Marika NiihoriReality-Centric AI Agents: Reasoning, Learning and Deciding in the Real World
Each year, the Conference on Neural Information Processing Systems (NeurIPS) brings together leading researchers from around the world to explore breakthroughs in machine learning, AI and computational science. It’s one of the field’s most prestigious conferences, taking place next month in San Diego.
This year, the van der Schaar Lab will be presenting six papers.
Our work pushes towards reality-centric AI Agents and systems that can reason, adapt and collaborate with humans in dynamic, uncertain environments. Below, we highlight three papers that capture a reality-centric vision, and how these advances can be applied to build more intelligent, trustworthy systems in healthcare and beyond.
Simulating Viva Voce Examinations to Evaluate Clinical Reasoning in Large Language Models
Clinical reasoning lies at the core of medical decision-making. It is a dynamic, hypothesis-driven process where clinicians iteratively gather and interpret information to reach a diagnosis. Yet, most existing benchmarks for large language models (LLMs) evaluate only static knowledge recall: models are given complete information upfront and rewarded for correct answers, without assessing how they reason under uncertainty.
To bridge this gap, we developed VivaBench, a large-scale standardised benchmark designed to evaluate the sequential reasoning abilities of LLMs in realistic diagnostic contexts. Inspired by oral “viva voce” medical examinations, the VivaBench comprises over 1,700 physician-curated, interactive vignettes where models must question, test, and synthesize information across multiple steps to reach a diagnosis. Each case requires active decision-making, asking relevant questions, choosing appropriate investigations, and revising initial hypotheses as new findings emerge.
Figure 1: Action and reasoning trace of two evaluated models on our simulated viva voce examination. Given the initial scenario (green), agents (blue) are tasked to diagnose the patient (orange).
Our results show that while state-of-the-art LLMs possess strong factual knowledge, performance degrades significantly when reasoning with incomplete or evolving information. Models display systematic errors similar to those observed in clinical practice, including fixation on initial hypotheses, excessive or inappropriate test ordering, premature diagnostic closure and failure to screen critical conditions. These patterns highlight that despite their fluency, current LLMs struggle to manage uncertainty and gather information sequentially, lacking the structured reasoning characteristic of expert clinical judgment.
We developed a benchmark to test whether LLMs can perform clinical reasoning where they have to actively gather information over multiple steps. We observe that current models possess the medical knowledge to diagnose patients, but struggle with asking the right questions and navigating uncertainty in this simulated clinical setting.
— Chris Chiu
By establishing a standardised framework for evaluating diagnostic reasoning, VivaBench provides foundation for advancing reality-centric AI. Systems that can reason interactively, manage uncertainty and support informed decision-making. Beyond clinical applications, it provides a controlled environment for studying how AI agents reason, plan, and adapt in complex real-world scenarios.
Timely Clinical Diagnosis through Active Test Selection
The next step toward reality-centric AI is enabling models not only to reason, but to take action – deciding what information to seek and when. In clinical settings, this means choosing the right tests at the right time, under uncertainty and resource constraints.
Diagnosis is inherently dynamic: clinicians gather evidence step by step, deciding which questions to ask and which investigations add value. However, most machine learning systems still treat diagnosis as a static classification task, assuming a complete dataset and overlooking the practical realities of uncertainty, cost, and time.
ACTMED addresses this challenge by combining Bayesian Experimental Design (BED) with LLMs to create an adaptive diagnostic framework that reflects how clinicians reason in practice. At each step, the system identifies the test expected to yield the greatest reduction in diagnostic uncertainty for a given patient, while accounting for resource constraints and clinical priorities. The LLM acts as a flexible simulator, generating plausible patient states and updating beliefs as new information becomes available, without task-specific training.
Figure 2: Overview of ACTMED.
Clinicians remain integral to the process: they review test suggestions, interpret intermediate results, and guide the model’s next move. This ensures transparency and interpretability, keeping human expertise central to the diagnostic loop. In evaluations on real-world, ACTMED improved both accuracy and efficiency, demonstrating how probabilistic reasoning and LLMs can support adaptive, data-efficient clinical-decision making without losing the nuance of clinical expertise.
ACTMED rethinks clinical AI by treating diagnosis as a dynamic process of information acquisition and uncertainty reduction, much like how clinicians’ reason through each new test result. In contrast, most existing models make static predictions from a fixed set of inputs, without adapting as new information becomes available
— Silas Ruhrberg Estévez
ACTMED represents a step toward probabilistic, uncertainty-aware AI systems that move beyond interpreting data to actively deciding what information to gather and when. By enabling clinician-in-the-loop decision-making, such frameworks can enhance interpretability, optimise resource use, and improve diagnostic quality. More broadly, approaches like ACTMED lay the groundwork for reality-centric AI – systems that plan, reason, and adapt alongside humans in complex, uncertain environments.
Cascaded Language Models for Cost-Effective Human–AI Decision-Making
As multi-agent AI systems increasingly take part in real-world decision-making, a key question arises: when should they act autonomously, when should they defer to a more capable model, and when should they abstain and let a human expert decide?
In fields such as healthcare, finance, and education, these decisions come with trade-offs. Powerful models can achieve high accuracy but are costly to run, while smaller ones are faster and cheaper yet less reliable on difficult or ambiguous cases. Effective systems need a way to balance accuracy, cost, and uncertainty, deciding how much expertise to involve in each situation.
Our work introduces a cascaded decision framework designed to do exactly that. It allows models of different capabilities to collaborate with each other and with human experts. Two key elements underpin this process: deferral and abstention. The deferral policy decides whether to accept a smaller model’s answer or pass the question to a larger model when confidence is low. The abstention policy determines when uncertainty is high enough that the system should defer to a human. Finally, an online learning component adapts and improves these policies over time, using feedback to adapt as the system encounters new data.
Figure 3: Decision flow of the two-tiered cascaded LLM system.
Evaluated across general and medical question-answering benchmarks, this framework improved both accuracy and efficiency, illustrating how structured deferral and abstention can make multi-agent AI systems more reliable and cost-effective.
Our work introduces a cascaded LLM framework that adaptively delegates tasks across a cheap base model, a more powerful (yet more expensive) large model, and human experts to balance prediction correctness, costs, and the decision of when to abstain. Interestingly, we then utilise the expert feedback to enhance the decision-making framework by incorporating online learning.
— Claudio Fanconi
By enabling AI systems to recognise when they are uncertain or out of their depth, this work takes a step toward more reliable, human-aligned decision support. Rather than acting alone, these systems learn to make context-aware choices about when to act, when to escalate, and when to defer to human judgement.
Together, these works show our progress toward reality-centric, human-empowering AI -systems that can reason, act, and decide responsibly in complex, uncertain settings.
Beyond these highlights, the van der Schaar Lab will also present additional papers and posters at NeurIPS 2025, spanning topics from digital twins and semantic similarity benchmarking.
