arXiv subjects

Eric Horvitz

Publications and source records attributed to Eric Horvitz.

At least 37 records · Page 2Linked to original sources

arXiv2025

Sequential Diagnosis with Language Models

Artificial intelligence holds great promise for expanding access to expert medical knowledge and reasoning. However, most evaluations of language models rely on static vignettes and multiple-choice questions that fail to reflect the complexity and nuance of evidence-based medicine in real-world settings. In clinical practice, physicians iteratively formulate and revise diagnostic hypotheses, adapting each subsequent question and test to what they've just learned, and weigh the evolving evidence before committing to a final diagnosis. To emulate this iterative process, we introduce the Sequential Diagnosis Benchmark, which transforms 304 diagnostically challenging New England Journal of Medicine clinicopathological conference (NEJM-CPC) cases into stepwise diagnostic encounters. A physician or AI begins with a short case abstract and must iteratively request additional details from a gatekeeper model that reveals findings only when explicitly queried. Performance is assessed not just by diagnostic accuracy but also by the cost of physician visits and tests performed. We also present the MAI Diagnostic Orchestrator (MAI-DxO), a model-agnostic orchestrator that simulates a panel of physicians, proposes likely differential diagnoses and strategically selects high-value, cost-effective tests. When paired with OpenAI's o3 model, MAI-DxO achieves 80% diagnostic accuracy--four times higher than the 20% average of generalist physicians. MAI-DxO also reduces diagnostic costs by 20% compared to physicians, and 70% compared to off-the-shelf o3. When configured for maximum accuracy, MAI-DxO achieves 85.5% accuracy. These performance gains with MAI-DxO generalize across models from the OpenAI, Gemini, Claude, Grok, DeepSeek, and Llama families. We highlight how AI systems, when guided to think iteratively and act judiciously, can advance diagnostic precision and cost-effectiveness in clinical care.

cs.CL↗

arXiv2025

Superhuman performance of a large language model on the reasoning tasks of a physician

A seminal paper published by Ledley and Lusted in 1959 introduced complex clinical diagnostic reasoning cases as the gold standard for the evaluation of expert medical computing systems, a standard that has held ever since. Here, we report the results of a physician evaluation of a large language model (LLM) on challenging clinical cases against a baseline of hundreds of physicians. We conduct five experiments to measure clinical reasoning across differential diagnosis generation, display of diagnostic reasoning, triage differential diagnosis, probabilistic reasoning, and management reasoning, all adjudicated by physician experts with validated psychometrics. We then report a real-world study comparing human expert and AI second opinions in randomly-selected patients in the emergency room of a major tertiary academic medical center in Boston, MA. We compared LLMs and board-certified physicians at three predefined diagnostic touchpoints: triage in the emergency room, initial evaluation by a physician, and admission to the hospital or intensive care unit. In all experiments--both vignettes and emergency room second opinions--the LLM displayed superhuman diagnostic and reasoning abilities, as well as continued improvement from prior generations of AI clinical decision support. Our study suggests that LLMs have achieved superhuman performance on general medical diagnostic and management reasoning, fulfilling the vision put forth by Ledley and Lusted, and motivating the urgent need for prospective trials.

cs.AI↗

arXiv2025

MedHELM: Holistic Evaluation of Large Language Models for Medical Tasks

While large language models (LLMs) achieve near-perfect scores on medical licensing exams, these evaluations inadequately reflect the complexity and diversity of real-world clinical practice. We introduce MedHELM, an extensible evaluation framework for assessing LLM performance for medical tasks with three key contributions. First, a clinician-validated taxonomy spanning 5 categories, 22 subcategories, and 121 tasks developed with 29 clinicians. Second, a comprehensive benchmark suite comprising 35 benchmarks (17 existing, 18 newly formulated) providing complete coverage of all categories and subcategories in the taxonomy. Third, a systematic comparison of LLMs with improved evaluation methods (using an LLM-jury) and a cost-performance analysis. Evaluation of 9 frontier LLMs, using the 35 benchmarks, revealed significant performance variation. Advanced reasoning models (DeepSeek R1: 66% win-rate; o3-mini: 64% win-rate) demonstrated superior performance, though Claude 3.5 Sonnet achieved comparable results at 40% lower estimated computational cost. On a normalized accuracy scale (0-1), most models performed strongly in Clinical Note Generation (0.73-0.85) and Patient Communication & Education (0.78-0.83), moderately in Medical Research Assistance (0.65-0.75), and generally lower in Clinical Decision Support (0.56-0.72) and Administration & Workflow (0.53-0.63). Our LLM-jury evaluation method achieved good agreement with clinician ratings (ICC = 0.47), surpassing both average clinician-clinician agreement (ICC = 0.43) and automated baselines including ROUGE-L (0.36) and BERTScore-F1 (0.44). Claude 3.5 Sonnet achieved comparable performance to top models at lower estimated cost. These findings highlight the importance of real-world, task-specific evaluation for medical use of LLMs and provides an open source framework to enable this.

cs.CL↗

arXiv2025

Navigating Rifts in Human-LLM Grounding: Study and Benchmark

Language models excel at following instructions but often struggle with the collaborative aspects of conversation that humans naturally employ. This limitation in grounding -- the process by which conversation participants establish mutual understanding -- can lead to outcomes ranging from frustrated users to serious consequences in high-stakes scenarios. To systematically study grounding challenges in human-LLM interactions, we analyze logs from three human-assistant datasets: WildChat, MultiWOZ, and Bing Chat. We develop a taxonomy of grounding acts and build models to annotate and forecast grounding behavior. Our findings reveal significant differences in human-human and human-LLM grounding: LLMs were three times less likely to initiate clarification and sixteen times less likely to provide follow-up requests than humans. Additionally, we find that early grounding failures predict later interaction breakdowns. Building on these insights, we introduce Rifts, a benchmark derived from publicly available LLM interaction data containing situations where LLMs fail to initiate grounding. We note that current frontier models perform poorly on Rifts, highlighting the need to reconsider how we train and prompt LLMs for human interaction. To this end, we develop a preliminary intervention aimed at mitigating grounding failures.

cs.CL↗

arXiv2025

Steering Language Model Refusal with Sparse Autoencoders

Responsible deployment of language models requires mechanisms for refusing unsafe prompts while preserving model performance. While most approaches modify model weights through additional training, we explore an alternative: steering model activations at inference time via amplifying sparse autoencoder (SAE) features that mediate refusal. This work uncovers a fundamental tension between SAE steering-based safety improvements and general model capabilities. While feature steering successfully improves robustness against both single-turn and challenging multi-turn jailbreak attempts, we discover that this comes at a previously underexplored cost -- systematic degradation of performance across multiple benchmark tasks, even on safe inputs with no apparent connection to refusal behavior. This suggests that features mediating refusal may be more deeply entangled with general language model capabilities than previously understood. Our findings reveal important open questions about the nature of safety-relevant features in language models and the feasibility of isolating them for targeted intervention. While SAE-based steering shows promise as a flexible approach to enhancing language model safety, our results highlight the critical need to understand and address the mechanisms behind these capability tradeoffs before such techniques can be practically deployed.

cs.LG↗

arXiv2025

Improving Instruction-Following in Language Models through Activation Steering

The ability to follow instructions is crucial for numerous real-world applications of language models. In pursuit of deeper insights and more powerful capabilities, we derive instruction-specific vector representations from language models and use them to steer models accordingly. These vectors are computed as the difference in activations between inputs with and without instructions, enabling a modular approach to activation steering. We demonstrate how this method can enhance model adherence to constraints such as output format, length, and word inclusion, providing inference-time control over instruction following. Our experiments across four models demonstrate how we can use the activation vectors to guide models to follow constraints even without explicit instructions and to enhance performance when instructions are present. Additionally, we explore the compositionality of activation steering, successfully applying multiple instructions simultaneously. Finally, we demonstrate that steering vectors computed on instruction-tuned models can transfer to improve base models. Our findings demonstrate that activation steering offers a practical and scalable approach for fine-grained control in language generation. Our code and data are available at https://github.com/microsoft/llm-steer-instruct.

cs.CL↗

arXiv2025

Utility-Directed Conformal Prediction: A Decision-Aware Framework for Actionable Uncertainty Quantification

Interest has been growing in decision-focused machine learning methods which train models to account for how their predictions are used in downstream optimization problems. Doing so can often improve performance on subsequent decision problems. However, current methods for uncertainty quantification do not incorporate any information about downstream decisions. We develop a methodology based on conformal prediction to identify prediction sets that account for a downstream cost function, making them more appropriate to inform high-stakes decision-making. Our approach harnesses the strengths of conformal methods -- modularity, model-agnosticism, and statistical coverage guarantees -- while incorporating downstream decisions and user-specified utility functions. We prove that our methods retain standard coverage guarantees. Empirical evaluation across a range of datasets and utility metrics demonstrates that our methods achieve significantly lower costs than standard conformal methods. We present a real-world use case in healthcare diagnosis, where our method effectively incorporates the hierarchical structure of dermatological diseases. The method successfully generates sets with coherent diagnostic meaning, potentially aiding triage for dermatology diagnosis and illustrating how our method can ground high-stakes decision-making employing domain knowledge.

cs.LG↗

arXiv2025

JSONSchemaBench: A Rigorous Benchmark of Structured Outputs for Language Models

Reliably generating structured outputs has become a critical capability for modern language model (LM) applications. Constrained decoding has emerged as the dominant technology across sectors for enforcing structured outputs during generation. Despite its growing adoption, little has been done with the systematic evaluation of the behaviors and performance of constrained decoding. Constrained decoding frameworks have standardized around JSON Schema as a structured data format, with most uses guaranteeing constraint compliance given a schema. However, there is poor understanding of the effectiveness of the methods in practice. We present an evaluation framework to assess constrained decoding approaches across three critical dimensions: efficiency in generating constraint-compliant outputs, coverage of diverse constraint types, and quality of the generated outputs. To facilitate this evaluation, we introduce JSONSchemaBench, a benchmark for constrained decoding comprising 10K real-world JSON schemas that encompass a wide range of constraints with varying complexity. We pair the benchmark with the existing official JSON Schema Test Suite and evaluate six state-of-the-art constrained decoding frameworks, including Guidance, Outlines, Llamacpp, XGrammar, OpenAI, and Gemini. Through extensive experiments, we gain insights into the capabilities and limitations of constrained decoding on structured generation with real-world JSON schemas. Our work provides actionable insights for improving constrained decoding frameworks and structured generation tasks, setting a new standard for evaluating constrained decoding and structured generation. We release JSONSchemaBench at https://github.com/guidance-ai/jsonschemabench

cs.CL↗

arXiv2024

Challenges in Human-Agent Communication

Remarkable advancements in modern generative foundation models have enabled the development of sophisticated and highly capable autonomous agents that can observe their environment, invoke tools, and communicate with other agents to solve problems. Although such agents can communicate with users through natural language, their complexity and wide-ranging failure modes present novel challenges for human-AI interaction. Building on prior research and informed by a communication grounding perspective, we contribute to the study of \emph{human-agent communication} by identifying and analyzing twelve key communication challenges that these systems pose. These include challenges in conveying information from the agent to the user, challenges in enabling the user to convey information to the agent, and overarching challenges that need to be considered across all human-agent communication. We illustrate each challenge through concrete examples and identify open directions of research. Our findings provide insights into critical gaps in human-agent communication research and serve as an urgent call for new design patterns, principles, and guidelines to support transparency and control in these systems.

cs.HC↗

arXiv2024

From Medprompt to o1: Exploration of Run-Time Strategies for Medical Challenge Problems and Beyond

Run-time steering strategies like Medprompt are valuable for guiding large language models (LLMs) to top performance on challenging tasks. Medprompt demonstrates that a general LLM can be focused to deliver state-of-the-art performance on specialized domains like medicine by using a prompt to elicit a run-time strategy involving chain of thought reasoning and ensembling. OpenAI's o1-preview model represents a new paradigm, where a model is designed to do run-time reasoning before generating final responses. We seek to understand the behavior of o1-preview on a diverse set of medical challenge problem benchmarks. Following on the Medprompt study with GPT-4, we systematically evaluate the o1-preview model across various medical benchmarks. Notably, even without prompting techniques, o1-preview largely outperforms the GPT-4 series with Medprompt. We further systematically study the efficacy of classic prompt engineering strategies, as represented by Medprompt, within the new paradigm of reasoning models. We found that few-shot prompting hinders o1's performance, suggesting that in-context learning may no longer be an effective steering approach for reasoning-native models. While ensembling remains viable, it is resource-intensive and requires careful cost-performance optimization. Our cost and accuracy analysis across run-time strategies reveals a Pareto frontier, with GPT-4o representing a more affordable option and o1-preview achieving state-of-the-art performance at higher cost. Although o1-preview offers top performance, GPT-4o with steering strategies like Medprompt retains value in specific contexts. Moreover, we note that the o1-preview model has reached near-saturation on many existing medical benchmarks, underscoring the need for new, challenging benchmarks. We close with reflections on general directions for inference-time computation with LLMs.

cs.CL↗

arXiv2024

How to Build the Virtual Cell with Artificial Intelligence: Priorities and Opportunities

The cell is arguably the most fundamental unit of life and is central to understanding biology. Accurate modeling of cells is important for this understanding as well as for determining the root causes of disease. Recent advances in artificial intelligence (AI), combined with the ability to generate large-scale experimental data, present novel opportunities to model cells. Here we propose a vision of leveraging advances in AI to construct virtual cells, high-fidelity simulations of cells and cellular systems under different conditions that are directly learned from biological data across measurements and scales. We discuss desired capabilities of such AI Virtual Cells, including generating universal representations of biological entities across scales, and facilitating interpretable in silico experiments to predict and understand their behavior using virtual instruments. We further address the challenges, opportunities and requirements to realize this vision including data needs, evaluation strategies, and community standards and engagement to ensure biological accuracy and broad utility. We envision a future where AI Virtual Cells help identify new drug targets, predict cellular responses to perturbations, as well as scale hypothesis exploration. With open science collaborations across the biomedical ecosystem that includes academia, philanthropy, and the biopharma and AI industries, a comprehensive predictive understanding of cell mechanisms and interactions has come into reach.

q-bio.QM↗

arXiv2024

MedFuzz: Exploring the Robustness of Large Language Models in Medical Question Answering

Large language models (LLM) have achieved impressive performance on medical question-answering benchmarks. However, high benchmark accuracy does not imply that the performance generalizes to real-world clinical settings. Medical question-answering benchmarks rely on assumptions consistent with quantifying LLM performance but that may not hold in the open world of the clinic. Yet LLMs learn broad knowledge that can help the LLM generalize to practical conditions regardless of unrealistic assumptions in celebrated benchmarks. We seek to quantify how well LLM medical question-answering benchmark performance generalizes when benchmark assumptions are violated. Specifically, we present an adversarial method that we call MedFuzz (for medical fuzzing). MedFuzz attempts to modify benchmark questions in ways aimed at confounding the LLM. We demonstrate the approach by targeting strong assumptions about patient characteristics presented in the MedQA benchmark. Successful "attacks" modify a benchmark item in ways that would be unlikely to fool a medical expert but nonetheless "trick" the LLM into changing from a correct to an incorrect answer. Further, we present a permutation test technique that can ensure a successful attack is statistically significant. We show how to use performance on a "MedFuzzed" benchmark, as well as individual successful attacks. The methods show promise at providing insights into the ability of an LLM to operate robustly in more realistic settings.

cs.CL↗

arXiv2024

"Is This It?": Towards Ecologically Valid Benchmarks for Situated Collaboration

We report initial work towards constructing ecologically valid benchmarks to assess the capabilities of large multimodal models for engaging in situated collaboration. In contrast to existing benchmarks, in which question-answer pairs are generated post hoc over preexisting or synthetic datasets via templates, human annotators, or large language models (LLMs), we propose and investigate an interactive system-driven approach, where the questions are generated by users in context, during their interactions with an end-to-end situated AI system. We illustrate how the questions that arise are different in form and content from questions typically found in existing embodied question answering (EQA) benchmarks and discuss new real-world challenge problems brought to the fore.

cs.MM↗

arXiv2024

Reading Between the Lines: Modeling User Behavior and Costs in AI-Assisted Programming

Code-recommendation systems, such as Copilot and CodeWhisperer, have the potential to improve programmer productivity by suggesting and auto-completing code. However, to fully realize their potential, we must understand how programmers interact with these systems and identify ways to improve that interaction. To seek insights about human-AI collaboration with code recommendations systems, we studied GitHub Copilot, a code-recommendation system used by millions of programmers daily. We developed CUPS, a taxonomy of common programmer activities when interacting with Copilot. Our study of 21 programmers, who completed coding tasks and retrospectively labeled their sessions with CUPS, showed that CUPS can help us understand how programmers interact with code-recommendation systems, revealing inefficiencies and time costs. Our insights reveal how programmers interact with Copilot and motivate new interface designs and metrics.

cs.SE↗

arXiv2024

When to Show a Suggestion? Integrating Human Feedback in AI-Assisted Programming

AI powered code-recommendation systems, such as Copilot and CodeWhisperer, provide code suggestions inside a programmer's environment (e.g., an IDE) with the aim of improving productivity. We pursue mechanisms for leveraging signals about programmers' acceptance and rejection of code suggestions to guide recommendations. We harness data drawn from interactions with GitHub Copilot, a system used by millions of programmers, to develop interventions that can save time for programmers. We introduce a utility-theoretic framework to drive decisions about suggestions to display versus withhold. The approach, conditional suggestion display from human feedback (CDHF), relies on a cascade of models that provide the likelihood that recommended code will be accepted. These likelihoods are used to selectively hide suggestions, reducing both latency and programmer verification time. Using data from 535 programmers, we perform a retrospective evaluation of CDHF and show that we can avoid displaying a significant fraction of suggestions that would have been rejected. We further demonstrate the importance of incorporating the programmer's latent unobserved state in decisions about when to display suggestions through an ablation study. Finally, we showcase how using suggestion acceptance as a reward signal for guiding the display of suggestions can lead to suggestions of reduced quality, indicating an unexpected pitfall.

cs.HC↗

arXiv2024

Now, Later, and Lasting: Ten Priorities for AI Research, Policy, and Practice

Advances in artificial intelligence (AI) will transform many aspects of our lives and society, bringing immense opportunities but also posing significant risks and challenges. The next several decades may well be a turning point for humanity, comparable to the industrial revolution. We write to share a set of recommendations for moving forward from the perspective of the founder and leaders of the One Hundred Year Study on AI. Launched a decade ago, the project is committed to a perpetual series of studies by multidisciplinary experts to evaluate the immediate, longer-term, and far-reaching effects of AI on people and society, and to make recommendations about AI research, policy, and practice. As we witness new capabilities emerging from neural models, it is crucial that we engage in efforts to advance our scientific understanding of these models and their behaviors. We must address the impact of AI on people and society through technical, social, and sociotechnical lenses, incorporating insights from a diverse range of experts including voices from engineering, social, behavioral, and economic disciplines. By fostering dialogue, collaboration, and action among various stakeholders, we can strategically guide the development and deployment of AI in ways that maximize its potential for contributing to human flourishing. Despite the growing divide in the field between focusing on short-term versus long-term implications, we think both are of critical importance. As Alan Turing, one of the pioneers of AI, wrote in 1950, "We can only see a short distance ahead, but we can see plenty there that needs to be done." We offer ten recommendations for action that collectively address both the short- and long-term potential impacts of AI technologies.

cs.CY↗

arXiv2024

AI-Powered Reminders for Collaborative Tasks: Experiences and Futures

Email continues to serve as a central medium for managing collaborations. While unstructured email messaging is lightweight and conducive to coordination, it is easy to overlook commitments and requests for collaborations that are embedded in the text of free-flowing communications. Twenty-one years ago, Bellotti et al. proposed TaskMaster with the goal of redesigning the email interface to have explicit task management capabilities. Recently, AI-based task recognition and reminder services have been introduced in major email systems as one approach to managing asynchronous collaborations. While these services have been provided to millions of people around the world, there is little understanding of how people interact with and benefit from them. We explore knowledge workers' experiences with Microsoft's Viva Daily Briefing Email to better understand how AI-powered reminders can support asynchronous collaborations. Through semi-structured interviews and surveys, we shed light on how AI-powered reminders are incorporated into workflows to support asynchronous collaborations. We identify what knowledge workers prefer AI-powered reminders to remind them about and how they would like to interact with these reminders. Using mixed methods and a self-assessment methodology, we investigate the relationship between information workers' work styles and the perceived value of the Viva Daily Briefing Email to identify users who are more likely to benefit from AI-powered reminders for asynchronous collaborations. We conclude by discussing the experiences and futures of AI-powered reminders for collaborative tasks and asynchronous collaborations.

cs.HC↗

arXiv2023

The Rise of the AI Co-Pilot: Lessons for Design from Aviation and Beyond

The fast pace of advances in AI promises to revolutionize various aspects of knowledge work, extending its influence to daily life and professional fields alike. We advocate for a paradigm where AI is seen as a collaborative co-pilot, working under human guidance rather than as a mere tool. Drawing from relevant research and literature in the disciplines of Human-Computer Interaction and Human Factors Engineering, we highlight the criticality of maintaining human oversight in AI interactions. Reflecting on lessons from aviation, we address the dangers of over-relying on automation, such as diminished human vigilance and skill erosion. Our paper proposes a design approach that emphasizes active human engagement, control, and skill enhancement in the AI partnership, aiming to foster a harmonious, effective, and empowering human-AI relationship. We particularly call out the critical need to design AI interaction capabilities and software applications to enable and celebrate the primacy of human agency. This calls for designs for human-AI partnership that cede ultimate control and responsibility to the human user as pilot, with the AI co-pilot acting in a well-defined supporting role.

cs.HC↗