Visual generative models increasingly blur the line between synthesizing observations and simulating environments. Yet visual plausibility alone does not constitute successful world modeling: compelling outputs may lack persistent state, physical dynamics, or causally consistent responses to interactions. We define a generative world model as a predictive dynamical system that maintains a persistent, history-conditioned representation of an environment, updates that representation in response to actions or interventions, and generates temporally coherent trajectories that reflect outcomes under observed and counterfactual interventions. From this definition, we derive an intelligence-oriented framework spanning perceptual grounding, persistent state modeling, controllable transformation, dynamics-aware prediction, and closed-loop interaction. We present WorldFoundry, open-source infrastructure for reproducible inference and evaluation across video, 3D and 4D, embodied action, and interactive world systems. WorldFoundry unifies model and benchmark discovery, runtime and asset preparation, execution orchestration, artifact management, and interactive visualization, while modality-specific adapters preserve native inputs, execution semantics, output representations, and evaluation protocols. Explicit manifests, durable artifact contracts, and evidence-aware scorecards separate execution readiness from evaluative claims, making each claim traceable to its supporting artifacts. The catalog contains 270 model entries, 74 benchmarks, and 62 registered metric entries, with readiness and execution evidence recorded separately. WorldFoundry enables reproducible comparison of persistent, controllable dynamics across heterogeneous systems, advancing the field from perceptual synthesis toward measurable world intelligence.
Pretrained encoders from self-supervised representation learning are increasingly applied to sonar imagery of the seafloor, where they are evaluated with protocols that take one embedding per target. A target patch of seabed, by contrast, is re-imaged from an alternate look on overlapping survey passes, so each embedding is one measurement of that target. Standard protocols cannot say how consistent those measurements are, or how much of their spread belongs to the patch and how much to the acquisition geometry. We treat this as a measurement-reliability problem. Using a synthetic-aperture sonar (SAS) survey, we measure the repeat-look reliability of embeddings from a frozen natural-image encoder (DINOv3 ViT-S/16). Standard single-look metrics, effective rank, clustering and a linear probe for coral, are all stable to the look, and the probe performs well (AUC 0.93). Yet the repeat-look reliability of the same embeddings is poor, with a distance-based intraclass correlation (dbICC) of 0.20 and nearest neighbours and clusters that barely agree across looks. This contradiction resolves as we find the probe learns a stable direction, whose ICC of 0.60 is 3.0 times that of the full embedding. Variance components, calculated from generalizability theory, attribute under a tenth of the within-cell spread to look acquisition geometry. The rest is cell-specific and cannot be separated without repeat passes at matched geometry. A sonar-pretrained encoder (BenthicDINO) and a classical SAS texture descriptor show the same pattern. For this dataset and the tested representations, uses that depend on patch identity fail, such as cross-pass comparison and retrieval, while uses at the level of habitat type (e.g. coral classification) do not. Because the standard metrics did not detect this, we argue that repeat-look reliability should accompany any evaluation of representations of the seafloor.
Search allows a robotic policy to evaluate candidate futures before committing to an action. Motivated by test-time planning for vision-language action (VLA) models, we study selective search over action chunks in PushT using a frozen Action Chunking with Transformers (ACT) policy. Beam search generates perturbed alternatives around the policy prediction and scores them through rollouts from cloned simulator states. Across 50 evaluation episodes, beam search raises task success from 0% to 18% and increases average cumulative reward from 39.44 to 103.2. We introduce One Step Ahead (OSA), a gate that simulates the nominal policy chunk and invokes full search only when its predicted coverage drop exceeds a threshold. In the reported ablations, OSA preserves or improves success while reducing runtime by approximately 3x relative to full beam search. Ablations over tree depth, branching factor, beam width, and replanning frequency show that larger search budgets do not consistently improve performance. The results support selective test-time search for action-chunking policies; evaluation on language-conditioned VLAs and physical robots remains future work.
Multimodal Large Language Models (MLLMs) have shown promise as interactive agents, yet precise, programmatic 3D graphics editing remains difficult to train because existing systems largely rely on closed-source model APIs and lack deterministic, state-verifiable execution environments. We present BlenderFORGE, a trainable data-environment-optimization framework for open-weight MLLM agents that edit explicit Blender scenes through Blender Python (bpy) code generation. BlenderFORGE targets four fundamental state-verifiable editing task families: object placement, shape-key editing, lighting adjustment, and material modification. It integrates three components: (1) a scalable perturbation-based data generation pipeline built from curated assets and large-scale 3D scene repositories; (2) an isolated Blender Sandbox for deterministic code execution, structured state export, and closed-loop reward computation; and (3) a two-stage training pipeline that uses offline teacher trajectories for cold-start Supervised Fine-Tuning (SFT) and sandbox-computed rewards for multi-turn tool-agentic Reinforcement Learning (AgentLoop-RL). Experiments across the four task families show that the resulting open-weight Qwen3-VL-8B and Qwen3.5-9B BlenderFORGE agents substantially improve over their base models and achieve competitive performance with several zero-shot proprietary MLLM baselines on selected metrics. Additional cross-benchmark evaluations and camera-reframing experiments provide evidence of transfer beyond the original training setting.
This paper introduces Cascading Circle Loss (CasCL), a contrastive learning framework for legal retrieval tasks with graded relevance. CasCL combines an InfoNCE objective for positive-negative separation with a cascade of ranking boundaries that model ordered relevance levels between queries and legal clauses. The framework also includes a negative-aware self-paced optimization mechanism and a relevance gain weighting strategy to improve ranking quality. Experiments on legal passage retrieval benchmarks show that CasCL consistently outperforms existing methods, including InfoNCE, Circle Loss, MNRL and Margin MSE, across both binary retrieval and graded ranking metrics.
Medical foundation models (Med-FMs) are increasingly adapted to clinical tasks through prompt learning, which tunes a small set of prompt tokens while the backbone remains frozen. Recent work has shown that this efficiency introduces a vulnerability: jointly optimizing learnable prompts and an imperceptible noise trigger during few-shot adaptation implants a backdoor that leaves clean performance intact but steers triggered inputs to an attacker-chosen class. Since prompts are compact and readily shared, this raises two questions: whether a backdoored prompt remains effective on a different Med-FM, and whether it can be repaired without retraining the backbone. We address both in BAPLe++. First, we present the first study of cross-model transferability of prompt-based backdoors, quantified by two proposed metrics, Clean Accuracy Difference (CAD) and Backdoor Accuracy Difference (BAD). Second, we propose DePLe (Defense via Prompt Learning), which uses the infected prompt as a fixed anchor and learns a new prompt on a small clean set that preserves the clean task while diverging from the compromised embedding region. Experiments on four Med-FMs and six chest X-ray and histopathology datasets show that backdoors transfer across models, but asymmetrically and in a dataset-dependent manner. Compared with four training-free input transformations, DePLe reduces backdoor accuracy more consistently while preserving, and in several settings improving, clean accuracy. Our code can be accessed on https://github.com/Ufaqkhan/Baple-Plus-Plus
Deep neural networks achieve universal function approximation by stacking interconnected nonlinearities, but as a consequence they are sensitive to noise and resist mechanistic interpretation. In contrast, the local update rules of cellular automata, such as Conway's Game of Life, give rise to not merely interpretable but programmable dynamics. We explore a bridge between these worlds: a lateral inhibition in neural networks based on continuous Game-of-Life rules that derives inhibitory and excitatory responses from the layer input and blends them into the activation map in a state-conditioned way. The mechanism is differentiable and architecture-agnostic, and its regulation is imposed rather than learned: the rule's parameters are fixed and are not optimised, while the rule's values and derivatives shape training throughout the network. Wired into fully connected, convolutional, wide-residual and graph convolutional networks, it gives rise to several properties. Activation maps become coherent across space, depth and seeds, with coarse structure persisting through depth while fine detail is renewed at every layer. The co-activation graph of the hidden units departs from an Erd\H{o}s--R\'enyi random graph: it is laid out on the image and its giant component grows in steps, as in a spatial graph. Networks started from nearly identical initialisations, under label noise, disagree on thirty times fewer test images. Saliency maps rest on the object rather than the background, and resilience to pixel noise not seen in training is substantially improved.
Skeleton-based rehabilitation assessment faces fundamental limitations of fixed topologies which cannot model compensatory patterns, uniform joint treatments overlook clinically important deviations, and lack of geometric understand- ing. We propose Concave-Convex Decomposition Sparse Topological Graph Convolutional Network (CC2-STG), a multi-stream model that combines geo- metric decomposition, sparse feature learning, and adaptive topology. Our framework decomposes joint relationships into complementary convex-concave distance functions to model both local and global movement characteristics, with sparse regularization with proximal optimization selecting clinically relevant geo- metric features, while multi-scale topological analysis captures non-anatomical joint dependencies in pathological movements. Experiments on the KiMoRe and UIPRMD datasets validate the effectiveness of our approach, achieving state-of-the-art performance on clinical pathological movement assessment.
Large Language Models (LLMs) are increasingly deployed as shared multi-tenant services, relying on advanced optimizations like KV cache and prefix sharing to improve throughput and reduce latency. However, these optimizations introduce a critical vulnerability: attackers can exploit time-to-first-token (TTFT) variations between KV cache hits and misses to infer sensitive prompt tokens via prefix-matching side-channel attacks. Existing defenses indiscriminately restrict cross-tenant cache sharing, which incurs prohibitive performance overhead and negates the fundamental efficiency benefits of the KV cache. We propose Mirage, a topology-aware emission obfuscation framework that provably bounds timing leakage without sacrificing the performance gains of cache sharing. Instead of stalling the GPU or forcing redundant recomputations, Mirage delays token emission at the gateway so the attacker observes a natural cache miss distribution, while the GPU seamlessly continues serving other requests at full speed. Mirage uses a Privacy Odometer that persistently tracks the topological novelty of prefix fan-out expansions to distinguish aggressive probing from benign traffic. Once a prefix exceeds a safety threshold, Mirage engages a two-phase policy that uses a dynamic estimator to pad the TTFT to a calibrated cache-miss baseline. On a 500-sample public corpus across three models, Mirage suppresses the timing side-channel to near the random-guessing floor, while maintaining a significantly lower TTFT than defense baselines. We further show that Mirage withstands a strong adaptive adversary and the same principle carries over to semantic caching layers.
Diffusion and flow models enable powerful generative modeling, but controlled editing remains challenging due to the need to balance source fidelity and edit compliance along the generative trajectory. We propose a principled, inversion-free framework for flow-based image editing from a trajectory centric perspective. Rather than relying on inversion or heuristic guidance, we introduce a one-step look-ahead mechanism that evaluates the effect of candidate updates on both edit and fidelity rewards estimated via denoising, and adjusts the current update based on this evaluation. At each step, our method applies reward-driven corrections guided by the look-ahead evaluation under an explicit fidelity constraint, resulting in stable and coherent trajectory evolution. This look-ahead design mitigates gradient interference between edit and fidelity objectives by anticipating their interaction in future states, leading to improved trade-offs. Experiments demonstrate that the proposed approach achieves stronger semantic edits while better preserving source content compared to existing strong baselines.
Aligning Large Language Models (LLMs) with human preferences via Reinforcement Learning from Human Feedback (RLHF) has driven substantial improvements across a wide range of tasks, but the cost of acquiring preference labels remains the dominant bottleneck in the alignment pipeline. While many recent alignment methods improve sample efficiency through active exploration over candidate responses, they typically query the oracle uniformly across contexts, even for contexts where the learned reward estimates are already reliable enough to provide synthetic labels. We propose Adaptive Ensemble-disagreement Routing for Oracle feedback (AERO), a fully online RLHF algorithm built on a Dyna-style model-based RL architecture, in which an ensemble-based Epistemic Reward Model (ERM) serves both as a reward posterior for exploration and as a learned preference-feedback model for synthetic labeling. AERO uses ensemble vote entropy, a Query-by-Committee disagreement measure, combined with an adaptive thresholding mechanism to decide which contexts require oracle feedback and which can instead receive ERM-based synthetic preference labels, concentrating expensive labels where they are most informative. Empirical results across different model families and scales show that AERO achieves highly competitive alignment performance while reducing the required oracle-query budget compared to online DPO and strong active-exploration baselines.
Agent harnesses -- runtime scaffolds that wrap an LLM's reasoning loop with tool dispatch, verification, and persistent memory -- have become the standard substrate for deploying LLMs as autonomous agents, as demonstrated by Claude Code and Codex in software engineering. In trajectory planning, however, this paradigm has not taken hold: existing LLM-based trajectory planning methods either keep the LLM in the decision loop, exceeding the millisecond-level control budget, or offload it to offline rules or training-time supervision -- leaving the runtime planner non-evolving. Intermediate hybrids still lack the verification-and-consolidation loop that makes harnesses reliable. We ask: what runtime harness would let an LLM planner meet real-time constraints while self-evolving into a deployable one? We propose REINS, a harness built on three LLM-planner couplings: LLM outputs are (i) grounded to a learned variable-level dynamics model, (ii) verified via calibrated predicates and forward rollout, and (iii) consolidated into an indexed rule memory that serves future scenes in O(log n) without re-invoking the LLM. The harness closes a self-evolution loop: familiar scenes resolve by memory lookup; novel scenes invoke the LLM, whose verified outputs enrich the memory. Without modifying LLM weights, REINS attains 98%+ compliance, nearly 50% collision reduction, and 1.9-4.6 ms latency on simulated and real-world benchmarks at the tactical decision layer, with LLM invocation rate monotonically declining as memory matures. Code: https://github.com/ZhihongCui/REINS
Lane-level map updating requires temporally consistent, physically valid, and auditable road-scene semantics despite occlusions, ambiguous lane markings, temporary obstacles, and viewpoint changes. We present RoadState, a production-deployed vision-language system that transforms sequential road-scene observations into structured lane-level semantics for Baidu Maps' map-updating pipeline. Its core model, RoadState-VLM, formulates road-scene understanding as autoregressive state estimation, predicting the current structured road state from the previous observation and symbolic state together with the current observation. We introduce Geometric Chain-of-Thought (Geo-CoT), which uses a directed acyclic graph of semantic dependencies as a domain prior to improve semantic consistency and reasoning fidelity across interdependent road-state dimensions. Extensive experiments on a real-world dataset show that RoadState-VLM achieves 94.33% mean accuracy across six road-state dimensions, outperforming the strongest fine-tuned VLM baseline. Ablation studies further demonstrate the complementary effectiveness of Geo-CoT and temporal transition context. A cascaded routing pipeline directs routine cases to lightweight models while reserving complex cases for RoadState-VLM. Since its deployment in March 2025, RoadState has supported lane-level map updating across more than 360 cities and reduced overall production costs by 30%. Together with human-in-the-loop auditing, the system maintains robust performance under stress-test conditions. These results demonstrate how domain-prior-enhanced reasoning, structured state interfaces, selective routing, and human quality control can bridge the gap between VLM-based road-scene understanding and large-scale lane-level map updating.
Handedness is a soft-biometric cue for forensic document examination and writer profiling, yet reported accuracies from offline handwriting rarely keep writers, texts and image acquisition apart, so models may recognise the writer or capture conditions rather than the writing hand. Such controlled evidence is missing for Bangla. Using our BanglaHand dataset of 42 writers (21 left- and 21 right-handed), who each wrote six shared prompts and two own texts under controlled collection and photography, we evaluate handedness identification with all images of each test writer held out, whole-writer permutation tests and photometric controls. An ink-gradient descriptor identified the handedness of unseen writers with 77.4% balanced accuracy (95% writer-bootstrap interval 67.7-86.5%; p=0.001) and frozen ImageNet embeddings with 69.5-72.0%. Part of this signal is stroke darkness: in a post hoc analysis, right-handed writers' strokes were darker at their centres for equal stroke width, consistent with higher pen pressure and not explained by lighting drift. With ink density and brightness regressed out, the descriptor still reached 64.5% (p=0.027), whereas writing-only shape features were weaker (60.8%; p=0.090). Handedness is thus identifiable from offline Bangla handwriting under controlled acquisition, and photometric controls are essential to interpret such cues.
Vision-Language Models (VLMs) can generate procedures that look convincing on paper, but do they work in practice? In this paper, we propose crochet as an ideal testbed for physically grounded procedural reasoning in VLMs. Crochet cannot be reliably simulated or automated: the only way to know what a pattern produces is to make it by hand, yet verification is tangible and cheap, requiring nothing but a hook and a ball of yarn. We introduce FoundationChain, a deceptively simple test of procedural reasoning: given an image of a basic crochet object, can VLMs infer how it was made and generate a pattern that produces the same result? We crochet 40 patterns generated by four VLMs by hand. Even at this scale, physical execution not only reveals failures that automated proxies cannot detect, but also enables a more precise diagnosis of their causes and consequences: while models write fluent, convincing patterns, only four are free of errors, and our analysis exposes deficiencies in visual, geometric and procedural reasoning.
Stroke is a disorder of the circulation, yet the haemodynamic state that immediately precedes it has rarely been observed, because continuous physiological recordings before a stroke are seldom available. Photoplethysmography (PPG), recorded by every pulse oximeter and by hundreds of millions of wearable devices, captures the peripheral pulse beat by beat, but its morphology is largely discarded. Here we ask whether pulse waveform morphology encodes haemodynamic signatures of impending stroke. Using in-hospital stroke as a natural experiment that provides continuous PPG before physician-adjudicated stroke recognition, we studied intensive-care (MIMIC-III, $n=176$) and emergency-department (MC-MED, $n=154$) cohorts and described each beat with interpretable descriptors of pulse timing, amplitude, contour and variability. At a nominal 360-min horizon, morphology separated pre-recognition from stable windows with $F_1$ scores of 0.941 and 0.989 (sensitivity 0.998 and 0.980), with models transferred unchanged to the second cohort, and outperformed heart rate alone on identical windows. The signature was dominated by systolic crest time and its change from each patient's own baseline, feature attribution was concordant across cohorts (Spearman $\rho=0.865$), and morphology recorded in the first two hours of care stratified the subsequent clinical course ($F_1$ 0.714 and 0.691). In high-risk patients without stroke, 29% and 14% of windows were positive, but five-window persistence rules limited alerts to 17% and 2% of patients. The results demonstrate that the routinely recorded pulse thus carries haemodynamic imprint of the period preceding stroke, positioning pulse morphology as a candidate continuous cerebrovascular biomarker.
Large language models (LLMs) often do not produce calibrated uncertainty estimates for their answers. We propose CAELUM, a framework that builds semantic prediction sets while separating aleatoric from epistemic uncertainty. CAELUM samples answers using an ensemble of clarification prompts, clusters them by semantic equivalence, and estimates the aleatoric and epistemic components via kernel-based von Neumann entropy before calibrating nonconformity scores with split conformal prediction. We use a weight $w_\lambda(x)=1+U_{\mathrm{ale}}(x)+\lambda\,U_{\mathrm{epi}}(x)$ to control for the contribution of the epistemic uncertainty of each instance and fit a coefficient $\lambda^\star$ to minimize prediction set size. When the objective is nonflat and selection is stable across splits, $\lambda^\star$ could be a diagnostic, where large values may indicate epistemic dominance (retrieve, abstain) and $\lambda^\star\to 0$ aleatoric dominance (clarify). For individual queries, a log ratio $R(x)$ compares the weighted epistemic component with the aleatoric component. Theoretically, we demonstrate that the coverage is capped by matchability, the fraction of items whose reference answer appears in any sampled cluster. Empirically, this matchability can be predicted from the original question features using a lightweight classifier. We use those predictions to define a stratum and derive a conformal coverage bound within it. Across 30 LLM and dataset combinations (4-32B), mean coverage is 0.74 in the learned stratum versus 0.65 overall, averaged over the 78 evaluations using the learned classifier. Project page: https://unco3892.github.io/caelum/
A brain is in one state at a time, yet most of what decides its next state is absent from it: the patterns its world never produces, the statistics of a lifetime, a goal that lies ahead. In mechanics a wire acts on a bead only through a force that appears when the bead presses on it, and Lagrange computed such forces with one multiplier for each constraint. We propose that a brain follows the same mechanics. Everything a circuit knows, is told or wants is a constraint on its single state and acts through its own multiplier, which accumulates the violation and pushes the state back. Neurons and synapses carry the multipliers, each with its own time constant: inhibitory interneurons carry what never happens, adaptation the cost of staying in one state, synapses the statistics of experience, and feedback to apical dendrites the future. From this hypothesis follow Dale's law, the reciprocal wiring and balance of excitation and inhibition, the paradoxical responses of cortex and the frequencies of brain rhythms. Belief becomes the time one trajectory spends in each state, learning becomes sleep proposing laws and waking refuting the false ones, and planning becomes a wave from the goal meeting a wave from the present. Connectionism and symbolism are the limits of frozen and of instantaneous multipliers, and between them the time a circuit runs becomes the depth of its computation. A machine built on these principles, trained without gradients on 1000 solved puzzles, solves all 422,786 Sudoku-Extreme test puzzles with one trajectory each. It rediscovers the rules of Sudoku by sleeping and waking, plans shortest paths in mazes of 64 times the training area, and tracks permutations exactly on strings 25 times longer than in training, where Transformers fail. A mind, on this view, is life that learns its own constraints.
Multimodal large language models have made progress in general visual tasks. However, limitations remain in fine-grained visual tasks requiring evidence from minute image regions. Existing methods primarily train models using evidence-related regions while overlooking the influence of extensive background regions, resulting in a limited ability to resist interference from irrelevant regions. We propose selective positive-negative on-policy self-distillation (SPN-OPSD), which considers both evidence and background regions. Supervision from both views improves the model's ability to resist interference from irrelevant regions. Furthermore, selective token-level distillation provides more informative supervision, improving the model's ability to attend to evidence regions in Fine-Grained Visual Tasks. On Qwen3.5-2B, SPN-OPSD improves accuracy on V* Bench, ZoomBench, and MME-RealWorld-Lite from 76.44%, 45.92%, and 57.43% to 84.82%, 54.56%, and 59.56%, respectively. Similar gains are also observed on Qwen3.5-0.8B.
This work presents a one-class anomaly detection framework for vibration-based condition monitoring in machining processes using a convolutional neural network autoencoder. Three-axis vibration signals acquired during industrial broaching are transformed into time--frequency spectrograms and combined into a compact three-channel representation, enabling convolutional feature learning on multivariate time-series data. The autoencoder is trained exclusively on normal process data and detects anomalous strokes based on reconstruction errors. Labeled validation data are used for threshold calibration and an independent test set is reserved for final performance evaluation. To improve sensitivity to short-lived process instabilities such as chatter, a localized residual-based metric termed Segmented Vertical Error (SVE) is introduced. Experimental results on approximately 10{,}000 broaching strokes show a clear separation between normal and anomalous process conditions under strong class imbalance. The processing pipeline is additionally evaluated on an external public turning dataset as a proof of concept, indicating preliminary applicability beyond the original broaching setup. The results suggest that spectrogram-based autoencoder models can provide a practical and scalable basis for one-class anomaly detection in industrial vibration monitoring.