Each section below summarizes one paper's reported method, experiments, and limits from its full review. The studies use different tasks and metrics; their results are not directly comparable.
arXiv:2611.01866
The (Butter)fly Dream: Linear Separability of World-Model Rollouts in a Connectome-Constrained Visual Model
This study asks whether temporal defects in generated visual worlds are linearly decodable from the responses of a fixed, connectome-constrained fly visual-system model. The observer was trained for optic-flow estimation, while a logistic-regression two-sample test examined statistics of its cell-type responses to MPI-Sintel sequences and continuations from a persistence random walk, an MDN-RNN, and an RSSM. Across nine generator and temperature conditions, the trained observer separated dreams from real continuations with AUCs of 0.85–0.99, and learned-model dreams were less separable before observer training. Graded probes suggest that trained responses expose subtle temporal disorder: one swapped adjacent frame pair yielded AUC 0.73.
How it works
The observer is flyvis version 1.2.0, a hexagonal-lattice optic-lobe model whose connectivity follows the connectome. It contains 45,669 cells of 65 types and has 734 free parameters trained by the model’s authors for optic-flow estimation. The study uses all 50 released trained models without modification. Its untrained control restores no checkpoint and retains the connectome’s signs and synapse counts; five seeds differ in their cell-type bias draws.
What the paper evaluates
The empirical evaluation uses MPI-Sintel: 35 sequences, yielding 198 windows, including 17 sequences seen during observer training and 18 observer-unseen sequences. The generators are trained and evaluated through sequence-level folds; the classifier is evaluated by leaving out a generator fold. The primary metric is window-level AUC, with 0.5 meaning indistinguishable and higher values meaning greater classifier separability. The paper reports trained ensemble means and standard deviations, untrained-seed means, and sequence-cluster bootstrap intervals for selected comparisons. The main comparison covers all three generators at temperatures 0.25, 0.5, and 1.0.
Limit to keep in view
The evaluation covers MPI-Sintel luminance on a 721-column retina, with 14 generated frames; it does not establish performance on other visual domains or long-horizon rollouts.
arXiv:2611.08392
Learning a patient world model for health system simulation in intensive care
This paper presents a time-conditioned patient world model (pWM) learned from longitudinal MIMIC-IV records and coupled to a discrete-event ICU simulator. A frozen neural model encodes 64 clinical state variables and 25 event types, supports arbitrary-time queries and autoregressive rollouts, and supplies admission proposals, organ-support events, and discharge proposals to scenario-specific resource mappings.
How it works
The study uses MIMIC-IV v3.1, containing 364,627 patients and 94,444 ICU stays. Patients are partitioned deterministically, with 80% used for training and the remaining patients divided into validation and held-out test groups. The EHR representation stores irregular observations on a sparse two-dimensional grid. Continuous measurements are discretized into ten quantile bins; event labels are grouped into retained ATC and ICD categories plus an “other” category.
What the paper evaluates
The evaluation has four layers. First, six fixed-history clinical queries are tested on 1,024 held-out records each: hospital mortality, ICU mortality at admission, ICU mortality at 24 hours, 30-day ICU admission, 30-day hospital readmission, and 30-day ICU readmission. The reported AUROCs are 0.937, 0.881, 0.814, and 0.682 for in-hospital mortality, 30-day ICU admission, 30-day ICU readmission, and 30-day hospital readmission, respectively. Across 24 task-horizon queries, the median recalibrated slope is 0.988, with an interquartile range of 0.885–1.051.
Limit to keep in view
MIMIC-IV records clinical trajectories from Beth Israel Deaconess Medical Center rather than empirical queue logs or staff schedules, so arrival streams, service durations, and capacities are scenario specifications rather than locally measured workflow data.
arXiv:2611.08449
VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding
VFlash, or Flash Primitive Decoding (FPD), speeds visual planning by predicting a complete structured output—such as a bounding box or variable-length 2D/3D trajectory—in one parallel drafting pass. It trains a lightweight drafter on annotated spatial primitives and accepts outputs that satisfy the primitive’s structural constraints, without requiring coordinate-token agreement with the autoregressive (AR) target. On the paper’s open-loop robot trajectory tasks, VFlash reduces prediction errors relative to its shared AR target; the result depends on the task and training variant. In the downstream MolmoSpaces setup, the VFlash-conditioned world action model reaches 70.2% average success versus 62.6% for the stated baseline.
How it works
VFlash’s FPD uses a frozen target VLM and a shallow non-causal drafter. The schema describes primitive type, coordinate dimensions, and extent. Fixed-size boxes can be determined by type; a variable-length trajectory additionally uses a tool call to declare extent. This means the decoder knows the primitive’s boundaries and coordinate layout before predicting its values. The model can represent joint 2D image-plane and 3D metric coordinates in one primitive, and the paper also describes dense motion blocks.
What the paper evaluates
The paper evaluates open-loop 2D trajectories on ShareRobot-Bench and VABench-V using DFD, HD, and RMSE in normalized image coordinates, all lower-is-better. The controlled comparison uses the same 4B target for Baseline (AR) and VFlash. VFlash lowers all six errors: for example, VABench-V RMSE is 0.089 versus 0.093, and ShareRobot-Bench HD is 0.162 versus 0.175. The authors report reductions of 7% on HD and 3–6% on DFD and RMSE across the two benchmarks. Table 2 preserves the paired values. For category-conditioned boxes, the datasets are COCO and Dense200. The metrics are F1 at IoU 0.50 and 0.95, mean F1, and boxes per second (BPS), with higher values better.
Limit to keep in view
The reported single-rollout speed estimate of approximately 25× for 3D point motion forecasting is tied to the paper’s setup; the supplied main-body evidence does not establish the same speedup across other rollout conditions or deployment workloads.
arXiv:2611.04566
A research world model of power engineering learns its dynamics and reveals what-ifs
This paper builds a research world model for power engineering from 1.20 million abstracts, 11.1 million citation events, and term usage in seven neighboring fields. A distilled language-model reader converts papers into role-typed concepts, graph-temporal growth and link cores learn how the field evolves, and intervention operators simulate bursts, new terms, adoptions, releases, and other actions. In temporally held-out tests, the system forecast concept growth and new method–target links better than tuned baselines in most settings, recovered effects from held-out natural experiments, transferred to semiconductors and materials science, and answered natural-language what-if questions by compiling them into executable paired simulations.
How it works
The corpus contains 1,196,720 power-engineering abstracts harvested from 33 OpenAlex topics, reduced by deduplication to 1,142,498 papers. The reader extracts concepts with five roles—problem, method, object, setting, and metric—and their typed relations. The state also includes role-specific counts, typed edges, citation attention, external supply from 58.5 million papers in seven other OpenAlex fields, and time-respecting concept text embeddings.
What the paper evaluates
The main dataset is the cleaned power-engineering OpenAlex corpus, with citation events and external-supply terms from seven other OpenAlex fields. Replication tests use semiconductors and materials science. Forecast evaluation uses walk-forward cutoffs from December 2013 to December 2021, with Spearman correlation over the 1,000 most active concepts for growth and typed average precision for links. Baselines include default, horizon-matched, and tuned LightGBM, a tuned multilayer perceptron, 11 link-prediction baselines, Adamic–Adar, and zero-shot Qwen3.8-27B.
Limit to keep in view
The operators rank intervention effects and determine directions more reliably than they estimate effect magnitudes; the paper reports systematic underestimation of size.
Read the full reviews
- The (Butter)fly Dream: Linear Separability of World-Model Rollouts in a Connectome-Constrained Visual Model · arXiv:2611.01866
- Learning a patient world model for health system simulation in intensive care · arXiv:2611.08392
- VFlash: Fast, Accurate Visual Planning via Flash Primitive Decoding · arXiv:2611.08449
- A research world model of power engineering learns its dynamics and reveals what-ifs · arXiv:2611.04566
Source note: findings and limitations above come from the linked full paper reviews. Follow each original arXiv paper for the authors' complete methods and results.