Safe offline reinforcement learning (RL) seeks to learn policies that maximize return while satisfying safety constraints, using a static dataset. To characterize state-wise safety, reachability analysis identifies the feasible region, defined as the set of states from which constraints can be satisfied at every future step. Policies are commonly extracted from learned value functions with advantage-weighted regression (AWR). However, AWR relies on exponentiated weights that amplify value estimation errors, whereas behavior-regularized policy gradient methods directly utilize the action gradient of the value function and often make more effective use of learned critics. We propose Feasibility-Optimized Conditional Actor Learning (FOCAL), a method that incorporates behavior-regularized policy gradients into a conditional variational autoencoder (CVAE) actor. FOCAL optimizes this policy using a first-order objective partitioned by feasibility, updating the actor to increase return on transitions estimated to be feasible and to reduce constraint violations on transitions estimated to be infeasible. During inference, FOCAL dynamically adjusts the scale of the latent prior according to the estimated feasibility of the state to adjust the conservatism of action sampling. Theoretically, we show that for AWR, the policy update variance depends exponentially on the advantage error, whereas for FOCAL it scales linearly with the variance of the value estimation error. Across the DSRL benchmark, FOCAL satisfies the constraints on all evaluated tasks while achieving competitive or higher returns compared to baselines.
Knowledge distillation (KD) between architectural families, e.g., from a Transformer teacher to a CNN student, is difficult because the two networks learn different representations. Logits share a label space, but heterogeneous teachers differ in logit offset, scale, and sharpness, and their non-target classes are less semantically organized. Conventional logit distillation matches these values sample by sample, asking the student to imitate how another architecture expresses its predictions. We argue that a heterogeneous teacher should pass on the relational structure of its predictions, i.e., how it arranges samples relative to one another, and propose Heterogeneous Asymmetric Contrastive Distillation (HACD) to transfer this structure. HACD maps teacher and student logits to centered, temperature-softened probability directions on a shared unit hypersphere, which removes architecture-dependent offsets and lets the temperature control how sharply peaked outputs enter the representation while keeping informative non-target classes. Each student output must then identify its own teacher target within the batch. Because every teacher target also acts as a negative for the others, HACD computes the targets on a moderately augmented view and contrasts both a moderate and a strong student view against them. HACD adds no learnable parameters and needs one teacher pass per sample. It achieves the best or second-best accuracy on 11 of 12 heterogeneous pairs on CIFAR-100, without ground-truth labels, and on 14 of 15 pairs on ImageNet-1K, improving over the reported vanilla KD results by 8.82 and 1.33 percentage points on average. Checkpoints are available at https://github.com/BTAEYANG/HACD-Heterogeneous-Asymmetric-Contrastive-Distillation; the code will be released upon publication.
Discovering causal structures in multivariate time series (MTS) is critical in domains such as finance and neuroscience, yet remains challenging in practice due to three distinct obstacles. First, MTS are routinely corrupted by missing values, where naive impute-then-discover pipelines distort the underlying data distribution and yield spurious causal graphs. Second, identifying instantaneous effects is difficult even on complete data, and missingness further entangles imputation with graph recovery. Third, real-world data frequently exhibits heteroscedastic noise with variance driven by both instantaneous and lagged causes, violating the core assumptions of standard discovery methods. Existing approaches fail to address these challenges concurrently, typically presuming complete data, homoscedastic noise, or strictly time-lagged relationships. Crucially, we show through ablation studies that resolving these issues sequentially, rather than jointly, leads to a substantial degradation in performance. To overcome this, we introduce Cheesefill, an optimal transport framework for causal structure learning under missingness and heteroscedastic noise. Cheesefill parameterizes causal mechanisms using conditional normalizing flows and jointly learns a stochastic correction map that refines naive imputations into trajectories consistent with the inferred causal mechanisms. Theoretically, we prove that optimizing over stochastic correction maps is equivalent to a conditional kernel reformulation of the Kantorovich problem, from which our neural parameterization yields a tractable objective. Extensive experiments on synthetic and real-world benchmarks demonstrate that Cheesefill consistently outperforms existing baselines across diverse missingness regimes.
Low-light conditions reduce visible contrast and obscure details, while infrared images lack color and texture, motivating the fusion of infrared and visible information for visual tasks. Existing fusion methods primarily optimize image appearance or task-specific features without explicitly disentangling the coupled factors underlying thermal observations, limiting their physical interpretation under thermal degeneracy. Hyperspectral thermal measurements enable physical decomposition, but they require specialized imaging hardware, while pixel-level physical annotations for supervision or evaluation incur additional costs. We propose TeX-Gen, a physically guided framework for TeX generation from paired visible and infrared images in low-light visual tasks. TeX Generation constructs complementary multi-view features and employs independent branch-specific gates to route cues into the Temperature, Emissivity, and Texture maps. TeXMetric integrates radiometric reconstruction, channel structure, and material consistency to assess the generated maps without requiring pixel-level physical labels. TeX Tune improves performance through task-guided adjustments, jointly optimizing the adjustment modules and task network while keeping the generator frozen. Experiments show that TeX-Gen achieves competitive task performance in low-light environments on the FLIR, LLVIP, M3FD, and MSRS datasets. Experiments on HADAR also examine the association between TeXMetric scores and downstream detection performance.
This paper presents enhancements to the DINOBot framework for imitation learning of object-object interactions. We introduce two key contributions aimed at improving the accuracy and efficiency of alignment in imitation-based robotic tasks. First, we integrate RANdom SAmple Consensus (RANSAC) for outlier rejection during the alignment phase, effectively mitigating the impact of noisy depth data and spurious feature matches. This improves robustness when transferring demonstrations to real-world applications. Second, we refine the semantic correspondence process by focusing on task-relevant interaction regions, ensuring that the learned representation prioritizes critical object features that influence successful manipulation. We evaluated our approach in a range of real-world manipulation tasks, including picking, insertion, shaking, and pouring, using a robotic system. The experimental results demonstrate an overall improvement of 13.8% in task success rates compared to the baseline, along with faster convergence in the alignment process. These results highlight the effectiveness of our enhancements in improving precision and efficiency for object-object interaction imitation in robotics.
Latent world models enable visual control by planning action sequences in compact representation spaces, but a gap remains between how latent dynamics are learned and how they are used for planning. First, teacher-forced training does not expose the dynamics model to the recursively predicted latent states encountered during planning. Second, sampling-based planners typically initialize action search from goal-agnostic proposals, without exploiting the action-conditioned dynamics learned by the world model. We introduce Dynamics-Informed Action Learning (DIAL) to bridge these two gaps. During training, rollout alignment exposes the forward model to its own autoregressive latent predictions, while multi-step inverse supervision trains a lightweight inverse dynamics head to infer actions from latent-state transitions. We further apply the inverse head to model-generated transitions with its parameters held fixed, encouraging predicted rollouts to retain action-relevant information. At test time, DIAL autoregressively couples the inverse and forward dynamics to construct a dynamics-informed action sequence toward the goal, which initializes the mean of the Cross-Entropy Method (CEM) proposal distribution. The subsequent CEM search then refines this initialization using the standard sampling and updating procedure. Experiments across four visual control benchmarks demonstrate that DIAL outperforms existing latent world models and planning baselines in both standard and long-horizon scenarios.
Diarization-conditioned target-speaker ASR (TS-ASR), in particular Diarization-conditioned Whisper (DiCoW), is an accurate and data-efficient approach to multi-talker transcription. It has, however, three structural limitations: the model transcribes each speaker independently, operates on short (30 s) windows, and optimizes its diarization and TS-ASR components separately. We address the first two. We integrate the diarization conditioning into a longform Parakeet backbone (DiCoP) and add cross-speaker attention, which lets the model attend across the speaker dimension so that speakers are no longer decoded in isolation, and an STNO mask refiner that re-estimates the diarization-derived conditioning. As conversational data with genuine multi-speaker interaction is scarce, we scale training with a pseudo-labeling pipeline that mines 5.4k hours of in-the-wild multi-speaker audio from YODAS2. Combining these with synthetic and real data yields state-of-the-art results across several meeting benchmarks. We open-source the resulting models, DiCoP and Pandemonium, and the pseudo-labeled data.
Cellular world models aim to predict how cells respond to genetic perturbations, analogous to world models in reinforcement learning. Perturbation response models are evaluated almost exclusively by global Pearson correlation between predicted and observed transcriptomes. We show this metric is dominated by thousands of non-responsive genes and severely inflates apparent performance. On the Replogle K562 screen (509 unseen perturbations, 5000 genes), GEARS achieves global Pearson 0.977, but its delta Pearson is only 0.076. A trivial baseline predicting mean training delta achieves 0.267, which is 3.5x better than GEARS. Our primary contribution is methodological: we identify and quantify this delta gap as a systematic evaluation artifact, and propose delta-space evaluation as the primary benchmark. We provide two post-hoc corrections (RWR on co-expression graphs and kNN matching) and show that 2.8x of the remaining gap lies in perturbation representation geometry, not decoder capacity. We further show the gap is partially addressable end-to-end: delta-aligned training (lambda = 3.0) raises delta Pearson to 0.179 on Replogle (2.35x over GEARS) and surpasses the trivial baseline on direction accuracy (0.680 vs 0.567) and gene recovery (1.52 vs 1.19); combined with kNN it reaches 0.217. An amplitude-calibration variant (eta = 0.3) further raises gene recovery to 2.16 while reaching 0.218. Critically, these representations transfer zero-shot across cell lines: from Replogle K562 to unseen Norman A549, the delta-aligned model reaches 0.328 on 741 shared genes, 4.7x a trivial baseline and 2.1x standard GEARS. This is the first systematic quantification of the delta gap, and the first diagnostic framework localizing the residual gap to perturbation representation geometry via an oracle upper bound.
Driven by increasingly complex real-world applications, Retrieval-Augmented Generation (RAG) has evolved from textual pipelines to multimodal settings, comprising two complementary dimensions: knowledge organization and retrieval enhancement. However, existing methods still face challenges of high construction cost, context window limitations, and the lack of unified training across offline and online knowledge sources. To address these challenges, we propose MMGraph-Agent, a unified multimodal RAG framework based on cache-inspired multimodal knowledge hypergraphs. MMGraph-Agent enables extremely lightweight hypergraph construction with zero API cost and dynamic memory-based retrieval that alleviates long-context issues. Experiments across offline, online, and joint retrieval settings demonstrate consistent improvements in performance, efficiency, and architectural unification. Our software and data are publicly available at https://github.com/LHRLAB/MMGraph-Agent.
Dense prediction and generative modeling in resource-constrained scenarios remain challenging, as repeated downsampling substantially reduces spatial detail and degrades boundaries. Although recent attention mechanisms have alleviated the problem of semantic ambiguity, they often fail to preserve high-frequency features or do so at the expense of a very high computational cost. In this paper, we propose the Frequency Guided Depthwise Separable Attention for Efficient Dense Prediction (FG-DSA), a novel framework for efficient dense visual understanding and generation. FG-DSA employs a 2D FFT high-pass filtering module to explicitly inject frequency domain priors into the spatial attention mechanism, allowing lightweight networks to recover sharp boundary transitions with linear complexity. A range of experiments show that FG-DSA is effective at recovering fine object boundaries and leads to a consistent performance gain in the task of semantic segmentation. Importantly, we show that FGA-DSA can be used as a module in diffusion-based generative U-Nets, leading to a significant improvement in the detail of the generated images with very little additional parameter and Floating Point Operation (FLOP) cost.
3D Gaussian Splatting (3DGS) provides high-quality real-time novel-view synthesis, but deployment budgets vary across devices and network conditions. Most compact methods target one operating point, so another budget requires a separate representation. SHIN is a scalable 3DGS hierarchy with a reusable Base and additive enhancement layers, so each higher-quality operating point extends rather than replaces the decoded representation. Constructing this hierarchy requires deciding which Gaussians each stage adds and how much spherical-harmonic (SH) appearance each retains, because these choices remain fixed in later prefixes. SHIN derives one shared Gaussian order from multi-scale position-scale sensitivity and divides each stage budget between adding Gaussians and retaining SH bands using calibrated gradient-based scores per byte. Lightweight opacity updates adjust compositing while transmitted geometry and SH coefficients remain unchanged. Across three benchmarks, SHIN progressively improves one cumulative representation and achieves the best PSNR-size efficiency in every overlapping comparison with the multi-rate and progressive baselines.
Transformer-based autoregressive multilingual language models (LMs) are often assumed to store factual knowledge in an interlingua: a language-agnostic latent space. If this assumption holds, a parameter update applied in a source language during knowledge editing should inherently propagate to its semantic equivalents in target languages. However, empirical benchmarks show that cross-lingual propagation of edits often degrades substantially, a degradation that prior work treats largely as an algorithmic symptom rather than a structural bottleneck. In this work, we investigate whether representational anisotropy creates a structural bottleneck for source-only associative-memory editing, and perform a geometric dissection of the cross-lingual propagation gap. Using associative memory-based "locate-then-edit" methods, including ROME, MEMIT, and AlphaEdit, as analytical frameworks, we derive a closed-form decomposition of this gap into key-space propagation, parallel residual mismatch, and orthogonal residual mismatch. Our empirical findings show that cross-lingual propagation is primarily bottlenecked by a large orthogonal residual component, indicating that source-language edits often fail to span the target-language update directions required for successful transfer. Furthermore, we show that post-hoc correction using only source-language residual statistics encounters a large residual-prediction barrier. Finally, we perform a structural realignment via a targeted contrastive alignment objective and find that it substantially improves key-space propagation and partially improves cross-lingual edit transfer, while leaving value-space mismatch as a remaining bottleneck.
AI agents can generate as-built Building Information Models (BIMs) from reality-capture data, but the evidence required for accurate reconstruction remains unclear. This paper presents Agent2BIM, an agent-based Scan-to-BIM framework, and evaluates seven capture configurations in a research laboratory and a mechanical, electrical, and plumbing (MEP) scene. Inputs include 360{\deg} panoramas, smartphone video, smartphone LiDAR, terrestrial laser scanning (TLS), TLS with posed panoramas, and imagery supplemented by prior IFC models. Outputs are compared against TLS-verified reference models for component detection, geometric accuracy, and MEP connectivity. In the laboratory, component recall ranged from 84% to 100%, but only 9% of components reconstructed from panoramas were within 20 cm of their reference positions, compared with 86% using smartphone LiDAR. In the MEP scene, scale-free inputs recovered 11 - 43% of components, versus 75 - 93% for metric inputs. Prior IFC models improved MEP reconstruction but could propagate inaccuracies. The results highlight the complementary roles of visual and metric evidence in agent-based Scan-to-BIM.
Visual citations show where an answer is supported, but not how the cited content contributes to it. In scientific document question answering, a model may read a value directly, compress several observations, or combine evidence to infer a conclusion. We study typed visual provenance, where a system generates an answer together with source regions and explicit support relations. We introduce ReViE, a dataset of 3,941 questions over complete scientific documents, with page, section, and region annotations labeled as Quotation, Compression, or Inference. We further propose VisProve, which augments joint answer and provenance generation with differentiable Spatial Evidence Reading and Evidence-Answer Relation Feedback. VisProve conditions next-token prediction on localized visual evidence and its estimated support relation through a gated norm-bounded residual. On the document-disjoint ReViE test set, VisProve improves Typed-F1 from 3.52 to 5.48 and Section-F1 from 45.99 to 53.24 over joint supervised fine-tuning with the same backbone, while Page-F1 changes only marginally. These results suggest that coupling visual grounding with support relation modeling mainly improves fine-grained provenance, while the low absolute Typed-F1 reveals a substantial gap between answer generation and precise visual evidence attribution.
Simulations of LLM agent collectives usually summarise interaction as a pairwise contact network. We test that summary with a contact-matched intervention: the members, their initial stances and every pairwise contact count stay fixed while the meetings are reorganised, and each member gives a one-step support readout after reading the arguments of everyone present. Across six open-weight LLMs and 14 debate topics, merging triads into five-person meetings (resize) moves group support more than reshuffling who meets in triads (regroup), although both move the same 60 dyadic exposures. The root-mean-square change is 4.36 versus 2.59 percentage points (pp), larger for resize in all six models and in 81 of 84 model-topic cells. Resizing also has a consistent direction: it amplifies the initial majority of the merged members by 1.72 pp, in every model, whereas regrouping has no direction (<= 0.03 pp). A contact network cannot represent regrouping at all, yet pairwise rules applied meeting by meeting explain up to 76% of its effect, and with identical weights they raise resize R^2 by 0.51-0.58 for two models. We show that the gap between the two placements is an exact Jensen gap of the sigmoid readout, bounded by how much a member's predicted logits vary across meetings; this dispersion tracks the gap with log-log r=0.97 and tells a modeller when contact counts are a safe summary. Meeting format is a first-class variable for LLM social simulation.
Deep reinforcement learning (RL) is often used to decide where Internet of Things (IoT) tasks with different latency limits run in the edge-cloud continuum, yet some methods give every missed task the same penalty and few compare their agents with an optimal policy. This benchmark study builds four offloading systems small enough to solve exactly by dynamic programming and measures reward and learner against the optimum. A bonus for a met limit makes each miss cost its penalty plus the bonus, which shifts the class weights of published per-class rewards. A class-blind penalty stays 65% to 118% above the optimal cost even with a perfect learner, while a corrected per-class penalty comes within 0.13% of the optimum. With a per-class reward, tuned proximal policy optimization (PPO) agents come within 9% to 11% of the optimum in the smallest system, and the corrected one beats a tuned reservation rule in 9 of 10 seeds. In larger systems the agents stay 41% to 76% above the optimum, and a tuned rule beats their mean. Training ten times longer does not close this gap, and a PPO-Lagrangian agent with reachable targets does no better than the per-class reward.
Full projector compensation adjusts a projector's input image to compensate for both geometric and photometric distortions introduced by the projection surface. Existing methods generally treat the viewing condition as a static camera viewpoint. When applied to a moving viewer, existing methods face two key challenges: (1) view-dependent methods require model retraining or recalibration to compensate for novel viewpoints; and (2) NeRF/GS-based view-agnostic methods use structure from motion (SfM) to calibrate the projector-camera systems (ProCams) with an unknown global scale, while view-aware tracking requires calibrating a camera oriented opposite to the projector, which introduces an additional calibration challenge. To address these challenges, we introduce VAComp, the first view-aware framework for full projector compensation for a moving viewer in a naked-eye setting. Using a projector, VAComp enables the viewer to perceive a virtual object that visually replaces the real object or to observe view-consistent modifications of the real object's appearance and material properties. To resolve the SfM scale ambiguity, we use metric calibration as a reference to recover the global scale; to calibrate the oppositely facing camera, we use two checkerboards, one facing the RGB camera and the other facing the depth camera. To balance compensation quality and runtime efficiency, we develop two non-iterative hybrid methods, VAComp-SD and VAComp-SI. Both combine 2D Gaussian-based view-conditioned geometric correction with learned photometric compensation, predicting each compensation image through a single forward pass without per-frame iterative optimization. Real-world experiments show that both methods achieve competitive compensation quality, with VAComp-SD enabling real-time updates and VAComp-SI operating at interactive rates. Project page: https://realkaizhou.github.io/VAComp/.
Sparse attention accelerates long-context decoding by reading a subset of the KV cache, but existing methods are open-loop: the kernel commits to its selection with no signal indicating whether enough attention mass was preserved and no recovery path when it was not. The needed signal already exists inside the kernel. The fraction of softmax mass on the selected tokens, which we call coverage, satisfies an exact algebraic equality with per-step output error and serves a dual role as the acceptance criterion and the interpolation weight for recovery. Coverage-Verified Sparse Attention (CVSA) turns coverage into a training-free, closed-loop verify-then-correct decode loop that accepts high-coverage drafts and recovers only the heads that fall short, with no sampling, no draft model, and no speculative-decoding infrastructure. Quality varies by only $0.14$ points across the threshold sweep; budget and scheduling ablations characterize the deployed operating point. On LongBench at 7B and 70B, CVSA comes within $1.13$ points of dense attention, statistically matching it at 70B, and it loses $0.8$ points on RULER; measured median decode-step speedup reaches $6.21\times$ at 128K context.
Quantified Boolean formula (QBF), the quantified version of Boolean satisfiability (SAT), is a fundamental problem in constraint solving. Due to the $\mathsf{PSPACE}$-complete nature of QBF, designing theoretically efficient QBF algorithms is much harder and receives less attention in academia. In this work, we present a new algorithm based on memoized search, which solves QBF with $n$ variables and $m$ clauses in $1.685^m\cdot\text{poly}(n)$ time and space. This is the first improvement over the $1.709^m\cdot\text{poly}(n)$ bound by Williams in over 20 years. Our technique is based on the minimum vertex cover of the conflict graphs, which was proposed to study the parameterized complexity of SAT. We extend this idea to QBF and further show that QBF is slicewise polynomial-time solvable using this parameter.
We introduce Latent-WAM++, a representation-centered World Action Model (WAM) for end-to-end autonomous driving. Existing WAMs are primarily trained to predict future states or final actions, leaving the intermediate world and action representations that connect prediction to planning only indirectly constrained. Latent-WAM++ matches the supervision of each representation to its role in planning. World Representation Construction combines scene-ego future prediction, SIGReg-based distribution stabilization, and frozen geometric priors to learn a temporally predictive, non-degenerate, and geometry-aware world representation without an EMA target encoder. Trajectory MeanFlow Planner directly predicts multimodal trajectory proposals while being optimized with an Improved MeanFlow velocity-space objective, enabling one-step sampling. Visual-Action Representation Alignment further aligns the resulting pre-output action features with a stop-gradient target derived from current visual semantics. Latent-WAM++ achieves state-of-the-art performance on NAVSIM v2 and under zero-shot evaluation on HUGSIM, reaching 90.1 EPDMS on NAVSIM v2 and 47.0 average RC and 30.1 average HD-Score on HUGSIM. Training the planner end to end costs 190 GPU-hours for Latent-WAM++ versus 1,600 for Latent-WAM, a reduction of 88.1%.