Transit-Oriented Development (TOD) is a key strategy for advancing sustainable urban development by promoting high-density, mixed-use land development within walking distance of public transit stations. However, traditional TOD planning often relies on heuristic methods, making it difficult to efficiently identify quantitatively optimized transit-oriented land use patterns. This paper presents DRL4TOD, a deep reinforcement learning framework that automatically optimizes land use types and floor area ratios (FARs) around transit stations in alignment with TOD principles. DRL4TOD incorporates a graph neural network to capture spatial interdependencies among land parcels, road networks, and metro stations, and uses a DRL-based optimization strategy to assign land-use types and FARs to selected parcels. It jointly optimizes multiple objectives, including land use density, land use diversity, and transit capacity utilization, while minimizing reconstruction costs to account for development feasibility. Applied to two real-world station areas in Shanghai, DRL4TOD consistently outperforms rule-based approaches and genetic algorithms, achieving more effective TOD outcomes with lower reconstruction costs. By adjusting objective weights, the framework enables flexible prioritization, such as placing greater emphasis on cost minimization when financial feasibility is a concern. These results demonstrate DRL4TOD's potential as a decision-support tool for generating coordinated land-use and FAR configurations under planner-specified priorities, rather than prescribing a single fixed design. For policymakers, the framework provides parcel-level, constraint-aware alternatives that make trade-offs among competing objectives explicit and support targeted decisions about development intensity and feasibility.
Safety-aligned language models are known to suffer degradation in refusal behavior following domain-specific fine-tuning, even when the fine-tuning data is entirely benign. The predominant mitigation strategy, preference optimization on the fine-tuned model, is computationally expensive and requires domain-specific preference data that is often unavailable in practice. We propose SPECTRA, a training-free, weight-level method that restores refusal behavior by transferring the base model's refusal direction into the fine-tuned model's weights via an activation-aware, low-rank update, requiring no gradient-based optimization or domain-specific safety data. The low-rank design restricts the intervention to the subspace most causally active on harmful inputs, leaving benign behavior largely undisturbed by construction. Empirically, SPECTRA substantially improves refusal behavior across a comprehensive safety evaluation, while preserving general coherence, domain-specific knowledge, and producing negligible change in false refusal rates.
Purpose: To develop and validate a scalable adversarial evaluation framework for assessing the robustness of large language models (LLMs) that generate patient-friendly summaries of radiology reports. Methods: We introduced a clinically grounded adversarial framework that uses controlled term substitutions and impression removals across intensity levels. Using a radiologist-validated adversarial dictionary, we generated adversarial variants of CT and MRI reports. Six LLMs summarized original and adversarial impressions using 2 prompt styles, yielding 18,000 evaluation pairs. Robustness was evaluated with an automated LLM-as-a-judge framework scoring factual faithfulness, sensitivity to adversarial edits, certainty fidelity, and hallucination rate. Failure mode analysis quantified vulnerabilities in incorporating clinical modifiers. Agreement with two radiologists was assessed on 105 cases. Results: GPT-5-mini achieved the highest factual faithfulness and certainty fidelity under both adversarial edits. Under term substitution, hallucinations occurred in 28.1-30.1% of GPT-5-mini summaries, compared with 35.3-36.4% for Claude Sonnet 4.5 and 41.4-41.9% for MedGemma. Impression-removal edits were handled more robustly than term substitutions. Even for the highest-scoring summarizer, substitutions of anatomical terms were not reflected in 23.7% of summaries, diagnostic certainty in 22.3%, and descriptor terms (e.g., "substantial") in 21.2%. Judge-radiologist agreement (macro-F1: 0.55-0.56) was comparable to inter-radiologist agreement (0.53). Conclusions: Adversarial evaluation exposes clinically meaningful failure modes, including missed anatomical modifiers, altered diagnostic certainty, and omitted severity descriptors.
Low-budget molecular optimization is often framed as Bayesian optimization in a learned latent space, but standard benchmarks typically allow hundreds or thousands of evaluations, whereas a realistic wet-lab round may allow only a few dozen. We revisit this regime using a SELFIES molecular variational autoencoder (VAE) with a Gaussian prior and GuacaMol objectives at N = 30 evaluations. The main finding is not a new surrogate model, but an interaction in candidate generation. The encoder means of a reference set of molecules concentrate near a narrow spherical shell, while many high-dimensional optimizers propose candidates from boxes or box-shaped trust regions. We study a simple constraint, the on-sphere scaffold, that restricts candidate generation to this shell. We then introduce two minimal optimizers to probe what remains useful once the support is fixed. GeoWalk performs geodesic local search on the shell without a surrogate. After 30 evaluations, it shows no detected difference from a Gaussian-process (GP) surrogate on the shell, which is ahead earlier in the budget. GeoCMA adds covariance learning in a random subspace. It does not improve on GeoWalk in the molecular benchmark. On LassoBench it has a lower mean loss than GeoWalk, but full-dimensional CMA-ES remains stronger on most tasks. A factorial experiment with a single GP crosses the candidate support (shell or box) with the acquisition maximizer (a finite candidate set or a continuous optimizer). In this benchmark, shell candidates outperformed box candidates when acquisition was maximized over a candidate set, but not with continuous maximization by L-BFGS-B. These results come from a controlled study of one Gaussian-prior VAE benchmark and do not recommend VAEs for molecular optimization. The practical lesson is to control the candidate support and the acquisition maximizer before attributing a gain to the surrogate.
Languages use different morphosyntactic strategies to express the same meaning. Speakers of multiple languages are able to apply strategies from one language to another, either by mistake or intentionally, to produce a literal translation, or calque. Considering that large language models (LLMs) have been shown to generalize crosslingually, despite being predominantly trained on English, we investigate whether they overgeneralize constructional strategies from English to other languages. To test this, we create a dataset of 5583 sentence pairs across seven languages, matching literal translations of strategies from English into other languages, and literal translations of the other languages' strategies into English. We then quantify constructional bleeding by measuring how often strategies from English are assigned a higher probability than native strategies. We find that models often fail to assign higher probabilities to native strategies and tend to prefer English-derived strategies even when evaluating sentences in other languages. Our findings highlight the limitations of the current English-dominant training paradigm.
LLMs are increasingly used to support decisions that require moral consideration. While LLMs perform well on morality benchmarks from psychology and philosophy, producing appropriate outputs on these benchmarks (moral performance) does not establish that an LLM weighs competing moral values at stake in a decision (moral competence). We test a minimal, necessary condition for such weighing: that an ordinal ranking can represent LLM moral choices. Such a ranking requires transitivity. To assess transitivity, we presented LLMs with pairwise trade-offs among moral value violations in vignette choice sets and extracted resulting choice probabilities from first-token distributions. Tested against three stochastic transitive choice models, most choice sets admitted weak or unstable orderings. Aggregating an LLM's responses, a common practice in LLM moral evaluation, produced transitive structure that matched no individual choice set. Evaluations of AI morality should establish that choices are structured before asking whether they are aligned.
A camera becomes an absolute pose sensor when its images can be relocalized against a prebuilt map of the scene. On mobile robots and augmented-reality devices, that map must be small and quick to build. Scene coordinate regression (SCR) stores the scene in a compact network that predicts a 3D coordinate for every pixel, and mixture-of-experts (MoE) variants extend its capacity to larger outdoor scenes. In existing MoE-based SCR, however, the router assigns each pixel to an expert from the same local descriptor that the experts regress from, so repeated facades and windows at different locations receive similar routing inputs. We present DAC-MACE, which conditions soft expert routing on recalibrated and contextualized features. A residual context adapter (RCA) reweights the channels and locations of frozen encoder features, a global context encoder (GCE) mixes them with multi-scale self-attention, and an image-level pooled descriptor is added to the router input. A dual-threshold RANSAC--Levenberg--Marquardt estimator recovers the pose. On Cambridge Landmarks, DAC-MACE attains a mean per-scene median translation error of 13.0 cm over three seeds, the lowest among the RGB-only SCR methods compared (GLACE 13.6 cm, MACE 14.4 cm), and reduces the error on the largest scene from 24 to 19 cm relative to MACE. In a cumulative study, the MoE head, GCE, RCA, and the dual-threshold estimator lower the mean error from 17.0 cm to 14.4, 14.0, 13.6, and 13.0 cm. On 7-Scenes, adding GCE to a single ACE head raises Recall@(5 cm, 5$^\circ$) from 86.56% to 87.60%.
Domain generalization (DG) for object detection aims to improve model performance on unseen target domains. Existing detectors, however, suffer substantial degradation on visually distinct unseen domains, as models confined to limited training styles tend to overfit superficial statistical cues rather than object-level semantics. Motivated by human perception, which attends primarily to semantically relevant features while ignoring irrelevant style cues, we propose Origin-Anchored Multi-Style Consistency (OAMSC), a source-only framework that couples multi-style data synthesis with cross-style consistency learning to suppress style-sensitive shortcut learning. For data preparation, we leverage off-the-shelf diffusion models to expand the source datasets into 20 distinct styles, with all generated images manually verified to ensure that object instances and annotations remain consistent across style variants. During training, we impose Origin-anchored consistency constraints at scene and object levels to reinforce the learning of domain-stable features while suppressing the influence of style-sensitive cues. Extensive experiments and ablation studies on multiple challenging DG benchmarks demonstrate the consistent effectiveness and strong generalization performance of OAMSC across diverse domain shifts. Notably, our approach applies to both one-stage and two-stage detectors, while its data-centric design is complementary in principle to other DG methods. The training code and two style-transfer datasets constructed for the experiments, comprising approximately 180K images in total, are publicly available on GitHub.
The phrase recursive self-improvement (RSI) refers to the idea of an AI system designing and creating another, better AI system, which in turn designs and creates an even better one, and so on. While the idea has a long history, it has always been an abstract, futuristic idea -- until recently. Today's frontier AI has attained a set of abilities, including in programming, mathematics, research, tool use, and working coherently over extended periods of time, that make RSI seem plausible. AI companies are starting to pursue it, but many of the involved people are also expressing concerns, asking whether humanity should really set this in motion and under what conditions. I will discuss the situation, argue that we do not even have clear frameworks for thinking about the likely outcomes of RSI, and do a quick game-theoretic analysis of how to choose whether to initiate RSI when another party is making the same choice.
Accelerated magnetic resonance imaging (MRI) forms images from incomplete Fourier measurements, shifting acquisition burden to computational reconstruction. Deep unrolled networks interleave learned regularization and data consistency, yet typically estimate the full unmeasured spectrum at every cascade. We propose STAIRS (Staging the Inverse Problem via Admissible-Support Scheduling), which projects intermediate reconstructions onto nested Fourier supports that expand to the full spectrum and conditions each update on its support transition. Conditional increments of a regularized log-determinant score of training covariance define a normalized spectral coordinate, accounting for dependencies between newly and previously admitted frequencies. Dyadic targets allocate larger coordinate increments to earlier cascades and are mapped to fixed supports before training, enabling coarse-to-fine reconstruction with unchanged sampling masks and data consistency. On fastMRI and Stanford2D, STAIRS improves matched Transformer (HUMUSNet) and convolutional (NAFNet) backbones in single- and multi-coil reconstruction, including radiologist-annotated pathology regions. It also outperforms predictor-based progression with lower parameter overhead and inference latency. Stage-wise analysis shows repeated error correction by projection and lower intermediate reconstruction errors than the backbones alone throughout support expansion. Ablations favor dyadic allocation over uniform expansion for square and line-band supports and over manual front-loading for square windows. Support projection alone improves reconstruction, and transition conditioning strengthens the gain. Cross-dataset evaluation shows improvements for both backbones in both directions. Support allocation thus complements regularizer design in unrolled MRI reconstruction.
Diffusion large language models (dLLMs) have recently emerged as a strong alternative to autoregressive models, generating responses by iteratively refining masked tokens in parallel rather than decoding sequentially. While enabling more efficient and flexible inference, dLLMs remain vulnerable to jailbreak attacks, including recent attacks that exploit properties unique to dLLMs such as bidirectional conditioning and parallel masked-token refinement. In this work, we show that these same properties also expose a safety-relevant detection surface absent from standard autoregressive decoding: even when a jailbreak is difficult to identify from prompt representations alone, it can still be detected from hidden states over masked response positions before any tokens are generated. By training linear classifiers on prompt and pre-decoding response representations, we empirically show that neither view strictly dominates, with each detecting attacks that the other misses. We introduce $\texttt{ReFuse}$ (Representation Fusion), an inference-time detector that fuses the sigmoid harmfulness scores of the two classifiers. Without modifying model weights or the decoding procedure, $\texttt{ReFuse}$ reduces average ASR across multiple jailbreaks from 63.29% for the undefended model and 11.27% for a prompt-only classifier to 3.31%, while keeping average benign refusal on standard utility benchmarks below 1% on LLaDA-8B. Together, these results establish $\texttt{ReFuse}$ as a lightweight jailbreak detector for dLLMs that improves safety beyond prompt-only detection.
Surface-mounted vibration sensing provides non-invasive structural monitoring with high temporal fidelity across diverse ubiquitous computing scenarios such as smart farms, yet structure-borne wave models remain fragile to physical relocation and structural shifts. However, addressing this spatial data discontinuity remains limited by costly manual recalibration, insufficient physical grounding in data-driven synthesis, and substantial sim-to-real gaps in numerical digital twins. To resolve this spatial bottleneck without labor-intensive data recollection, we propose VeV, a physics-consistent spatial augmentation framework that synthesizes vibration responses at unmeasured target coordinates from sparse reference observations by coupling a monocular Vision-Derived Propagation Possibility Map for structural wave-guiding paths with a Visual Feature Map for transient source kinetics and implicit material dynamics. Evaluated on a 1,237 sensor-hour sow pig and piglet farm across disparate structures, VeV reduces waveform reconstruction MSE by over 57.5% compared to data-driven baselines and achieves up to 81.69% zero-shot classification accuracy, while reaching peak performance within just 30-50 calibration chunks to effectively eliminate cold-start retraining overhead and establish a scalable foundation extending to human-centric floor vibration sensing and ambient intelligence.
Deepfake videos generated by advanced generative models pose serious threats to digital security, misinformation control, and media forensics. Although numerous detection methods have been proposed, their reported performances are often based on heterogeneous preprocessing pipelines, inconsistent data splits, and model-specific training protocols, hindering fair architectural comparison. This work presents a controlled and reproducible comparative study of convolutional neural networks (ResNet-18 and EfficientNet-B0) and a transformer-based architecture (ViT-B/16) within a unified, face-centric deepfake detection framework. All models are trained using identical preprocessing, augmentation, optimization settings, and data splits on Celeb-DF v2 and FaceForensics++, and results are validated across three independent random seeds with bootstrapped confidence intervals. Experimental results show that while CNN-based models perform competitively at the frame level, their performance degrades after video-level aggregation. In contrast, ViT-B/16 demonstrates more stable video-level performance, achieving 92.8\% [95\% CI: 91.7--94.0] video-level accuracy and an AUC of 0.950 [95\% CI: 0.934--0.964] on Celeb-DF v2 (3-seed bootstrapped). Cross-dataset evaluation across the same three random seeds reveals a large, asymmetric generalization gap (AUC 0.79 FF++$\rightarrow$Celeb-DF vs. 0.64 Celeb-DF$\rightarrow$FF++), indicating that in-domain video-level stability does not transfer to unseen manipulation domains. These findings indicate that global self-attention improves robustness and video-level stability for deepfake detection within a given data domain, while highlighting the need for further work on cross-domain generalization.
Low-Rank Adaptation (LoRA) is a popular parameter-efficient fine-tuning (PEFT) method for fine-tuning of large language models (LLMs), especially in federated learning, due to its strong performance and communication efficiency. In practice, LoRA is applied to the query and value projections of transformer layers without considering the distinct roles of these components. In this work, we analyze the impact of these projections in federated fine-tuning of LLMs using LoRA and observe that the value projection converges significantly faster than the query projection. Motivated by this finding, we propose FedLoVA, a federated fine-tuning framework that trains LoRA adapters on both query and value projections locally while sharing only the value projection with the server. This approach reduces communication overhead while improving convergence efficiency. Experiments on natural language understanding, multilingual sentiment analysis, and generation tasks show that FedLoVA reduces the communication cost by up to $7.8\times$ compared to existing federated LLM fine-tuning baselines while achieving comparable or better performance. The code is available at: https://github.com/EnsiehKhazaei/FedLoVA.
High-fidelity dynamic simulation is essential for scientific and engineering applications, but its computational cost limits large-scale scenario screening. We investigate the use of pretrained time-series foundation models (TSFMs) as computationally efficient surrogates to address this limitation. In-context learning and post-training strategies, including retrieval-augmented reference sample selection and physics-informed regularization, are considered in the investigation. Results are benchmarked against a traditional AI model representative of the state-of-the-art in the domain. Experiments using an open-source TSFM and data from a simulated 9-bus power system indicate promising results. Retrieval augmentation improves in-context learning performance, and fine-tuning leads to superior results compared to the traditional AI model, especially when including physics-informed regularization. The fine-tuned TSFMs achieve lower errors and improved computational costs compared to the traditional AI model, and a 563$\times$ speedup compared to the original physics-based simulation.
Computer vision (CV) for policing has centred on recognising objects such as seized weapons, while body-worn video (BWV) is still reviewed by people, with existing AI tools using only its audio. Multimodal video AI could make BWV review faster and cheaper. At the object level, we present the KnifeHunter dataset, developed with UK law enforcement, with over 25,000 images of 543 knife classes and a protocol reflecting investigative conditions, and KHNet, a retrieval architecture fusing local and global features, which with about 125,000 distractor images attains 85.7% and 63.4% mean Average Precision (mAP) on the medium and hard protocols. At the event level, we present an evaluation protocol for video understanding pipelines, which describe a video as events and a transcript, that scores transcription, speaker attribution, event detection, temporal localisation and unsupported events. Whether each event is supported by the recording is labelled by a judge validated against blinded human labels. We apply it to complete pipelines, three 2026 omni-modal models and our hybrid NarrativeBridge-PAL-GPT-5 pipeline, on UKBWV-27, a pilot stress-test set of 27 publicly released UK BWV recordings (118 minutes). On this set the hybrid pipeline leads on five of six measures, yet it leaves 13.6% of predicted events unsupported and misattributes 27.7% of utterances, and every pipeline degrades on recordings from edited releases, long recordings and noisy audio. We propose readiness levels separating assistive tasks from those requiring verification of every output.
Feed-forward 3D reconstruction turns an uncalibrated monocular stream into a dense point cloud in a single inference pass, orders of magnitude faster than classical structure-from-motion (e.g., COLMAP), and reliable on the weakly textured, repetitive surfaces where keypoint matching fails outright. However, raw point clouds lack the object identity, physical properties, and watertight volume required for robot learning. This gap widens under robot-native capture. A camera mounted on a robot constrained to a few degrees of freedom sees each object only partially, and the repetitive, weakly textured structure of an office makes distinct workstations nearly indistinguishable, leaving redundant and partial observations that downstream reconstruction cannot resolve. Existing image-to-3D generation methods struggle with these conditions, hallucinating geometry under incomplete views and confusing repeated instances. We present OfficeTwin, a robot-native framework that converts these incomplete observations into simulation-ready digital twins. OfficeTwin first imposes metric and structural constraints, then exploits the office's structural redundancy by registering repeated cubicles into a shared canonical frame, where complementary observations recover geometry occluded in any single view without hallucination. On a real office scan, OfficeTwin recovers every cubicle as an individually watertight, metric solid, outperforming feed-forward reconstruction and generative baselines that fail to produce simulation-ready assets.
Long-horizon LLM agents accumulate sessions of up to millions of tokens, far more key-value (KV) cache than one GPU holds, so serving systems keep a small budget of tokens and evict the rest. Eviction methods differ in which tokens they keep but share one goal: that the tokens a later step needs are still cached. We show this is not enough. In 100K-token coding sessions on Qwen3-8B with a 2,048-token budget, a planner-guided policy that keeps over 99% of the needed tokens resident still fails late steps increasingly often: its accuracy falls from 93% to 64% as the gap between a definition and its use grows from 2K to 90K tokens, at residency 0.996. The needed tokens are resident but unretrievable: what grows is their rotary-position distance to the query. We present PACE (Planner-Aware Context Engine), a KV-cache engine that couples plan-aware retention, in which the agent's planner tells the cache which context future steps will read, with bounded positions, which renumber kept tokens so their rotary distance to the query stays below 8K at any session length. In a pre-registered evaluation on fresh tasks, PACE passes 97% of hidden unit tests at 100K tokens, against 42% for H2O and 3% for SnapKV under the same budget, and needs both components: either alone reaches 63% (the combination beats each, p < 10^-4). The result holds with length: in 1M- and 2M-token sessions on one 24 GB GPU, with definitions written 0.9M and 1.8M tokens before use, PACE passes 99% and 100% of checks and recalls every session fact, while bounded positions without foresight falls from 33% to 13%; PACE holds under 0.7 GB of KV state where the full cache would need 147-295 GB. Renumbering is exact, costs no measurable perplexity, and adds 9% wall time. On Mistral-7B, whose coding ability caps it at 40% even with nothing evicted, PACE stays at that ceiling while retention alone falls below it.
Context selection can reduce input length during long-input inference but can remove evidence required to answer a query. We introduce CoRA-Bench, a benchmark for evaluating evidence preservation under shared token budgets. We compare head and tail truncation, BM25, dense retrieval, dense retrieval with reranking, LLMLingua-2, and Provence across BABILong, LOONG Financial Spotlight, a custom NoLiMa diagnostic, and LongMemEval. Evidence-preservation results reveal task-dependent tradeoffs: BM25 achieves 0.921 supporting-fact recall on BABILong at 128K, dense retrieval and dense reranking each achieve 0.908 needle recall on NoLiMa at 64K, and dense retrieval achieves 0.993 answer-bearing-turn recall on LongMemEval at 64K. Evidence, token retention, and complementary reader evaluations use the same selected contexts. Reader evaluations use dataset-native answer metrics and three runs on the same questions, with paired bootstrap confidence intervals. At a 32K budget, BM25 achieves 46.20% BABILong answer accuracy and dense retrieval achieves 71.13% LongMemEval accuracy. The results show that method rankings vary across evidence types and between evidence recall and downstream accuracy.
Linear State Space Models (SSMs) and linear autoregressive sequence models reuse the same parameters across a sample-dependent number of steps. As a result, different examples impose constraints through different predictors, even though all predictors share the same weights. We study this phenomenon through a general model of multiple non-homogeneous polynomial predictors with shared parameters, in a classification setting. Under standard interpolation and directional-convergence assumptions, together with a trajectory-dependent effective-degree condition, we show that gradient descent on the exponential loss converges in direction to a KKT point of an effective-degree max-margin problem. In this problem, only examples with minimal effective degree impose hard margin constraints; examples with faster-growing margins are asymptotically inactive except for feasibility. Our result provides a first-order characterization of the implicit bias in linear SSMs and linear autoregressive models. It exposes a mechanism of length bias: a longer input sequence or additional autoregressive steps may increase the example's effective degree, and the examples with the smallest effective degree dominate the limiting classifier. Our result highlights how non-homogeneity and parameter sharing alter classical homogeneous implicit-bias results.