This study introduces an intriguing phenomenon in Video LLMs: rather than merely translating frames into textual embeddings, Video LLMs establish a continuous manifold, token interface, allowing visual tokens to operate as standalone entities within the architecture. Exploiting this discovery, we propose V-LynX, a scalable framework that integrates novel modalities into Video LLMs by repurposing the internalized interface. Departing from conventional paradigms that necessitate heavy modality-specific encoders or paired supervision, V-LynX employs a lightweight auxiliary pathway in parallel with the frozen vision encoder. Our method integrates new sensory inputs with intrinsic video priors by aligning both attention responses and statistical distributions using unpaired unimodal data sets. This ensures manifold compatibility while preserving the integrity of the Video LLMs. Extensive benchmarks demonstrate that V-LynX achieves SOTA and efficiency across audio-visual QA, 3D reasoning, high-frame-rate, and multi-view video understanding. The code is available at https://github.com/park-jungin/lynx.

Recent advances in Speech Large Language Models (Speech LLMs) have significantly enhanced spoken language understanding and reasoning. However, their contextual awareness is limited, struggling to perform speech recognition that effectively reflects the speaker's intent and topical context. In this paper, we propose LaSR (Latent Speech Reasoning), a novel training paradigm featuring a context-aware reasoning trajectory that leverages the latent reasoning process. Instead of generating explicit intermediate tokens, LaSR aligns chain-of-thought (CoT) supervision around the acoustic feature region of the targeted word, and introduces latent reasoning periods for context information grounding and transcriptional transition. Furthermore, to effectively benchmark contextual recognition on specialized vocabulary, we propose Spoken Darwin-Science, a large-scale corpus focusing on academic terminologies. Preliminary experiments on Fun-Audio-Chat demonstrate that LaSR significantly improves terminology recognition without introducing additional latency and consistently outperforms standard supervised fine-tuning baselines. Our findings highlight the potential of latent reasoning in building efficient, context-aware speech assistants.

Energy consumption prediction is essential for efficient grid management, demand-side optimization, and sustainable energy planning. Although advanced machine learning methods have been employed for better prediction performance, existing works have two key limitations: (1) they usually formulate this task as a purely time-series prediction problem without explicitly modeling the spatial dependencies among different regions, and (2) they fail to provide reliable predictions with uncertainty estimates under abnormal situations such as extreme weather events. To advance existing research, we propose EnergyMamba, an uncertainty-aware spatiotemporal learning framework for accurate and reliable energy consumption prediction, which comprises two key components: (i) a novel Graph-Enhanced Selective State Space Model (GE-Mamba) that injects spatial context learned from the grid topology into the temporal dynamics, enabling coupled spatiotemporal modeling, and (ii) an Adaptive Sequential Conformalized Quantile Regression (AS-CQR) module, which includes locally adaptive normalization and an online feedback mechanism to dynamically calibrate prediction intervals under potential distribution shifts. We evaluate EnergyMamba on four large-scale real-world datasets from Florida, New York, and California. Results show EnergyMamba achieves around 5% improvement in prediction accuracy and 6% improvement in uncertainty quantification over 15 state-of-the-art baselines.

Modifying an attribute in tabular data often introduces an unnatural instance by breaking its relationships with other attributes. The modified instance must be both natural and minimally changed from the original instance. This paper addresses the challenge of generating such a modified instance. We identify key limitations in existing approaches: generative models either don't support instance-level attribute editing or, in the case of methods like CVAE, retain attribute information in the latent space, leading to unnecessary modifications. To solve this, we propose TabChange, an approach that analyzes the relationship between the attribute of interest and other attributes in the dataset. If the relationship is weak, it simply flips the attribute; if it is strong, it uses an adversarial framework that removes information about the attribute in the latent space representation. This removal enables precise modifications, making only the necessary adjustments to maintain naturalness. Our experiments across seven datasets show that TabChange generates counterfactuals in attributes that are comparable in naturalness and are more proximal to their original instances. This leads to a higher number of valid counterfactuals and a lower number of invalid counterfactuals compared to the baselines.

Video generation models are becoming a scalable form of world models, but they mainly generate plausible motion rather than proactively control or optimize the underlying dynamics. As a result, an object in the generated video may follow trajectories that are unsafe, not smooth, inefficient, or physically inconsistent. In this work, we propose \textbf{OptiWorld}, a framework that brings classical optimal control into video generation at inference time. OptiWorld first extracts a compact, task-relevant world state, then plans an optimal trajectory under physical constraints, and finally renders the video conditioned on this trajectory. We formulate planning as a geometric problem on a continuous manifold, which converts 3D geometry and task-dependent physical constraints into a unified planning geometry. By adding this optimal-control layer, OptiWorld generates videos with preferable dynamics, demonstrating strong potential in multiple tasks including goal-conditioned image-to-video generation, video dynamics editing, and counterfactual generation.

Oscillatory neural signals such as electroencephalography (EEG) and local field potentials (LFPs) show phase relationships that coordinate communication across brain regions. Modern recordings capture hundreds of channels across many frequency bins, yet standard phase analyses are restricted to only a few variables. The Torus Graph (TG) model, an exponential-family distribution over phases whose univariate and pairwise potentials generalize von Mises distributions, infers principled structure among oscillations but models only static, undirected dependencies and is limited to $\sim \! 100$ variables because its score matching inference scales as $\mathcal{O}(d^{6})$. We introduce a stochastic score matching procedure that reduces the per-iteration cost to $\mathcal{O}(d^{2})$, enabling inference on datasets with thousands of variables. This scalable foundation supports analyses of 1,860 frequency-phase features from multi-electrode LFPs and enables two extensions previously inaccessible to TGs or classical circular statistics: (i) a TG Hidden Markov Model capturing state-dependent phase-coupling changes (e.g., spindle-related states during sleep) and (ii) an autoregressive TG inferring directional interactions via transfer-entropy estimation. Applied to LFP recordings, these models reveal state-dependent phase-interaction patterns between wakefulness and NREM sleep. Together, they enable systematic, large-scale mapping of dynamic and directional phase relationships across brain and cognitive states.

Using a diffusion model for parallel drafting is a promising approach for speculative decoding. By predicting tokens at multiple future positions in a single forward pass, diffusion drafters substantially reduce drafting latency. However, this shifts the bottleneck to verification: verifying a single sequence limits acceptance length, while verifying large draft trees incurs excessive target-model latency. We identify a key mismatch in existing draft-tree methods: existing diffusion-tree methods rank nodes by the marginal probability, ignoring that verification is prefix-conditioned. As a result, they may verify unreachable descendants of rejected prefixes, increasing latency with limited acceptance gains. To address this, we propose TAPS, a target-aware prefix selection method that turns diffusion marginals into path-conditioned acceptance estimates. TAPS then selects a compact prefix-closed subtree under a fixed verification budget, improving the acceptance-cost tradeoff rather than simply expanding the draft tree. Experiments across diverse datasets and model families demonstrate that TAPS achieves up to 7.9x lossless end-to-end speedup over vanilla autoregressive decoding, outperforming state-of-the-art DFlash and DDTree by 1.36x and 1.74x respectively. Our work is available at https://anonymous.4open.science/r/TAPS-EMNLP2026-53DD

Unified multimodal models (UMMs) have emerged as a promising paradigm for general-purpose multimodal intelligence. As they are deployed in real-world applications, effectively updating internal knowledge becomes critical. While knowledge editing has matured for text-only models, it remains unclear whether edits that successfully modify textual outputs also transfer to image generation in UMMs. To study this question, we introduce UniKE, the first benchmark for cross-modality knowledge editing in UMMs, comprising 2,971 edit subjects spanning attribute and relation edits. Using VQA-based visual verification, we reveal a striking modality gap: text-side efficacy can reach approximately 92%, whereas the best overall VQA accuracy under direct image generation is only 18.5%. We further propose Reasoning-augmented Parameter Editing, which explicitly activates edited knowledge before generation and improves overall VQA accuracy for all evaluated model-editor pairs, with gains up to 18.6 percentage points. Mechanistic analysis shows that this gap is associated with partial alignment between edited textual representations and the conditioning pathways for visual generation, where edits sufficient for text outputs may remain too weak or misaligned to steer image synthesis. These findings show that textual knowledge edits do not guarantee reliable cross-modality transfer and motivate modality-aware editing methods. Our code and data are available at https://github.com/gxx27/UniKE.

Do LLM agents act on the reasoning they state? This question of process fidelity is central to using LLMs in social simulation, yet it is hard to measure where no reference for correct behavior exists. We study it in acontrolled setting, a Texas Poker simulator with a verifiable reference action for every decision by decomposing the faithfulness gap into two steps: reasoning-conclusion and conclusion-action. The two steps behave oppositely.

Conventional tissue image analysis software provides foundational capabilities for cellular analysis, including segmentation, basic morphological feature extraction, and spatial organization analysis. However, these tools often require manual intervention and are not well integrated with code-driven automation, limiting efficiency and scalability for complex spatial tissue studies. In addition, they offer limited flexibility for custom analyses, as they typically support only a fixed set of pre-implemented spatial cellular features. To address these limitations, we propose CodeCytos, a coding-based reasoning agent framework that enables dynamic, programmable interaction with spatial molecular imaging data to improve automation and customization. CodeCytos is designed to streamline the exploration of custom spatial cellular features and adapt to diverse research needs. We demonstrate its utility through case studies on four expert-curated datasets from distinct tissue types: frontal cortex, non-small-cell lung cancer, pancreas, and tonsil. We evaluate CodeCytos under a realistic minimal prompt setting, where bioscientists pose simple questions without task-specific instructions or contextual information about spatial cellular analysis, and benchmark multiple LLM backbones with strong coding capabilities. We further show that incorporating tailored, domain-agnostic few-shot in-context coding-reasoning examples (randomly sampled demonstrations outside the spatial analysis domain) can substantially improve performance without requiring costly, expert-crafted in-domain demonstrations. Overall, CodeCytos outperforms baseline approaches, highlighting the potential of code-action agents to assist with custom feature exploration in spatial molecular imaging and to accelerate biomarker discovery.

Accurate pedestrian trajectory prediction is essential for safe navigation in autonomous driving and intelligent transportation systems. Despite substantial progress made by recent methods, most existing approaches are limited in fully exploiting diverse observations and often overlook the scale dependency of future motion, treating multiscale features uniformly regardless of underlying motion dynamics. This limits their robustness across diverse pedestrian behaviors. To address these challenges, we propose a Predicted-MUltiSCale-Aware Network (MUSCLE-NET) for Pedestrian Trajectory Forecasting that integrates complementary multimodal cues with scale-adaptive prediction mechanisms. The proposed framework is built upon a Multiscale Multimodal Feature Extraction (MMFE) module, which combines multiscale representation, modality-aware recalibration, and directional cross-modal fusion to construct semantically aligned representations from bounding boxes, velocities, and pose information. Building on these features, a Multiscale Enhanced Hierarchical Prediction (MEHP) module performs prediction-aware future-motion refinement via a probabilistic coarse predictor, scale-aligned fusion, and progressive refinement, adaptively selecting scale-relevant cues to mitigate spatial drift. Extensive experiments on the JAAD and PIE benchmarks demonstrate that the proposed MUSCLE-Net achieves competitive performance and consistent gains compared with state-of-the-art trajectory prediction methods.

A passive monolithic compliant grasping mechanism that works based on the eversion of an elastically deformable bistable shell is conceptualized. It comprises grasping arms made of beam segments that work in conjunction with the everting shell. The grasper is capable of picking up a stiff object of any shape up to a maximum size and weight. The bistable shell everts upon contact with the object to enable the grasping arms envelop the object forming an enclosure. The mechanism then stays in that configuration until it is actuated again to turn the shell back to its original configuration and thereby opening the enclosure to release the object. The stiffness of the arms decides the payload of the mechanism. The size of the arms decides the largest object that can be grasped and held. The arms have distributed compliance so that they can conform to the shape of the object without applying undue force on it.

Large Language Models (LLMs) are increasingly used for zero-shot annotation and LLM-as-a-judge tasks, yet their reliability hinges on how model-internalized priors interact with user-provided instructions. We investigate three dimensions of this interaction: (1) how an LLM's familiarity with data and task definitions affects performance, (2) the extent to which additional information in prompts can correct zero-shot errors ("decision stickiness"), and (3) model susceptibility to misaligned task definitions. Through experiments on toxicity detection across diverse datasets (spanning social media, gaming, news, and forums) using both dense and mixture-of-experts models, we find that nearly two-thirds of zero-shot errors are resistant to correction, with an overall rescue rate (fraction of initial errors corrected by prompting) of only 34.8%. High-confidence errors prove especially resistant to correction. When given misaligned definitions, LLMs follow them while maintaining confidence levels unchanged from the aligned condition. Crucially, we introduce Definition-Specific Familiarity (DSF), which measures alignment between a model's internal concept and the task definition. After controlling for dataset-level confounds, DSF shows a positive association with model performance (partial r = +0.41), while three distinct memorization metrics (ROUGE-L, BERTScore, and embedding cosine similarity) all fail to show a positive association. These findings show the limitations of prompt-based correction in annotation tasks, highlighting the importance of definition alignment over text-level memorization.

Short-form text rewriting is a constrained variant of paraphrasing in which limited context and high semantic density leave little room for variation. While large language models perform well on general paraphrasing, small language models (SLMs) often struggle with semantic fidelity and hallucination robustness in short-form settings. In this work, we present an empirical study of adapting an SLM, Phi Silica, for short-form rewrite through dataset curation, prompt distillation, parameter-efficient fine-tuning, and evaluation. We curate a dataset of short presentation-style text from public slide decks and use GPT-5-chat both to generate rewrite supervision and to conduct LLM-as-a-judge evaluation. Our results show that finetuning improves semantic fidelity, reduces hallucinations, and increases preference win rate against GPT-5-chat rewrites. The findings suggest that targeted adaptation for SLMs can substantially narrow the gap to cloud models and provide practical guidance for adapting SLMs to precision-critical rewrite tasks.

Tremor is a common movement disorder associated with conditions like Parkinson's disease and Essential tremor, traditionally diagnosed through expert clinician assessment. Current automated detection methods rely on frequency-domain features informed by clinical expertise. In this work, we present an explainable, two-stage hierarchical framework for tremor detection in the time domain that learns tremor patterns directly from 3D kinematic marker time-series data across entire tremor-provoking trials. Our framework combined a deep convolutional and long short-term memory network to learn tremor representations from short, discrete, non-overlapping time segments of kinematic time series data from trials, which are then processed by a vision transformer that models their long-term temporal dynamics of time segment features for trial (session) level classification. Evaluated across nine body parts, the framework achieved F1-scores of 0.594 - 0.947 depending on body parts (average: 0.765), falling short of the frequency-domain state-of-the-art performance (0.909) while requiring minimal preprocessing. Attention weights and gradient-based class activation maps (Grad-CAM) identified time-domain features of tremor across body parts. This proof of concept demonstrated the feasibility of data-driven time-domain modeling for tremor detection across anatomically diverse body parts, while reducing reliance on expert-engineered spectral features and providing posthoc interpretability of temporal and anatomical patterns of tremor.

Speech-aware large language models often generalize poorly to out-of-domain settings. We propose SALSA (Speech-Aware LLM Adaptation via Learned Steering Activations), a lightweight adaptation method that learns layer-wise steering vectors. Unlike commonly used steering approaches that rely on contrastive activation differences, SALSA directly optimizes steering vectors using a supervised objective. Across children's speech, multilingual speech, and Mandarin-English code-switching benchmarks, SALSA substantially improves performance over zero-shot inference and speech in-context learning baselines, achieving up to 46.8% relative improvements over zero-shot. Analysis further demonstrates that steering the encoder, particularly the later layers, is more effective than steering the LLM backbone. These findings suggest that steering improves downstream ASR performance by adapting higher-level acoustic and phonetic representations to better align with the pretrained language model representation space, rather than by modifying the decoder itself.

Compliant force or torque control are approaches often investigated to achieve safe physical human-robot interaction (pHRI). However, these approaches have limitations. Force control requires a robot to be equipped with external force sensors to track the amplitude and direction of applied forces. Torque control requires torque sensing or estimation in each joint. As this is not available on every robot, energy-based approaches offer a promising alternative. Such approaches aim to achieve safe pHRI by limiting the mechanical energy of the robot. Current schemes leveraging an energy-based approach tend to have a complex implementation, and some may require further stability verification. We hence propose an adaptive proportional-derivative (PD) controller that can limit a robot's energy under any given limit to achieve safe pHRI. The proposed controller can limit both the kinetic and potential energy of a robot, and the behaviour of the controller gains can be shaped using various parameters, defining precisely the cutoff limit and sharpness. We construct a stability proof for the controller and define a condition to ensure the controller's stability. The proposed controller's behaviour and compliance are tested on the TALOS robot from PAL Robotics both in simulation and on hardware, verifying the expected compliant and energy-limiting behaviour of the controller.

Dynamic scene reconstruction via 3D Gaussian Splatting (3DGS) has emerged as a compelling approach for representing evolving environments, yet understanding trade-offs between methodologies remains crucial. This paper presents a comprehensive analysis of dynamic 3DGS methods, categorizing them into two paradigms: structure-guided methods employing auxiliary representations (deformation fields, canonical spaces, grids) to model temporal changes, and gaussian-centric methods encoding dynamics directly into primitives via continuous functions or 4D representations. We evaluate representative methods from both paradigms on the D-NeRF benchmark. Our findings reveal that structure-guided methods achieve superior reconstruction fidelity and compact model sizes, while gaussian-centric approaches demonstrate significantly higher rendering speeds enabling real-time performance, though with greater quality variability and potentially substantial storage overhead. This analysis highlights a fundamental trade-off between reconstruction quality/compactness versus rendering speed, providing insights to guide future research and application development in dynamic scene reconstruction.

Protein-language systems are often evaluated by whether they generate plausible biological text, but a structural question has a sharper semantics: it denotes a measurement in a 3D coordinate system. We introduce ProtStructQA, an executable benchmark for protein structural question answering in which each natural-language question is generated from a hidden typed domain-specific language (DSL) program and the answer is obtained by executing that program on an AlphaFold-predicted structure. ProtStructQA releases 382.2K questions covering confidence, distances, predicted aligned error (PAE), solvent exposure, secondary structure, topology and contacts, and held-out compositions: a 330K active benchmark over 10K proteins from four species, plus a 52.2K hard-negative robustness pool. Without fine-tuning, we evaluate Qwen3 models from 0.6B to 8B under direct prompting, chain-of-thought, grammar-constrained executable voting, executable voting with chain-of-thought, and multi-turn ReAct-style tool use, and replicate the headline finding on Gemma-3-1B and Gemma-3-12B. We find a capability-dependent denotation threshold between Qwen3-1.7B and Qwen3-4B: below it, tool-mediated ReAct dominates because models often fail to produce executable denotations; above it, chain-of-thought flips from mostly harmful to strongly beneficial and becomes the strongest strategy on most splits. Parse-failure and family-level analyses show that the threshold is a transition from unparseable language to executable structural denotation, while grammar and execution remain selectively valuable for PAE and secondary-structure queries. ProtStructQA reframes scientific QA as compilation from language to measurement and provides a diagnostic testbed for when language models can map words to executable 3D structural measurements.

3D Gaussian Splatting (3DGS) is a technique for creating and rendering 3D scenes, however its performance depends heavily on the quality of initial seed points. To improve 3DGS initialization, this study presents and evaluates several point cloud upsampling approaches: linear interpolation, triangular interpolation, spline-based surface reconstruction, moving least squares surface fitting, and Voronoi-based point generation. Additionally, this research introduces a depth-guided point lifting method that leverages depth maps to maintain geometric consistency with Structure-from-Motion (SfM) reconstructions. Through extensive experiments on the Mip-NeRF360 and Replica datasets, the proposed methods demonstrate improvements in reconstruction quality across diverse scene types. Results indicate that different upsampling strategies excel in different scenarios: surface reconstruction methods perform better with organic, detailed scenes, while simpler interpolation approaches are more suited for scenes dominated by piecewise-smooth geometries. In comparison, the depth-guided approach shows promise for adding geometry-aware points across the entire scene, importantly in texture-less regions. These findings, which provide preliminary practical guidelines for selecting appropriate upsampling methods based on scene characteristics and computational constraints, advances the understanding of how point cloud initialization affects 3DGS quality.

Orientation-aware manipulation is essential in post-harvest agricultural processing, where produce must be grasped and placed in consistent configurations. This paper presents ROG-Grasp, a geometry-based robotic grasping and placement framework that estimates the produce orientation from root surface geometry using RGB-D perception. A YOLO-based root detector and point cloud plane fitting are used to infer the root normal, enabling stable grasp pose generation and orientation-constrained Cartesian motion planning. Experiments on tomatoes and onions demonstrate high success rates and stable execution time in both isolated and cluttered scenarios. Compared with vision-language-action (VLA) policies, the proposed method achieves more reliable and accurate grasp completion with faster execution. These results highlight the effectiveness of geometry-driven perception for practical orientation-controlled manipulation tasks. A video of our paper is available online https://youtu.be/Ir2UtGODdMo.

Open-vocabulary 3D scene segmentation usually assumes RGB-D video, calibrated multi-view imagery, or a reconstructed mesh. GeoSAM-3D studies a lighter setting: a user uploads a short monocular video, clicks or names an object in one frame, and receives a propagated 3D mask over a Gaussian scene. The implementation combines frozen image and video foundation models with a monocular 3D Gaussian Splatting reconstruction and a differentiable graph-geodesic propagation kernel over Gaussian centroids. The central design choice is to propagate prompts by heat-kernel distance on the reconstructed scene graph, rather than by Euclidean nearest neighbors in 3D. This preserves continuity around curved surfaces and reduces leakage across nearby but disconnected objects. This paper describes the repository state, the mathematical kernel implemented in geosam3d.propagate, the feature head trained from Segment Anything masks, and the validation already present in the codebase. The evaluation protocol separates implementation validation, graph propagation quality, leakage control, and interactive latency.

Dark vessel detection requires fusing what vessels report through AIS with what satellites observe through radar and optical sensors. DarkVesselNet is a multi-modal remote sensing stack that combines Sentinel-1 SAR, Sentinel-2 optical imagery, geospatial foundation model backbones, AIS trajectory reasoning, TGARD-style gap detection, and a Pi-DPM-inspired anomaly head. The repository exposes the system as a tested Python package and a public Hugging Face Space. The paper presents the sensor stack, backbone abstraction, fusion path, anomaly head, and current validation. The evidence currently available is software-grounded: tests for SAR speckle filtering, optical band ratios, Haversine distance, TGARD gap emission, sensor coregistration, backbone token shapes, and differentiable anomaly scoring.

Integrating dynamic 3D reconstructions into physics simulation requires fixed mesh topology for efficient collision detection, but state-of-the-art methods like DG-Mesh produce varying topology optimized for geometric quality. We investigate whether topology conversion can enable physics integration while preserving reconstruction fidelity. We propose a dual-representation framework combining fixed-topology meshes for physics with Gaussian splatting for rendering, achieving 4.65$\times$ speedup over varying-topology baselines through runtime vertex buffer updates. We evaluate two conversion strategies, temporal correspondence tracking and template-based projection, against native fixed-topology methods (MaGS) on the DG-Mesh dataset. Our evaluation reveals that both conversion approaches incur 65-80% geometric degradation, producing results inferior to MaGS despite DG-Mesh's superior initial quality. This demonstrates that high-quality reconstruction and physics-compatible topology represent fundamentally distinct objectives that cannot be reconciled through post-processing. Our findings inform future development of physics-aware reconstruction methods and our framework enables real-time simulation with any fixed-topology approach.

Many machine learning techniques rely on approximating a loss function's curvature, but this is notoriously hard to do at the scale of modern deep networks. Surprisingly, no previous work has exploited the curvature constraints that arise from well known weight-space symmetries in loss landscapes. By analytically averaging over group actions that leave the loss invariant, we construct structured Hessian approximations from single gradients that can be tractably estimated, stored, and inverted. The choice of user-specified symmetry group directly governs the trade-off between approximation accuracy and computational cost. Moreover, our framework provides a unifying theoretical lens for viewing existing methods; in particular, a specific choice of symmetry group recovers Shampoo/Muon-like curvature estimates. We validate our method on a range of network architectures, and deploy it to second-order optimization benchmarks, including a small language model. Our curvature estimation framework might find applications in other machine learning problems such as uncertainty estimation, continual learning, compression/pruning, training data attribution, and more.

Reinforcement learning with verifiable rewards has rapidly advanced reasoning in vision--language models. However, for chest X-ray report generation, the standard rewards (i.e. exact-match accuracy and step-level processes) are incompatible because the reports consist of unordered and orthogonal findings, rather than a causal reasoning chain. We address this gap with a set-based view: each report is split into sentences and embedded by a frozen sentence transformer, yielding unordered embedding sets. We propose the use of set-to-set distances between generated and reference embeddings as continuous, permutation-invariant rewards. Across two datasets and three vision--language models (Qwen3-VL-2B/4B, Gemma3-4B), post-training with set-to-set distance based rewards via GRPO consistently outperforms supervised fine-tuning and exact-match GRPO on all headline metrics (BERTScore, RadGraph F1 and CheXbert F1 by average \%6.80, \%7.82 and \%4.45 relative improvements respectively). The same set distances also enable test-time best-of-$N$ selection: scoring candidates by their distance to training-report embeddings outperforms random selection on our trained models as well as three closed-source LLMs (Mistral-Small, Gemini-2.5 Flash-Lite, GPT-4o-mini) with on average \%16.4 relative improvement on BERTScore. Used as a streaming signal, they support a more efficient form of test-time scaling: pruning low-scoring candidates mid-generation reduces generated tokens by over 50\% while preserving the Findings quality of full best-of-$N$ selection. Together these results establish set-distance rewards as a unified signal for both post-training and test-time scaling in chest X-ray report generation. Our code is publicly \href{https://anonymous.4open.science/r/Set-Distance-Rewards-CXR-BFDA}{available}.

A central challenge in visual intelligence is learning the physical structure of scenes from raw videos: how regions form objects and the laws that govern their interactions. Solving these tasks requires world models capable of inferring distributional states of the world from partial observations - capabilities that current architectures do not provide. We introduce a new class of probabilistic world models that support estimation of the probability of any visual variable, such as appearance and dynamics, conditioned on any other variables. Here, we identify that these models can be trained efficiently with autoregressive sequence modeling, yielding world models from which rich object understanding emerges. First, we demonstrate that our model captures the physical laws governing how objects move by generating multiple plausible future states of the world through sequential inference. Then, by analyzing motion correlations across these futures, we extract objects and articulated object subparts. Having discovered these objects, we show that our world model can manipulate them in 3D. Finally, we demonstrate how physical relationships between objects can be computed from the world model, enabling applications such as Visual Jenga.

Process reward models (PRMs) are widely used in language-model training with dense step-level supervision. They assume PRM scores are stable proxies for step correctness under label-preserving transformations. These transformations change reasoning structure but preserve final answers. We argue this assumption is not well validated. Such transformations can change how PRM scores relate to correctness signals, leading to different failure modes across models.To address this gap, we introduce \textbf{EST-PRM}, a stress-testing framework for dense process rewards. It applies three transformations: (1) step inflation, (2) dependency-aware step reordering, and (3) confidence markers. A vulnerability decomposition is defined that separates reward inflation from loss of correctness sensitivity. Five PRM-style models are evaluated on 4,687 reasoning chains from MATH-500, GSM8K, and PRMBench.The results indicate clear differences in vulnerability patterns across models. Math-Shepherd shows the strongest sensitivity to position perturbations, with a Pearson correlation drop of $0.152 \pm 0.038$ and a $32.8 \pm 4.9\%$ score inflation rate. Qwen2.5-Math-PRM is most affected by step inflation, reaching a $47.6 \pm 4.3\%$ inflation rate. Confidence-based perturbations also distort reward calibration, revealing inconsistencies in correctness estimation. Three mitigation strategies are evaluated, highlighting trade-offs between robustness coverage and false-positive rates.

As retrieval-augmented generation (RAG) systems scale, it becomes increasingly challenging to ensure faithful grounding in external evidence. Large language models may still prioritize parametric knowledge over retrieved information when conflicts arise. We propose a novel training-free decoding framework, \emph{Grounded Decoding}, designed to improve factual consistency in RAG without modifying model parameters. Unlike standard approaches that rely on a single conditional distribution, our method constructs two matched-prompt distributions at every generation step: (1) a full RAG distribution conditioned on the query, retrieved documents, and generated prefix, and (2) a retrieval-only distribution conditioned solely on retrieved evidence and the same prefix. The final next-token distribution is derived as the unique solution to a KL-barycenter objective over the probability simplex, yielding a normalized geometric fusion of the two distributions.This formulation naturally recovers standard RAG when the grounding weight is zero and smoothly shifts probability mass toward retrieved evidence as grounding strength increases. We further introduce a conflict-aware adaptive weighting scheme that dynamically adjusts grounding based on distributional disagreement and retriever confidence. Experiments on ALCE, Natural Questions, and FActScore demonstrate consistent improvements in factual accuracy and citation quality over standard RAG and competitive decoding-time baselines, while maintaining fluency. Our results indicate that probability-level fusion provides a strong and efficient alternative to logit-level intervention methods for faithful RAG decoding.

We prove a variance-sensitive regret bound for Thompson sampling in stochastic generalised linear bandits. The argument assumes a warm-up, after which the regret is controlled through using the Gaussian Poincaré inequality. This bypasses the point at which previous optimism-based analyses break down. Removing the warm-up while retaining the same variance-sensitive scaling remains open, and appears nontrivial.