Marika Niihori
Mihaela van der SchaarICML 2026 is an exciting moment for the van der Schaar Lab. The thirteen papers accepted this year reflect not only the outstanding work of our students and collaborators, but also a set of research directions that we believe are becoming increasingly important across machine learning, bio-science, and medicine.
Viewed together, the papers naturally cluster around five research frontiers the lab has been building towards over several years: AI for Medicine, AI for Bioscience, Causality, Next-generation AI agents (Genies), and LLMs Evaluation and Improvement.
While these themes do not capture the full scope of the van der Schaar Lab’s research, they reflect several areas becoming increasingly important within the wider international research landscape. In the lead-up to ICML 2026, we will use this series to explore each of these directions in more detail, introducing the papers, highlighting the researchers and collaborators behind them, and examining why these areas are generating so much excitement across machine learning and science more broadly.
AI for Bioscience
One of the most exciting directions emerging from this year’s ICML papers is AI for bioscience. This is becoming a major area of focus for the van der Schaar Lab, as it brings together two things that are often still separated: cutting-edge machine learning and the scientific process of discovering biological mechanisms.
Biology is full of difficult ML problems. We often do not observe the system directly. Instead of continuous trajectories; we see destructive snapshots, instead of clean mechanisms we observe high-dimensional and noisy measurements; instead of tracking the same cell over time, we measure populations at isolated timepoints. Yet these are exactly the settings where new machine learning ideas can have the greatest impact. If we can build models that help scientists infer hidden biological dynamics, design better experiments, discover interpretable equations, and reason about interventions, then ML becomes more than a prediction tool. It becomes a partner in scientific discovery itself.
Three ICML papers from the lab illustrate this new direction.
CellBRIDGE addresses a central challenge in single-cell biology: how can we infer cellular trajectories when single-cell RNA sequencing destroys the cells being measured, so that the same cell cannot be followed over time? Standard optimal transport methods align snapshots mostly by gene-expression similarity, but this treats cells as independent points and ignores the fact that cell populations evolve through communication. CellBRIDGE changes this by incorporating directed, typed ligand–receptor interactions into the alignment itself. In other words, it does not only ask which cells look similar across time; it asks which cells play similar roles in the communication structure of the population. This is an important ML contribution because it reframes the cost function in biological optimal transport as a place where mechanistic biological knowledge can enter the model. It is also important biologically because it allows inferred trajectories to reflect cell–cell signalling, not just expression geometry. The paper shows that these interaction-aware couplings improve trajectory inference across several downstream continuous-time models, and also enable interpretable in silico perturbations, where silencing specific ligand–receptor pairs in lung cancer data induces trajectory shifts consistent with targeted pathway inhibition.
Active Timepoint Selection for Learning Measure-Valued Trajectories tackles another fundamental bottleneck in experimental biology: we often cannot measure everything. In single-cell experiments, each timepoint can be expensive and destructive, so the key question becomes not only how to model a biological trajectory, but when to measure it. This paper formulates active learning for measure-valued trajectories, where each observation is not a single vector but an empirical distribution of cells. This is technically challenging because probability distributions live in a non-Euclidean Wasserstein space, and standard active learning methods do not directly apply. The paper addresses this by using Linearized Optimal Transport to map snapshots into a tangent space, building a warped Gaussian Process surrogate, and selecting new timepoints to reduce uncertainty about the whole trajectory. This is exciting because it brings together optimal transport, uncertainty quantification, active learning, and experimental design in a setting that is directly relevant to biology. It gives ML a concrete role in making biological experiments more informative, especially when dense sampling is too costly or impossible.
Influence-Guided Symbolic Regression (IGSR) asks a different but equally important question: can AI help discover interpretable biological mechanisms, not just fit black-box predictors? Symbolic regression has long been attractive for science because it can produce compact equations, but searching over equations is hard, especially in high-dimensional biological settings. LLMs can propose plausible functional forms, but if they are guided only by a single global loss, they do not know which parts of an equation are useful and which are noise. IGSR introduces granular per-term influence scores, so that each proposed symbolic term receives feedback about its marginal contribution to generalization. This creates a propose-and-prune loop in which the LLM generates candidate basis functions, statistical influence scores select the useful ones, and Monte Carlo Tree Search explores the space of possible equations. What makes this especially exciting for bioscience is that the paper does not stop at benchmark performance. In a high-dimensional biological case study, IGSR identified a relationship between DNA methylation and RNA Polymerase II pausing, which was subsequently supported by wet-lab experimentation. This is exactly the kind of loop we are excited about: ML generates hypotheses, statistical structure filters them, and biology tests them.
Together, these papers show why AI for bioscience is such an important and exciting frontier for the lab. The common thread is not simply ‘apply AI to biological data’. The deeper idea is to build ML methods that respect how biology is observed and studied. CellBRIDGE brings biological interaction structure into trajectory inference. Active Timepoint Selection brings uncertainty-aware experimental design into the space of cellular distributions. IGSR brings LLM-driven search together with rigorous statistical feedback to discover interpretable mechanisms that can be tested experimentally.
We see AI for bioscience as one of the most promising directions for the next stage of machine learning research. Biology presents some of the hardest and most meaningful ML problems: partial observation, sparse experiments, interacting systems, high-dimensional measurements, hidden dynamics, and the need for interpretable mechanisms. At the same time, advances in ML, including optimal transport, active learning, uncertainty quantification, LLM-driven scientific search, and causal or mechanistic modelling, are creating new ways to approach biological discovery. What makes this area especially exciting is the possibility that AI may not only analyse biological data after experiments are completed, but also help scientists decide what to measure, how to infer dynamics, which mechanisms may be operating, and what hypotheses should be tested next.
AI for Medicine and Healthcare
AI for medicine has long been central to the van der Schaar Lab’s research agenda, reflecting the group’s broader commitment to building AI systems that remain grounded in real-world impact. Across healthcare, the lab has consistently focused on developing methods that do not simply optimise benchmark performance, but can meaningfully support clinical reasoning, treatment planning, and patient-specific decision-making.
This year’s ICML papers in this area focus on two complementary questions: how can digital twins be trained to support better treatment decisions, and how can machine learning help improve the clinical scoring systems used every day in medical practice?
In Decision-Targeted Digital Twins (DT²), the researchers study how digital twins can better support decision-making in healthcare and other high-stakes settings. While digital twins are often intended to guide decisions, most machine learning-based approaches are still trained primarily to minimise simulation error, optimising whether trajectories appear accurate rather than whether the system correctly ranks treatment or policy choices.
To address this, the paper introduces DT², a new decision-targeted DT training paradigm for digital twins. Instead of focusing purely on trajectory reconstruction, DT² trains the model to preserve meaningful rankings between candidate policies using off-policy evaluation methods and ranking-aware objectives. In a cancer treatment case study involving five distinct therapeutic strategies, DT² produces substantially better treatment ranking performance than conventional digital twin training while maintaining nearly the same simulation fidelity. This highlights an important and very practical insight: a digital twin’s job is not just to simulate reality, but to guide decisions well.
Alongside this, AgentScore explores how clinical scoring systems themselves can be improved. Modern healthcare still relies heavily on compact rule-based scoring systems because they are interpretable, auditable, and easy to deploy in clinical workflows. Yet constructing these systems has traditionally depended on manually designed heuristics and rigid rules.
AgentScore proposes a framework in which LLMs generate candidate scoring rules, which are then refined through verification and constrained optimisation loops to produce compact, interpretable checklist models aligned with real clinical workflows. Rather than replacing clinical guidelines, the work demonstrates how machine learning can be used to systematically improve them, achieving performance comparable to more flexible interpretable models despite operating under far stronger structural constraints.
Together, these projects reflect a broader shift toward clinically grounded AI systems designed not only to predict outcomes, but to support reliable and actionable decision-making.
Causality
Causality has been a longstanding foundation of research within the van der Schaar Lab. This year’s ICML papers continue that agenda through work on treatment effect estimation and causal discovery, combining theoretical advances with methods designed for practical, high-dimensional settings.
Identifiable Nonlinear Differentiable Causal Discovery via Independence and Adaptive Group Sparsity explores how causal structure can be recovered from observational data without relying on restrictive linearity assumptions. The paper combines independence-based optimisation with adaptive group sparsity to make differentiable causal discovery more robust in realistic nonlinear settings.
Gradient-Based Causal Tree Ensembles: A Backbone Architecture for Heterogeneous Treatment Effects introduces a new architecture for estimating heterogeneous treatment effects- the question of who benefits from a treatment, and by how much. By combining the flexibility of gradient-based learning with the robustness of tree-based methods, GRACE provides a powerful new framework for treatment effect estimation on complex real-world data.
Together, these papers show why causality remains central to the lab’s research agenda: many of the most important questions in medicine, science, and policy are not only about what will happen, but about what would happen if we intervened differently.
Genies, Next-Generation AI Agents
Next-generation AI agents have been a longstanding research direction within the van der Schaar Lab. At this year’s ICML, the lab explores two important questions at the frontier of this area: how agents can continually acquire and compose new skills over time, and how large populations of AI agents may eventually compete, coordinate, and behave within emerging economic systems.
Skill Neologisms: Towards Skill-based Continual Learning explores a central question for the future of AI agents: can large language models learn new composable skills without updating their core model weights? The paper introduces ‘skill neologisms‘ which are vocabulary-integrated soft tokens trained on skill-centric data that allow models to acquire new procedural capabilities while preserving existing behaviours. More broadly, the work presents a vision of continual learning in which agents evolve over time not by repeatedly retraining from scratch, but by inventing new skills.
While AI Agents Compete for Jobs shifts the conversation around AI agents away from simply ‘how do agents work?’ towards a broader and increasingly important question: what happens to labour markets when they do? The paper introduces AI-Work, a simulated gig economy where LLM agents compete for jobs, build reputations, acquire skills, and adapt strategies under uncertainty and competitive pressure. Unlike human workers, AI agents can operate across multiple jobs simultaneously, learn rapidly, and function without traditional labour constraints, creating entirely new economic dynamics. Through this environment, the researchers study which strategic capabilities allow agents to succeed, including metacognition, competitive awareness, and long-horizon planning. More broadly, the work begins to explore the microeconomics of AI-only labour markets and what emerging agentic economies may look like as AI systems increasingly interact, compete, and coordinate with one another autonomously.
The two papers reflect the lab’s broader reality-centric approach to AI research: asking not only how increasingly capable AI systems can be built, but how they will behave, adapt, and interact within real-world environments, institutions, and economies.
LLMs Evaluation and Improvements
The lab approaches large language models from a different angle than much of mainstream AI research. Rather than focusing primarily on scaling models larger and larger, the emphasis is on how language models can become more trustworthy, controllable, and rigorously evaluated. Two recent papers illustrate this direction particularly clearly.
One paper this year, No More, No Less: Least-Privilege Language Models, explores least-privilege language models: systems that expose only the capabilities needed for a specific request, rather than granting full access to everything a model can do. The work introduces a framework for limiting internal computation at deployment time, controlling which capabilities are reachable during a model’s forward pass. This creates a new way to balance utility, safety, and capability exposure, allowing the same underlying model to behave differently depending on context or user needs. More broadly, the paper proposes a new deployment paradigm for language models, challenging the assumption that control can only exist at the output level of the model.
Nonparametric LLM Evaluation from Preference Data studies how language models can be compared and ranked using human preference data in a statistically rigorous way. Rather than relying on restrictive parametric assumptions, the paper introduces a nonparametric framework for estimating ranking scores and collecting preference data efficiently. Human preference comparisons are increasingly central to how language models are evaluated in practice, and the framework provides practitioners with powerful, state-of-the-art methods for comparing or ranking LLMs for leaderboards.
The two papers in this frontier reflect a broader shift in language model research.: the challenge is no longer only to make LLMs more capable, but also to understand how they should be evaluated, governed, and deployed in ways that are reliable, transparent, and controllable.
Looking Ahead
These five research frontiers reflect several of the major directions currently shaping both the van der Schaar Lab and the wider future of machine learning research. In the lead up to ICML 2026, we’re excited to share more about these works, diving deeper into the ideas, collaborations, and researchers behind them, and exploring why these directions are generating so much excitement across AI, science, and medicine.
Most of all, we hope this series will make it easier for others to engage with the papers, build on the ideas, and join the broader conversation about where machine learning is heading next.
