use a Generative Adversarial Net (GAN) to learn a latent space and suitable latent-space transformations. However, current approaches often suffer from attribute edits that are entangled, global image identity changes, and diminished photo-realism. To address these concerns, we learn multiple attribute transformations simultaneously, integrate attribute regression into the training of transformation functions, and apply a content loss and an adversarial loss that encourages the maintenance of image identity and photo-realism. We propose quantitative evaluation strategies for measuring controllable editing performance, unlike prior work, which primarily focuses on qualitative evaluation. Our model permits better control for both single- and multiple-attribute editing while preserving image identity and realism during transformation. We provide empirical results for both natural and synthetic images, highlighting that our model achieves state-of-the-art performance for targeted image manipulation.

Katherine Tsai, Mladen Kolar, Oluwasanmi Koyejo

We propose a flexible yet interpretable model for high-dimensional data with time-varying second order statistics, motivated and applied to functional neuroimaging data. Motivated by the neuroscience literature, we factorize the covariances into sparse spatial and smooth temporal components. While this factorization results in both parsimony and domain interpretability, the resulting estimation problem is nonconvex. To this end, we design a two-stage optimization scheme with a carefully tailored spectral initialization, combined with iteratively refined alternating projected gradient descent. We prove a linear convergence rate up to a nontrivial statistical error for the proposed descent scheme and establish sample complexity guarantees for the estimator. We further quantify the statistical error for the multivariate Gaussian case. Empirical results using simulated and real brain imaging data illustrate that our approach outperforms existing baselines.

Gaurush Hiranandani, Jatin Mathur, Harikrishna Narasimhan et al.

Metric elicitation is a recent framework for eliciting classification performance metrics that best reflect implicit user preferences based on the task and context. However, available elicitation strategies have been limited to linear (or quasi-linear) functions of predictive rates, which can be practically restrictive for many applications including fairness. This paper develops a strategy for eliciting more flexible multiclass metrics defined by quadratic functions of rates, designed to reflect human preferences better. We show its application in eliciting quadratic violation-based group-fair metrics. Our strategy requires only relative preference feedback, is robust to noise, and achieves near-optimal query complexity. We further extend this strategy to eliciting polynomial metrics -- thus broadening the use cases for metric elicitation.

Cong Xie, Shuai Zheng, Oluwasanmi Koyejo et al.

The scalability of Distributed Stochastic Gradient Descent (SGD) is today limited by communication bottlenecks. We propose a novel SGD variant: Communication-efficient SGD with Error Reset, or CSER. The key idea in CSER is first a new technique called "error reset" that adapts arbitrary compressors for SGD, producing bifurcated local models with periodic reset of resulting local residual errors. Second we introduce partial synchronization for both the gradients and the models, leveraging advantages from them. We prove the convergence of CSER for smooth non-convex problems. Empirical results show that when combined with highly aggressive compressors, the CSER algorithms accelerate the distributed training by nearly 10x for CIFAR-100, and by 4.5x for ImageNet.

Jacky Y. Zhang, Rajiv Khanna, Anastasios Kyrillidis et al.

Bayesian coresets have emerged as a promising approach for implementing scalable Bayesian inference. The Bayesian coreset problem involves selecting a (weighted) subset of the data samples, such that the posterior inference using the selected subset closely approximates the posterior inference using the full dataset. This manuscript revisits Bayesian coresets through the lens of sparsity constrained optimization. Leveraging recent advances in accelerated optimization methods, we propose and analyze a novel algorithm for coreset selection. We provide explicit convergence rate guarantees and present an empirical evaluation on a variety of benchmark datasets to highlight our proposed algorithm's superior performance compared to state-of-the-art on speed and accuracy.

Gaurush Hiranandani, Harikrishna Narasimhan, Oluwasanmi Koyejo

What is a fair performance metric? We consider the choice of fairness metrics through the lens of metric elicitation -- a principled framework for selecting performance metrics that best reflect implicit preferences. The use of metric elicitation enables a practitioner to tune the performance and fairness metrics to the task, context, and population at hand. Specifically, we propose a novel strategy to elicit group-fair performance metrics for multiclass classification problems with multiple sensitive groups that also includes selecting the trade-off between predictive performance and fairness violation. The proposed elicitation strategy requires only relative preference feedback and is robust to both finite sample and feedback noise.

Amar Budhiraja, Gaurush Hiranandani, Darshak Chhatbar et al.

In this paper, we study the problem of recommendation system where the users and items to be recommended are rich data structures with multiple entity types and with multiple sources of side-information in the form of graphs. We provide a general formulation for the problem that captures the complexities of modern real-world recommendations and generalizes many existing formulations. In our formulation, each user/document that requires a recommendation and each item or tag that is to be recommended, both are modeled by a set of static entities and a dynamic component. The relationships between entities are captured by several weighted bipartite graphs. To effectively exploit these complex interactions and learn the recommendation model, we propose MEDRES- a multiple graph-CNN based novel deep-learning architecture. MEDRES uses AL-GCN, a novel graph convolution network block, that harnesses strong representative features from the underlying graphs. Moreover, in order to capture highly heterogeneous engagement of different users with the system and constraints on the number of items to be recommended, we propose a novel ranking metric pAp@k along with a method to optimize the metric directly. We demonstrate effectiveness of our method on two benchmarks: a) citation data, b) Flickr data. In addition, we present two real-world case studies of our formulation and the MEDRES architecture. We show how our technique can be used to naturally model the message recommendation problem and the teams recommendation problem in the Microsoft Teams (MSTeams) product and demonstrate that it is 5-6% points more accurate than the production-grade models.

Zengjie Song, Oluwasanmi Koyejo, Jiangshe Zhang

This paper addresses two crucial problems of learning disentangled image representations, namely controlling the degree of disentanglement during image editing, and balancing the disentanglement strength and the reconstruction quality. To encourage disentanglement, we devise a distance covariance based decorrelation regularization. Further, for the reconstruction step, our model leverages a soft target representation combined with the latent image code. By exploring the real-valued space of the soft target representation, we are able to synthesize novel images with the designated properties. To improve the perceptual quality of images generated by autoencoder (AE)-based models, we extend the encoder-decoder architecture with the generative adversarial network (GAN) by collapsing the AE decoder and the GAN generator into one. We also design a classification based protocol to quantitatively evaluate the disentanglement strength of our model. Experimental results showcase the benefits of the proposed model.

Zengjie Song, Oluwasanmi Koyejo, Jiangshe Zhang

A crucial problem in learning disentangled image representations is controlling the degree of disentanglement during image editing, while preserving the identity of objects. In this work, we propose a simple yet effective model with the encoder-decoder architecture to address this challenge. To encourage disentanglement, we devise a distance covariance based decorrelation regularization. Further, for the reconstruction step, our model leverages a soft target representation combined with the latent image code. By exploiting the real-valued space of the soft target representations, we are able to synthesize novel images with the designated properties. We also design a classification based protocol to quantitatively evaluate the disentanglement strength of our model. Experimental results show that the proposed model competently disentangles factors of variation, and is able to manipulate face images to synthesize the desired attributes.

Cong Xie, Oluwasanmi Koyejo, Indranil Gupta et al.

When scaling distributed training, the communication overhead is often the bottleneck. In this paper, we propose a novel SGD variant with reduced communication and adaptive learning rates. We prove the convergence of the proposed algorithm for smooth but non-convex problems. Empirical results show that the proposed algorithm significantly reduces the communication overhead, which, in turn, reduces the training time by up to 30% for the 1B word dataset.