Searcharxiv⌕ Search

arXiv subjects

Marinka Zitnik

Publications and source records attributed to Marinka Zitnik.

At least 19 recordsLinked to original sources

AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation

Scientific research proceeds through iterative cycles of hypothesis generation, experiment design, execution, and revision. AI agents can automate parts of this process, but existing approaches typically follow a single research trajectory or coordinate through a central planner with fixed objectives. As a result, they struggle to sustain parallel exploration, adapt as experimental evidence changes, or preserve knowledge of failed directions over long-running experiments. We introduce AutoScientists, a decentralized team of AI agents for long-running computational scientific experimentation. Agents interpret a shared experimental state, self-organize into teams around promising hypotheses, critique proposals before using experimental compute, and share successes and failures to reduce redundant exploration. Under matched experimental budgets, AutoScientists improves over prior AI agents across biomedical machine learning, language-model training optimization, and protein fitness prediction. On BioML-Bench, spanning biomedical imaging, protein engineering, single-cell omics, and drug discovery, AutoScientists achieves a mean leaderboard percentile of 74.4% across 24 tasks, improving over the strongest AI agent by +8.33%. On GPT training optimization, AutoScientists reaches a target validation bits-per-byte 1.9x faster than Autoresearch and continues discovering improvements from a starting champion where the single-agent approach finds none (7 vs. 0 accepted improvements). On ProteinGym fitness prediction, AutoScientists discovers a method for ACE2-Spike binding that improves over the current state-of-the-art model by +12.5% in Spearman correlation. Applied without modification across all 217 ProteinGym assays, the same method improves over the prior state of the art by +6.5% (Spearman correlation).

cs.AI↗

Does Predictive Performance Translate to Causal Performance? Evaluating tabular foundation models as nuisance learners for causal effect estimation

Tabular foundation models based on prior-data fitted networks (PFNs) achieve strong predictive performance, but their value for causal inference is unclear. We evaluated whether PFNs, which predict through in-context learning, could match or outperform conventional nuisance learners in supporting accurate causal estimation and reliable uncertainty quantification (inference). Learners modeled propensity scores and outcome regressions within doubly robust estimators of average causal effects, with and without cross-fitting, across synthetic and semi-synthetic settings with varying confounding complexity and effect heterogeneity. Under simple parametric data-generating processes, the PFNs were competitive with correctly specified parametric regressions and ensembles containing them. In more complex settings, they outperformed all conventional learners evaluated, including richly specified Super Learner ensembles, for both accuracy and uncertainty quantification. As sample size increased, only the two PFN models (TabPFN and Mitra) consistently approached nominal confidence interval coverage, whereas under-coverage persisted for conventional learners. In some settings, PFNs achieved near-nominal coverage and well-calibrated standard errors even without cross-fitting. At the smallest sample size (n=200) and in the higher-dimensional semi-synthetic settings, PFNs and other flexible learners exhibited poor coverage without cross-fitting. For PFNs, this appeared to be driven by propensity score overfitting consistent with in-sample memorization of treatment labels. These findings highlight the promise of prior-data fitted networks as in-context nuisance learners for causal estimation, while underscoring the need for further evaluation in challenging settings.

stat.AP↗

Hindsight Relabeling is All You Need for Reach-Avoid Learning

Offline goal-conditioned reinforcement learning methods have shown promise for reach-avoid tasks, where an agent must reach a target state while avoiding undesirable regions of the state space. Existing approaches typically encode avoid-region information into an augmented state space and cost function, which prevents flexible, dynamic specification of novel avoid-region information at evaluation time. They also rely heavily on meticulously designed reward and cost functions, limiting transferability to novel environments and specifications. We introduce RADT, a decision transformer approach for offline, reward-free, goal-conditioned, avoid-region-conditioned RL. RADT encodes goals and avoid-regions directly as prompt tokens, allowing any number of avoid-regions of arbitrary size to be specified at evaluation time. Using only suboptimal offline trajectories from a random policy, RADT can learn reach-avoid behavior in a completely data-driven manner using our novel avoid-region hindsight relabeling approach, without the need for reward/cost supervision. We benchmark RADT against existing offline goal-conditioned RL models across 17 tasks, environments, and experimental settings. RADT generalizes in a zero-shot manner to out-of-distribution avoid-region sizes and counts, outperforming baselines that require retraining. In one such zero-shot setting, RADT achieves 35.7% improvement in normalized cost over the best retrained baseline while maintaining high goal-reaching success. We also apply RADT to cell reprogramming in biology, demonstrating its versatility.

cs.LG↗

Learning developmental dynamics from lineage context

A central goal of biology is to predict the dynamics of multicellular systems. Lineage context (the states of a cell's sisters, cousins, etc.) could inform these predictions, but how best to leverage it remains unclear. Here we introduce lineage inpainting, a generative modeling approach that uses lineage context with tree-structured transformers and stochastic interpolants to model cell state dynamics across lineage trees. Applied to a lineage-traced transcriptional atlas of mouse embryogenesis, it recovers dynamics not resolved by gene expression alone, including multiple transcriptional convergence events where distinct lineages are drawn to a common state. Despite being fit on unperturbed development, lineage inpainting can generate hypotheses about perturbation responses by simulating counterfactuals. These simulations agree in many cases with observed developmental phenotypes, including a single-cell measurement of the Tbx5 knockout response. Our results show that lineage information provides a rich source of context for learning multicellular dynamics.

q-bio.QM↗

Latent Decision Flow Rewards for Clinical Guideline Reasoning

Clinical practice guidelines encode evidence-based recommendations as conditional decision flows, which makes them a natural source of process supervision for training reasoning models. Two properties of clinical reasoning make this supervision difficult to use. A decision flow consistent with a guideline does not by itself determine the correct answer for a given patient, which we call the path--answer gap, and several valid decision flows can support the same clinical decision, which we call path multiplicity. Current training approaches address parts of this problem. Answer rewards supervise the final decision, and process rewards computed against a reference flow supervise one valid decision flow per case. Supervised fine-tuning on annotated flows teaches guideline-consistent reasoning but constrains the model to one trajectory per case and can reduce answer accuracy. We introduce GUIDER, a method that jointly optimizes a decision flow controller and a reasoning policy. The controller learns to generate plausible decision flows from guideline-derived supervision, and the policy learns to generate a decision flow and a final answer for each patient case. GUIDER computes the process reward for the policy over several controller-generated flows, which accounts for path multiplicity, and gates this reward on answer correctness, which addresses the path--answer gap. Joint training allows the process supervision to evolve with the policy. On MedGUIDE, GUIDER with Qwen3-14B achieves 82.5 weighted accuracy and 73.5 guideline adherence, 3.3 and 3.1 points above the strongest clinical reinforcement learning baseline. With Qwen3-14B and Gemma4-12B backbones, GUIDER also improves over CRPO, MRPO, and FaithMed on the out-of-domain benchmarks AMEGA-LLM, RealMedQA, and CancerGUIDE.

cs.LG↗

STRAND: Sequence-Conditioned Transport for Single-Cell Perturbations

Prediction of the transcriptional response of a cell to a genetic perturbation is a central problem in the computational modeling of gene regulation. Perturbations of the same gene can produce different transcriptional responses depending on the targeted locus, such as an alternative transcription start site or a distal regulatory element, and gene-level perturbation models represent these distinct interventions with a single gene identifier. We introduce STRAND, a generative model that encodes the regulatory DNA sequence at the perturbation locus and uses this encoding to parameterize a conditional transport process from control to perturbed single-cell states. Sequence conditioning enables zero-shot prediction for perturbation targets absent from training and supports locus-resolved inputs at 128 bp model resolution within a 524 kb sequence window. On CRISPRi screens in K562, Jurkat, and RPE1 cells, STRAND achieves the best average rank on few-shot and unseen-gene benchmarks, with gains of up to 33% in perturbation discrimination, and improves transfer to held-out cell lines by up to 0.12 in $Δ$Pearson. Ablation studies attribute these gains to sequence conditioning and to conditional transport. At the locus level, STRAND distinguishes alternative transcription start sites and validated regulatory elements that gene-identifier models represent as a single perturbation. STRAND is available at https://github.com/mims-harvard/STRAND.

q-bio.GN↗

Agentic discovery of blood biomarker from distilled private health records

Routine complete blood counts (CBCs) could yield new biomarkers, but the private records needed to evaluate candidates cannot be shared with frontier language model agents that excel at discovery. We distilled the evidence held in the Clalit Health Services panel of over 5.4 million patients into a released scoring tool: for each of 13 immune-mediated diseases, a graph attention network was trained inside the data boundary to predict the case-control AUC of candidate CBC expressions, and only the trained weights were released. The tool grounds an agent's propose-score-refine loop in real-world data without exposing any patient data. In external validation, agent-discovered expressions improved on their literature-seeded starting points by a median of 4.18 AUC percentage points, and across three independent cohorts, reranking the candidates of three frontier research tools improved on their first choices in most comparisons, with gains that varied by cohort. The released scorer supports privacy-preserving biomarker hypothesis generation.

cs.AI↗

When Do Biological Reasoning Models Use Their Biological Inputs?

Biological reasoning models use post-training to connect LLMs to biological foundation model representations and biological text. Their benchmark accuracy is taken as evidence that LLMs reason over these inputs. We test this assumption in six biological reasoning models across DNA, protein, and single-cell tasks. We perturb one biological input while holding the query and other inputs fixed, construct evidence conflicts that pair the foundation model representation of one genome, protein, or cell with the text of another, fit linear probes to the representations the language model receives, and analyze reasoning traces against the biological inputs. Evo2 and ESM3 contribute little to BioReason and BioReason-Pro performance on the evaluated tasks. Shuffling the DNA sequence barely changes BioReason disease prediction accuracy, and in evidence conflicts the two models follow the text in 97.9% and 99.7% of cases. Linear probes trained on the Evo2 and ESM3 representations predict the task targets, so these foundation models encode information relevant to the task, but provide limited overall performance improvement to BioReason and BioReason-Pro. In contrast, foundation model inputs contribute to ChatNT, Prot2Text-V2, and CellWhisperer performance, and differentially expressed genes in the gene sentence contribute to Cell2Sentence-Scale performance. Across SFT and RL checkpoints of BioReason-Pro and 42 BioReason checkpoints, increases in accuracy do not imply greater performance contributions from biological inputs. BioReason traces misstate nucleotide changes, while BioReason-Pro traces describe functions omitted from final predictions under evidence conflicts. We find that current post-training strategies do not ensure that foundation model representations contribute to task performance.

cs.LG↗

Distribution-Conditioned Transport

Learning a transport model that maps a source distribution to a target distribution is a canonical problem in machine learning, but scientific applications increasingly require models that can generalize to source and target distributions unseen during training. We introduce distribution-conditioned transport (DCT), a framework that conditions transport maps on learned embeddings of source and target distributions, enabling generalization to unseen distribution pairs. DCT also allows semi-supervised learning for distributional forecasting problems: because it learns from arbitrary distribution pairs, it can leverage distributions observed at only one condition to improve transport prediction. DCT is agnostic to the underlying transport mechanism, supporting models ranging from flow matching to distributional divergence-based models (e.g. Wasserstein, MMD). We demonstrate the practical performance benefits of DCT on synthetic benchmarks and four applications in biology: batch effect transfer in single-cell genomics, perturbation prediction from mass cytometry data, learning clonal transcriptional dynamics in hematopoiesis, and modeling T-cell receptor sequence evolution.

cs.LG↗

Multimodal AI predicts clinical outcomes of drug combinations from preclinical data

Predicting clinical outcomes from preclinical data is essential for selecting safe and effective drug combinations and for reducing late-stage failures. AI models use molecular structure and target annotations, and do not leverage the perturbation readouts that report how a compound acts in a cellular context. Here we introduce Madrigal, a multimodal AI model that learns from structural, pathway, cell-viability, and transcriptomic data. Madrigal aligns these modalities across 21,842 compounds into a shared latent space and predicts combination outcomes even for drugs observed in only a subset of the data modalities. Trained on 158 expert-curated and 795 patient-reported combination outcomes, Madrigal outperforms single-modality and state-of-the-art multimodal methods. Ablations show that modality alignment and multimodal input each improve predictive performance. Madrigal predicts elevated risk for combinations that share membrane transporters. In head-to-head trials that compare two combination arms,the arm with the higher observed incidence of neutropenia, anemia, alopecia, or hypoglycemia receives the higher predicted risk in 25 of 28 comparisons. In MASH, Madrigal ranks resmetirom among the candidates with favorable predicted safety when paired with type 2 diabetes drugs. Madrigal also improves adverse-event prediction in a longitudinal patient cohort and an independent oncology cohort and predicts efficacy in primary acute myeloid leukemia samples and patient-derived xenografts.

q-bio.QM↗

Unifying biomedical knowledge in a modern multimodal graph

Biomedical knowledge graphs (KGs) are widely used in the life sciences, yet many are derived from unstructured documents and therefore lack schema-level constraints, whereas graphs assembled from structured resources are difficult to harmonize into a unified representation. We present OptimusKG, a multimodal biomedical labeled property graph (LPG) built from structured and semi-structured resources to preserve factual, type-specific metadata across molecular, anatomical, clinical, and environmental domains. OptimusKG contains 190,939 nodes across 10 entity types, 21,818,752 edges across 27 edge types, and 67,070,490 property instances encoding 109,665,797 values across 145 distinct property keys, derived from 18 ontologies and controlled vocabularies. The graph enforces a top-level schema for nodes and edges and retains granular, type-specific properties, cross-references, and provenance. We assessed the validity of OptimusKG by evaluating whether graph relationships are supported by evidence from the scientific literature using a multimodal agent, PaperQA3. PaperQA3 identified supporting evidence for 70.0% of sampled edges, whereas 83.4% of sampled false edges received no supporting evidence. Edges without literature support were concentrated in associations derived from experimental and functional genomics resources, suggesting that OptimusKG captures biomedical knowledge that may precede synthesis in the scientific literature. OptimusKG is distributed as Apache Parquet files, providing a standardized resource for graph-based machine learning, knowledge-grounded retrieval with large language models, and biomedical discovery use cases such as hypothesis generation.

cs.AI↗

Qworld: Question-Specific Evaluation Criteria for LLMs

Evaluating large language models (LLMs) on open-ended questions is difficult because response quality depends on the question's context. Binary scores and static rubrics fail to capture these context-dependent requirements. Existing methods define criteria at the dataset level or generate them in a single pass, which limits their ability to explore the evaluation space implied by each question. We introduce One-Question-One-World (Qworld), a method that generates question-specific evaluation criteria using a recursive expansion tree. Given a question, Qworld decomposes it into scenarios, perspectives, and fine-grained binary criteria through hierarchical and horizontal expansion. The resulting criteria specify what a high-quality answer must address for that question. On HealthBench, Qworld covers 89% of expert-authored criteria and generates 79% novel criteria validated by human experts. Experts rate Qworld criteria higher in insight and granularity than those produced by prior methods. When applied to 11 frontier LLMs on HealthBench and Humanity's Last Exam, Qworld reveals capability differences in dimensions such as long-term impact, equity, error handling, and interdisciplinary reasoning that coarse rubrics do not capture. By generating evaluation criteria for each question, Qworld enables assessment of LLM responses that is tailored to the question rather than based on fixed task-level criteria.

cs.CL↗

Do LLMs Know What to Ask and When? Evaluating Multi-Turn Information Seeking

When a user question is underspecified, a capable model should recognize that its context is insufficient, identify the missing information, ask for it, and respond only once that information determines a unique answer. We formalize multi-turn information seeking as solving a k-underspecified constraint satisfaction problem, where k is the number of variables jointly required to determine the target and therefore measures the degree of missing information. We instantiate the formulation in MT-InfoSeek, a controlled evaluation suite of 5,251 problems and 9,006 task instances spanning mathematics, logic, biology, medicine, and general knowledge. We evaluate models along three axes: what they ask, when they ask it, and how the acquired information affects the final answer. Performance degrades across models and domains as underspecification increases. Models recognize that additional information is needed but underestimate how much, and in logical problems at k = 2 they under-predict the degree of missing information about four times as often as they over-predict it. They also fail to identify a minimal sufficient set of queries, improve only marginally when given the true k, and often stop before acquiring sufficient information. In tasks with ordered dependencies, an incorrect query order reduces final accuracy even when the model eventually acquires all necessary information. We measure information seeking directly through final sufficiency, which records whether the acquired information determines the target independent of answer generation. This separation shows differences between models that final accuracy alone does not capture, and indicates that the ability to seek information over multiple turns is distinct from the ability to generate answers and is not measured by current LLM evaluations.

cs.AI↗

ToolUniverse: An open platform for democratizing AI scientists

AI scientists are emerging computational systems that serve as collaborative partners in discovery. These systems remain difficult to build because they are bespoke, tied to rigid workflows, and lack shared environments that unify tools, data, and analyses into a common ecosystem. We present ToolUniverse, an open platform for building AI scientists from any language or reasoning model across open- and closed-weight models. ToolUniverse standardizes how AI scientists identify and call tools through an AI-tool interaction standard, in which every tool declares its purpose in natural language, a typed schema for its inputs and outputs, and a backend-agnostic invocation format, and applies that standard to more than 2,700 scientific tools and over 130 research skills spanning machine learning models, datasets, APIs, and scientific packages for data analysis, knowledge retrieval, and experimental design. It automatically refines tool interfaces for correct use by AI scientists, generates new tools from natural language descriptions, iteratively optimizes tool specifications, and composes tools into agentic workflows. In case studies, ToolUniverse was used to create AI scientists that carried out end-to-end analyses in target assessment, chemical strategy, and clinical-trial safety. The open-source ToolUniverse is available at https://aiscientist.tools.

cs.AI↗

MatrAIx: Simulating the World with 8.3 Billion Persona Agents

Human evaluation of AI systems and digital products is costly, slow, and difficult to scale. Offline evaluations are more scalable but often abstract away human diversity and interactive behavior. We therefore introduce MatrAIx, a population-scale simulated-user evaluation infrastructure for testing AI systems and digital products with heterogeneous users. MatrAIx has three core components: First, Persona 8B contains 8.3 billion persona records represented by 1,290 categorical dimensions. Records are either sampled from a dependency graph that preserves correlated attributes or derived from human-authored profiles. We release a quality-filtered coreset of approximately 1 million personas, comprising 599,847 human-grounded and 400,000 synthetic records. Second, the MatrAIx Playground provides four environments in which diverse users evaluate and interact with digital products: Survey, AI Chatbot, Web, and App. Third, MatrAIx provides 1,010 application tasks spanning more than 25 domains, including Commerce, Software, Finance, and Healthcare. We conducted 18,189 evaluation trials across eight representative tasks. Persona agents were powered by three LLMs: Claude Opus 4.8, GPT 5.5, and Claude Haiku 4.5. The resulting feedback captures how decisions and preferences vary across persona backgrounds, including hesitation after a price increase, willingness to continue after an AI assistant fails, and latency tolerance. We conducted two main validation studies: First, a 400-trial controlled study evaluated persona adherence across ten behavioral attributes and all four environments. The declared behavior was expressed or correctly suppressed in 366 trials (91.5%). Second, human and LLM judges evaluated the extraction quality of human-grounded personas. Overall, MatrAIx provides an end-to-end infrastructure for evaluating AI systems and digital products with diverse simulated human users.

cs.AI↗

Kidney function and kidney failure prediction in a large multiethnic population

Background: Patients with chronic kidney disease (CKD) experience worsening kidney function and develop subsequent kidney failure at different rates. Accurate estimates of a patient's current and future kidney function are needed for optimal clinical decision-making. Methods: To compare current and previously recommended equations for estimating kidney function and risk of kidney failure across a large multiethnic population, we conducted a retrospective multicenter cohort study of primary care, acute care, and hospital settings. Our study population comprised 1,909,042 adults with at least one recorded serum creatinine measurement during 2012-2014, with follow-up until January 2025. Our primary outcomes were the area under the receiver operating characteristic curve and prevalence of CKD by stage across regions of origin. Findings: GFR estimates from the two race-stratified equations (2006 MDRD and 2009 CKD-EPI) were similarly calibrated. GFR estimates from the two race-neutral equations diverged, with 2021 EKFC producing lower GFR estimates and 2021 CKD-EPI producing higher GFR estimates compared to prior equations. The oldest equation, 2006 MDRD, was the most discriminative of kidney failure within 5 years while the newer European equation, 2021 EKFC, was the least discriminative, with AUROCs of 0.862 (95% CI, 0.855-0.869) and 0.846 (95% CI, 0.838-0.853), respectively. The age-adjusted prevalence of CKD varied by the choice of eGFR equation (ranging from 8.5% for 2021 CKD-EPI to 10.8% for EKFC) and across regions of birth (ranging from 8.9% for East Sub-Saharan Africa to 15.3% for South Asia). Interpretation: Using 10 years of follow-up for nearly 2 million individuals, our study demonstrates how newly developed US and European equations diverge from previously recommended equations with broad clinical and epidemiological implications.

q-bio.QM↗

Adaptive Time Series Reasoning via Segment Selection

Time series reasoning tasks often start with a natural language question and require targeted analysis of a time series. Evidence may span the full series or appear in a few short intervals, so the model must decide what to inspect. Most existing approaches encode the entire time series into a fixed representation before inference, regardless of whether or not the entire sequence is relevant. We introduce ARTIST, which formulates time-series reasoning as a sequential decision problem. ARTIST interleaves reasoning with adaptive temporal segment selection. It adopts a controller-reasoner architecture and uses reinforcement learning to train the controller role to select informative segments and the reasoner role to generate segment-conditioned reasoning traces and final answers. During inference, the model actively acquires task-relevant information instead of relying on a static summary of the full sequence. We use a novel hierarchical policy optimization approach for post-training that allows the model to excel in both segment selection and question-answering behavior. We evaluate ARTIST on six time-series reasoning benchmarks and compare it with large language models, vision-language models, and prior time-series reasoning systems. ARTIST improves average accuracy by 6.46 absolute percentage points over the strongest baseline. The largest gains appear on rare event localization and multi-segment reasoning tasks. Supervised fine-tuning improves performance, and reinforcement learning provides additional gains by optimizing question-adaptive segment selection. These results show that selective data use drives effective time-series reasoning.

cs.LG↗

How Post-Training Shapes Biological Reasoning Models

Scientific reasoning models for biology combine language models with foundation models trained on multimodal biological data, including DNA, RNA, and proteins. These models are built through post-training, yet how each stage shapes reasoning and generalization remains poorly understood. We study when post-training improves performance and when it induces over-specialization. Across genomics, transcriptomics, and proteins, we train and evaluate more than 100 biological reasoning models under controlled variation in backbone, continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL), measuring both in-domain (ID) and out-of-domain (OOD) performance. We find that each post-training stage reshapes generalization in a distinct way rather than contributing uniform gains. CPT improves downstream performance by aligning models with biological language. SFT consistently increases ID performance but causes OOD performance to peak early and decline as models fit the training distribution. RL, when applied to strong SFT checkpoints with aligned rewards, improves OOD performance and partially recovers generalization. These results show that biological reasoning does not improve monotonically with additional supervision or compute. Instead, performance depends on how training stages are composed. Under fixed post-training budgets, the strongest ID-OOD trade-off comes from brief SFT, larger RL allocations, and asymmetric adaptation capacity across stages.

cs.LG↗