PyTorchKR Papers

최신 논문

총 523,801편
태그·기간 등 상세 필터

카테고리 탐색

arXiv:2608.19908v2cs.IT

A Layered Simplex Architecture for Large Alphabets

A learning architecture induces a prior over its expressible models. Recent work links architectural success to a broad complexity spectrum: model probabilities spanning many orders of magnitude. Understanding how depth shapes these priors and prediction may help explain the success of neural networks and transformers. We study this connection in large-alphabet prediction through the *layered simplex architecture* (LSA), which multiplies independent uniform simplex vectors coordinate-wise and normalizes. At depth zero, the prior is concentrated on the uniform distribution; at depth one, the Bayesian predictor is Laplace's rule. Each additional layer shifts prior mass toward more concentrated distributions. We derive exact Bayesian marginal likelihood and regret expressions using reusable depth-specific kernels. Mixing over depths up to a cutoff logarithmic in alphabet size yields an explicit predictor with rigorous guarantees. For targets uniform on a specified sparse support, the mixture's prior complexity is asymptotic to the support-description cost at every fixed positive KL resolution, establishing a broad complexity spectrum. We also prove a regret lower bound for every label-invariant predictor: the expected description length of the discovered symbol set minus the source's uncertainty about its identity given its size. For Zipf sources in a growing-alphabet regime, the depth mixture attains this bound to leading order. Thus, the symbol-discovery rate determines the regret exponent. Although introduced to study depth, the LSA predictor competes with established large-alphabet methods, including Good--Turing, Ristad's law and an adaptive Dirichlet mixture, while substantially outperforming classical add-constant rules on concentrated sources.

arXiv:2608.19490v2cs.RO

Fine-Tuning VLAs with Self-Demonstrated Generative Control for Multi-Task Manipulation

State-of-the-art vision-language-action (VLA) models such as pi0.5 exhibit strong semantic understanding, instruction following and task behavior. However, when deployed on new robots, even minor mismatches in hardware configuration relative to pretraining can cause severe performance drops. Finetuning the VLA on in-domain expert data from the new embodiment improves performance on the expert task but leads to a loss in its original instruction following and behavioral priors. In this paper, we propose a self-supervised method that generates online interaction rollouts from the zero-shot VLA as additional training data for finetuning. Our experiments show this yields strong multi-task manipulation policies on the target robot that (1) inherit pretrained tasks distilled from the zero-shot VLA without expert demos for them, (2) improve success and generalist instruction following beyond expert-only fine-tuning on expert-teleoperated tasks, and (3) mitigate forgetting of held-out tasks from the pretrained policy, not part of fine-tuning. We demonstrate the success of our approach across test sets probing generalization on a real ALOHA robot and a new simulation benchmark in RoboTwin. Video results are available at https://self-supervised-control.pages.dev/

arXiv:2608.18085v3cs.CL

Persona-Guided LLM Agents for Task-Oriented Dialogue

Prior work has shown that large language models can express diverse personality traits in open-ended text generation, but it is unclear whether they can do so in goal-directed dialogue without harming task completion, and whether adapting to a user's personality improves the interaction. We study these questions in task-oriented dialogue (TOD) with a training-free framework in which a personality-conditioned user agent interacts with a system agent under three conditions: Neutral (no personality information), Try (adapt to cues observed in the dialogue), and Oracle (the user's trait label is given explicitly). Across GPT-4o, Qwen3-Next-80B, and Gemini 2.0 Flash, two Schema-Guided Dialogue domains, and ten Big Five poles (9,000 dialogues), each scored by two independent LLM judges, user agents express most traits while task performance remains high (89.94% completion, 85.27% success). The exceptions are introversive and closed users: these traits manifest mainly as absent behavior and are rarely detected. Both adaptive conditions improve pairwise user satisfaction and lower truthfulness, revealing a personalization-grounding trade-off, but only cue-responsive Try also improves constraint satisfaction and inform rate. How adaptation is implemented matters: Oracle's gains depend on how visibly the trait is expressed, and for disagreeable users the label leads Qwen3 and Gemini to mirror hostility. Try avoids this failure, its satisfaction gains do not depend on trait visibility, and it is preferred over Oracle overall. A 175-pair human study supports the reliability of the LLM judges: on clear-winner cases, annotators agree with the satisfaction judges more often than with each other (76.3% vs. 65.1% agreement). Our results suggest that responding to observable behavior is a more robust basis for social adaptation than conditioning on psychological labels.

arXiv:2608.13573v2cs.AI

A Year in LLM Serving: Workload Evolution, Caching and Load-Balancing

Large Language Model (LLM) serving has become a critical cloud workload, and realistic traces are essential for motivating and benchmarking serving systems. However, existing LLM serving workload studies remain limited in scale and scope. They often observe short time periods and provide limited visibility into how users interact with models in production. As a result, they do not fully capture how LLM serving workloads evolve over time or how user-model interactions shape production traffic. In this work, we further the understanding of real-world LLM serving workloads through both a global characterization and a longitudinal study of a one-year production trace from CompanyX. Unlike prior studies, our trace captures full production behavior across many models and users, including both popular and long-tail models. We analyze the workload from aggregate, temporal, model-level, and user-level perspectives, revealing workload evolution and user-model structure that are typically hidden behind aggregate views. To support future research, we publicly release the full one-year trace, enabling downstream studies of production behavior without relying on sampled or synthetically generated workloads. The trace is available at https://github.com/HarvardMadSys/chutes_workload.

arXiv:2608.12231v3cs.LG

An Efficient Minimax-Optimal Algorithm for Adversarial $m$-Set Bandits

We study adversarial combinatorial bandits with $m$-set actions, where at each round the learner selects $m$ out of $d$ items and observes only the aggregate loss of the selected items. The resulting action set contains $K=\binom{d}{m}$ elements and can therefore be exponentially large. We propose a computationally efficient algorithm that exploits this structure without explicitly enumerating the action set. Against adaptive non-anticipating adversaries, it guarantees, with probability at least $1-\delta$, regret against the best action in hindsight of $R_T =O(\sqrt{dT\log(K/\delta)}). $ This matches the high-probability regret bound of the finite-action EXP3 with Kiefer-Wolfowitz exploration algorithm of Zimmert and Lattimore, whose direct implementation may require exponential space. Our algorithm instead represents each sampling distribution with $d$ parameters and runs in polynomial time without enumerating the action set. This resolves an open problem posed by Maiti et al. We complement this upper bound with a matching high-probability lower bound. For all sufficiently small $\delta$ and all sufficiently large horizons, every randomized algorithm admits an adaptive non-anticipating adversary for which, with probability at least $\delta$, $R_T=\Omega( \sqrt{dT\log(K/\delta)}). $ Thus, the rate is minimax optimal up to universal constants in this regime. In particular, setting $m=1$ shows that the $\log K$ factor in the high-probability regret bound for ordinary $K$-armed bandits against adaptive non-anticipating adversaries is unavoidable. This strengthens the lower bound of Gerchinovitz and Lattimore, closing the remaining $\sqrt{\log K}$ gap between high-probability upper and lower regret bounds.

arXiv:2608.12035v3cs.CV

How Far from Clinical Deployment? Evaluating the Complete Unsupervised Domain Adaptation Pipeline in Medical Imaging

Deploying unsupervised domain adaptation (UDA) in clinical practice requires choosing which algorithm to use and which of its trained models to ship. However, the deployment (target) domain is unlabeled, so models cannot be evaluated directly on it, leaving it unclear which to select. We address this by evaluating the complete UDA pipeline, considering both adaptation and label-free selection together. Our study covers eleven clinically relevant cross-domain scenarios from nine medical imaging datasets, with ten UDA algorithms and 13 label-free selection methods (validators), evaluating over 80,000 trained models in total. By this, we find that a capable adapted model usually exists, but identifying it without target labels is difficult: the validator-selected models leave a large and structural target performance gap to the best available one, with no evaluated validator consistently reliable. Towards closing it, we explore two strategies, ensembling and a small target-labeling budget; both narrow this gap but do not close it entirely. Overall, deployable UDA depends on the complete pipeline; addressing the less explored selection step could bring much of current UDA closer to clinical use. Our code is open sourced at \href{https://xiongyiheng.github.io/complete-uda-pipeline-site/}{Project Page}.

arXiv:2608.10775v2cs.AI

SkillLens: Visual Skill Cards for Retrieval-Augmented GUI Action Prediction and On-Policy Distillation

Graphical user interface (GUI) agents can reuse past interaction experience. We study how illustrated procedures support action prediction and students deployed without retrieval. We introduce SkillLens, which represents procedures, applicability cues, and interface views in a shared Visual Skill Card (VSC) schema. Retrieved cards provide a frozen executor with reference evidence alongside the current screen. CardDistill uses this schema to construct privileged teacher context from labeled training steps for on-policy distillation; the student learns from native inputs and requires no retrieval at deployment. With matched training data and update counts, students trained with full-card teachers improve step success rate by an average of 5.04 percentage points over those trained with compact-text teachers across three training seeds on 119 tasks from ten websites absent from training. An independent runtime control uses training-source memory and pre-action retrieval queries: full cards achieve 12/64 correct actions (18.75%) versus 4/64 (6.25%) for text-only cards. These results establish the value of the complete VSC teacher-context package for retrieval-free students and provide evidence for pre-action retrieval of illustrated procedures.

arXiv:2608.10288v2cs.LG

Power law graph attention: exact generalization of scaled dot-product attention, empirical collapse at inference

The Large Language Model from Power Law Decoder Representations (PLDR-LLM) and its attention, Power Law Graph Attention (PLGA), replace the fixed bilinear form of scaled dot-product attention (SDPA) with a learned, input-generated bilinear operator $G_{\mathrm{LM}}$, built from a positive tensor $A_{\mathrm{LM}}$ by elementwise power laws. The architecture is fully specified, verified against pinned reference releases; claims are labeled theorem, conditional theorem, measurement, or conjecture. Unconditionally: PLGA contains SDPA exactly at $G_{\mathrm{LM}}=I$; $A_{\mathrm{LM}}$ and $A_{P}$ are strictly entrywise positive, with Perron-Frobenius structure on $A_{\mathrm{LM}}$; the DAG regularizer has the NOTEARS walk-counting form and positivity obstructs exact acyclicity; and, under nonresonance (satisfied by standard rotary frequencies), a commutant criterion identifies which operators preserve relative-position dependence. An inference-collapse theorem: exact input invariance of the head score operators $G_{\mathrm{LM}}$ collapses inference to generalized SDPA with a constant operator per head. Measured invariance: relative fluctuations of $10^{-6}$ and below; perturbation bounds quantify but do not certify cached inference; the assembled proxy misses the decoding margin. A conditional three-stage mechanism (rotary twirl, concentration, row-map contraction) is measured on a released checkpoint. Blockwise training and scoring under the global Gram are stated with explicit target exposure; on tested samples, block and sequential scoring select identical answers and agree on the published TruthfulQA probability-mass metric within $5\times10^{-5}$ per item. Self-organized criticality enters as a phenomenological framework with an intrinsic order parameter; open claims become falsifiable conjectures. Selected proof cores are machine-checked in Lean 4.

arXiv:2608.10195v2cs.CV

More Accurate, Less Human: Gestalt Grouping in Vision Models

Human vision organizes what it sees into wholes: same-colored points group into series, similar marks cohere into categories, and shapes complete into recognizable objects. These are the Gestalt operations that visualization design builds on. Whether vision models organize visual content this way has not been systematically tested. We introduce a behavioral battery that scores models against human data from prior perception studies on four grouping tasks: mark-color odd-one-out, color-series counting, silhouette recognition, and object odd-one-out. We apply it to 45 models across five training families: supervised, self-supervised, and contrastive vision-language encoders, open-weight VLMs, and closed foundation models. The battery reveals that agreement with human responses captures aspects of perceptual organization that conventional performance metrics fail to distinguish, with several closed models exhibiting substantially lower alignment than their benchmark accuracy would suggest. Scoring against published perception data therefore gives visualization research a reusable yardstick, requiring no new user study, for auditing whether the models now entering visualization pipelines organize what they see the way their human audience does.

arXiv:2608.09579v2cs.CV

You Only Flow Once: Calibrated and Real-Time Radar Pose Estimation with Multi-Hypothesis Normalizing Flows

Sparse and noisy millimeter-wave radar point cloud observations often correspond to multiple plausible human poses, making deterministic pose estimation fundamentally ill-posed. Yet existing radar methods remain deterministic, collapsing this ambiguity into a single estimate. Diffusion-based alternatives can model multi-hypothesis distributions but require costly sequential denoising for each sample and lack calibrated uncertainty. We propose Multi-Hypothesis Normalizing Flow Pose Generator (MH-NFPG), which models pose distributions from radar point clouds using a conditional normalizing flow. Specifically, we combine a spatiotemporal transformer backbone with a normalizing flow that transforms a Laplace base distribution into an expressive posterior, generated in parallel through a single forward pass. Leveraging this efficiency, we outperform diffusion-based alternatives in calibration across three radar benchmarks (MM-Fi, mmRadPose, mRI) at comparable pose accuracy, while achieving up to 22x faster inference and reducing calibration error by up to 84%. We find that calibration degrades substantially for diffusion models, whereas our flow-based approach maintains reliable coverage, also across environments. These results demonstrate normalizing flows as a practical alternative to diffusion models for real-time, uncertainty-aware radar pose estimation. Our code is available at https://github.com/jonasmueler/YouOnlyFlowOnce_WACV.

arXiv:2608.07786v2cs.AI

Who Built This Model? Tracing LLM Lineage via Spectral Fingerprints in Weight Space

Open-weight large language models (LLMs) are increasingly developed through complex, multi-stage pipelines, leading to intricate lineage relationships that reflect model origin, ownership, and evolution. Understanding these relationships is important for model provenance, governance, and supply-chain integrity. In this work, we investigate the notion of LLM "biometrics" (analogous to human biometrics) to ask whether LLMs exhibit intrinsic fingerprints in weight space alone, without access to input data, that reveal their origin and lineage. We formulate this as a lineage discrimination problem, distinguishing among independent-origin, same-series, and shared-base models. To characterize these relationships, we propose a unified geometric fingerprinting framework that analyzes weight matrices from two complementary perspectives: (i) spectral energy, captured by singular value distributions to encode global magnitude patterns, and (ii) subspace alignment, quantified via subspace deviations to capture directional geometry. Our analysis uncovers a clear hierarchy of structural similarity in weight space: spectral energy reliably distinguishes independently trained models and different model families, while subspace alignment enables fine-grained discrimination among closely related models, including variations in dataset scale and post-training procedures. Extensive experiments on over 110 diverse open-weight LLM pairs demonstrate that weight-space geometry provides a robust and interpretable signal for model lineage, enabling coarse-grained regime separation and fine-grained discrimination within shared-base models.

arXiv:2608.07226v2cs.AR

Dual-Node NVIDIA DGX Spark over Tailscale: A Remote-Access Testbed for Distributed LLM Training and Cyber-Threat-Intelligence Fine-Tuning

Practical evidence for multi-node training on desktop-class accelerators remains limited. This report presents a proof-of-concept deployment of distributed NanoChat pretraining across two NVIDIA DGX Spark systems, each with a GB10 Grace Blackwell system-on-chip and 128 GB of unified memory, administered remotely over a Tailscale mesh VPN and connected for training by a dedicated 200 Gb/s QSFP56 direct fiber link. PyTorch torchrun, DDP, and NCCL were configured with one process per node, a depth-20 NanoChat model, a device batch of 32 sequences per node, a 2,048-token context, and eight gradient-accumulation steps, giving 1,048,576 tokens per optimizer step for the 896.5-million-parameter model. The run took 4,980 steps over about four days, averaging 63.4 s of measured training time per step (about 16,540 tokens/s) and processing about 5.22 billion tokens. We document link configuration, container setup, interface binding, a step-zero evaluation bug that triggered NCCL timeouts, checkpointing, and troubleshooting lessons. We also built a cybersecurity fine-tuning dataset from 77 CISA advisories (338 training, 37 validation conversations) and compared a baseline SFT checkpoint against a CTI-augmented checkpoint on 17 held-out prompts with an Ollama-hosted LLM judge: CTI-specific categories improved while general-knowledge categories regressed, for a small overall change from 2.06 to 2.29 on a 0-10 scale. Version 2 adds a stage-2 fine-tune with a MITRE ATT&CK technique dictionary under a deterministic evaluation: in-dictionary ID-to-name lookup reaches 98.5%, yet none of 50 held-out IDs is named correctly, knowledge of held-out threat groups, software, campaigns, and vulnerabilities stays near zero, and general ability drops measurably. The study establishes feasibility rather than a scaling-efficiency claim.

arXiv:2608.05548v2cs.CR

Behavioral Residualization for Unsupervised Intrusion Detection in Automotive CAN Networks

Modern vehicles exchange safety-critical messages over the Controller Area Network (CAN), a bus designed for low cost and real-time operation that offers neither authentication nor encryption. Once an attacker reaches the network, arbitrary frames can be injected. Existing detectors frequently rely on surface cues (previously unseen identifiers, constant payload bytes, or irregular DLC values). These cues yield near-perfect scores on public corpora whose attacks are separable by exactly those rules, yet they collapse when the adversary simply reuses legitimate identifiers. We introduce per-ID behavioral residualization (PIRD). Fourteen temporal, protocol and payload statistics are extracted from sliding windows belonging to each arbitration ID and then expressed as standardized residuals relative to that ID's own benign baseline. When the resulting residual vectors are supplied to six standard unsupervised detectors, mean F1 rises in the majority of experimental cells (21 of 24 on HCRL, 30 of 36 on ROAD, five random seeds). On targeted legitimate-ID attacks the method attains recall >= 0.99 and ROC-AUC near 0.99. A deployment-oriented Phase-1 protocol (cross-condition transfer, global threshold, alternative residualization strategies, window/latency sweep) additionally reports Recall at fixed FPR, FPR at target recall, false alarms per hour and detection latency. Two clear failure modes, novel-ID flooding (HCRL DoS, F1 = 0.02) and cross-ID fuzzing (ROAD, F1 = 0.27), are quantified and left as open problems.

arXiv:2608.01320v2cs.DS

Dense Language Generation Made Simple: Deterministic, Randomized, and Multi-Order Algorithms

Language generation in the limit is a theoretical framework for studying how a generator can learn to produce new valid strings from a stream of positive examples. In this model, an adversary chooses an unknown language from a countable family and enumerates its elements in an arbitrary order, while the generator must eventually output only elements of the language that have not yet appeared in the enumeration. Reliable generation is thus formalized through two eventual guarantees: validity and novelty relative to the observed data. To further quantify the breadth of the generator's outputs, Kleinberg and Wei (FOCS 2025, STOC 2026) introduced lower density as a measure of output coverage. Given an order representing the importance or relevance of possible outputs, lower density is the asymptotic lower bound, as $n$ grows, on the fraction of the first $n$ elements of the target language that the generator outputs before they appear in the data. Kleinberg and Wei showed that $1/2$ is the optimal lower-density guarantee for deterministic algorithms. We develop a simple and unified framework for obtaining optimal lower-density guarantees. We first give a deterministic algorithm that recovers the optimal guarantee of $1/2$ with a significantly simpler analysis than prior work. We then demonstrate the flexibility of our framework through two extensions. First, against an oblivious adversary, randomization raises the optimal guarantee to $1-1/e$. Second, for any finite collection of orders, the optimal deterministic and randomized guarantees can be achieved simultaneously with respect to every order, so accommodating multiple notions of importance or relevance entails no loss in the optimal guarantee.

arXiv:2608.00232v2cs.CV

Real-Time Visual Obstruction Detection in Surgical Augmented Reality

Surgical augmented reality (AR) can provide contextual guidance by overlaying virtual annotations, tool cues, and procedural information onto the surgical workspace. However, the virtual content may obstruct task-relevant real-world information, such as surgical instruments, and interfere with users' perception during time-sensitive surgical tasks. In this paper, we investigate visual obstruction detection for surgical AR and present a latency-aware pipeline that combines vision-language model (VLM)-based surgical-object recognition with segmentation-based obstruction reasoning. To reduce inference overhead, the system adopts a cascaded small-to-large VLM architecture with segmentation-guided early exiting and attention-based visual token pruning. The small VLM handles easy frames when its key-object prediction is supported by segmentation consistency, while difficult frames are forwarded to a large VLM with pruned visual tokens. We construct a pseudo-AR surgical obstruction detection benchmark by overlaying virtual content onto surgical-tool images and labeling whether the virtual content obstructs task-relevant instruments. Evaluation results show that the proposed system achieves 87.43% obstruction detection accuracy with an average end-to-end latency of 479 ms, reducing latency by 62.90% compared with a cloud large-model baseline. These results demonstrate the feasibility of latency-aware obstruction detection for surgical AR and motivate future work on dynamic surgical videos, multi-object scenes, and clinically grounded AR guidance content.

arXiv:2607.28907v2cs.IT

Domain-Adaptive Deep Joint Source-Channel Coding for Image Classification

Deep joint source--channel coding (Deep JSCC) can suffer from domain shift between training and deployment data. We study single-source domain adaptation for image classification over wireless channels. A Gaussian example shows that improving channel quality does not necessarily improve target-domain accuracy under a fixed source-trained classifier. We further proposed an information objective to reduce domain information while preserving class information. Motivated by the analysis, we develop a domain-adaptive Deep JSCC method that combines class-wise adversarial alignment with pseudo-label supervised contrastive learning. The former uses all target samples to reduce class-conditional domain differences, while the latter selects high-confidence target samples to obtain class-discriminative received representations. We relate these losses to the information objective, and analyze the effects of pseudo-label errors and confidence-based selection. Experiments on SVHN--MNIST and PACS over AWGN and Rayleigh channels demonstrate improved target-domain classification without additional inference-time networks. The proposed method achieves 98.15\% accuracy on SVHN $\rightarrow$ MNIST at 10~dB CSNR.

arXiv:2607.27384v3cs.CL

Same Illness, Different Voice: Clinical Language Models Lose Accuracy When Patients Speak for Themselves

Clinical language models are evaluated on vignettes written in a clinician's voice, but patients describe the same illness in their own. We pair 1,000 MedQA cases with three audited retellings (clinical, socioeconomic, cultural) and test 12 open models. Every model is less accurate when the patient speaks in a marked voice, by 9.6 points on average (95% CI 8.2 to 11.0), with no detectable shrinkage at larger scale. Matched controls trace the penalty. Generation temperature does not explain it, a plain first-person paraphrase that states every fact accounts for 2.4 points, and attaching the original vignette to the marked narrative recovers 7.0. We read this mainly as a failure to use facts expressed in patients' own words, not as a reaction to social cues, although our voices bundle both. Chain-of-thought and explicit debiasing instructions together close about a fifth of the gap. Having the same model first restate the account as a clinical vignette halves it and raises marked-voice accuracy by 3.9 points, whereas a structured-schema pipeline closes the same gap but is 7.1 points less accurate than the rewrite. Rewriting also invents plausible details, which we quantify. Code, data and analysis scripts: https://github.com/prabhjotschugh/narrativeshield-sdoh-medqa

arXiv:2607.25494v4math.NA

Automated Numerical Stability Analysis of Deep Learning Operators

Finite-precision arithmetic unavoidably introduces numerical approximation errors. Numerical computations may use insufficient precision or an improper formulation, which leads to numerical instability. In this paper, we introduce a unified software tool for stochastic numerical validation of deep-learning operators. The tool follows CESTAC on supported exposed operations and uses an operator-level data-perturbation approximation for GEMM-like kernels. Our developed software not only enables numerical validation with a single computation pass but also detects the sources of numerical instability and provides numerical stability monitoring during deep learning training and inference. We verified its effectiveness on the detection of polluted operators with injected numerical instabilities across various tasks. We believe that our developed method and tools provide valuable insights into developing numerically stable computing kernels, which are particularly critical for numerically stable and efficient deep learning training and inference.

arXiv:2607.25489v2cs.CV

Agentic AI in medicine: architectures, applications, evaluation, and challenges for clinical translation

Large language models and multimodal foundation models are enabling medical artificial intelligence (AI) systems to move beyond isolated prediction and undertake multistep clinical tasks that require planning, tool use, memory, iterative correction, and coordination among specialized agents. However, the scope of agentic AI in medicine remains unsettled, and current evaluation practices are not yet aligned with the requirements of clinical use. We conducted a scoping review with systematic evidence mapping across five electronic sources, screened 1,649 exportable records, and provisionally included 557 unique studies that met predefined criteria for goal-directed task execution, tool use, interaction with external resources, feedback-based refinement, or multi-agent collaboration. The included studies describe single agents that use external tools, workflows supported by retrieval and external knowledge, multimodal agents, and multi-agent systems applied to medical question answering, image interpretation, electronic health record analysis, drug safety, and clinical trial prediction. The evidence base remains dominated by public benchmarks, simulated settings, retrospective datasets, and small-scale expert evaluation. Process reliability, evidence traceability, uncertainty, safety, workflow impact, and external validity are evaluated less consistently. Clinical translation will depend on clearer definitions, reproducible evaluation, auditable oversight, interoperable system design, and prospective validation in real-world clinical workflows.

arXiv:2607.25242v3cs.CV

Medical world models in healthcare: capabilities, evidence, and trustworthy clinical translation

Medical world models offer a framework for extending medical artificial intelligence beyond static prediction by representing evolving patient states and modelling how they change over time and in response to clinical interventions. This Review uses structured narrative synthesis and reproducible evidence mapping to define conceptual boundaries, technical foundations, application domains, and clinical evidence requirements. We screened 1,455 unique records and assembled a corpus of 300 sources, including 14 studies that met a strict empirical definition of a medical world model. The field is organized around four capabilities: patient state representation, temporal dynamics modelling, intervention-conditioned simulation, and clinician-supervised planning. Evidence spans medical imaging, longitudinal electronic health records, treatment response modelling, physiological and multimodal state modelling, ultrasound and surgical interaction, and population and health-system simulation; clinical digital twins are treated as a cross-cutting integration framework. Studies provide early evidence of technical feasibility for trajectory forecasting and comparison of candidate interventions, but most remain retrospective, task-specific, or preclinical. We distinguish functional capability from clinical evidence maturity and specify evidence needed to support progressively stronger clinical claims, from temporal forecasting to intervention comparison and planning-oriented decision support. These requirements include precise intervention representations, robust causal and mechanistic grounding, calibrated trajectory-level uncertainty, safety-constrained planning, external validation, and prospective multicentre evaluation against clinically meaningful endpoints. This framework aims to distinguish technically plausible simulation from clinically supportable prediction and decision support.