Computer Science & Mathematics

Data distributions often possess one or more symmetries: for example, all orientations of any specific dog are included in the distribution of all possible dog pictures. Despite the omnipresence of such symmetries, modern generative modeling approaches may fail to incorporate or learn them, which can lead to lower sample efficiency and poorer generalization. To what extent do diffusion models fail to learn data distribution symmetries, and what sources of symmetry breaking are most responsible? Here, we answer this question in the theoretically tractable setting of linear diffusion models. We show that of three potential sources of symmetry breaking---a non-symmetric initial weight matrix, a noisy learning target, and finite-sample noise---target noise produces the most robust symmetry breaking, since it persists even in the sample-rich regime. Using local laws from random matrix theory, we derive formulas for average symmetry error in fairly general cases involving symmetry within a subspace and show that smaller symmetries are harder to learn.

Clinical trial prescreening requires distinguishing unmet conditions from insufficient evidence in patient records. Given patient records, criterion questions, and predefined answer options, we retrospectively compared JEV, a non-generative decision model returning native option probabilities, with schema-constrained Qwen Flash across 2,048 test patient-condition pairs from TrialGPT and question-answer datasets derived from MIMIC-III and MIMIC-IV, all evaluated under strict Zero Data Retention (ZDR) policies to ensure clinical data privacy. Both models assessed conditions individually or in batches sharing the same patient context, with performance measured against reference labels. In batch mode, JEV and Qwen achieved similar macro-F1 scores on TrialGPT (0.652 vs. 0.660) and MIMIC-IV (0.807 vs. 0.814), whereas JEV substantially outperformed Qwen on MIMIC-III (0.768 vs. 0.534; $Δ$ = +0.234, 95% CI: 0.187-0.281). This advantage largely reflected higher recall for conditions labeled as having insufficient information (56.3% vs. 9.9%). Across all three datasets, JEV also had lower observed median job latency (1.18-1.40 vs. 2.80-8.90 s) and 53%-74% lower estimated cost per condition. These findings position JEV as an initial assessment component that provides criterion-level judgments and flags potential information gaps for review. JEV's observed errors in interpreting clinical evidence underscore the need to verify its outputs against patient records before they inform eligibility decisions.

High-resolution video matting requires preserving fine details, such as hair strands and soft transparency, while maintaining consistency across frames. Memory-based methods achieve strong matting quality by matching current-frame features with information stored from earlier frames, but this process becomes increasingly costly as resolution grows. We present MatAnyone-E, an efficient extension of MatAnyone 2 that focuses computation where it is most needed. Our approach first performs low-resolution propagation to identify uncertain regions that require closer inspection and the memory candidates most relevant to them. It then performs high-resolution matching only for these regions using the selected memory, while reusing low-resolution matching results elsewhere. This strategy reduces both the number of locations requiring detailed matching and the amount of memory searched, while preserving the information needed to produce a complete matte. We further reduce the cost of generating high-resolution mattes with a lightweight decoder designed around measured runtime bottlenecks and trained to reproduce the original decoder's representations. Experiments show that MatAnyone-E maintains matting quality comparable to MatAnyone 2 while achieving an estimated 7.4x inference speedup and reducing peak GPU memory usage to 7.6% of the usage of MatAnyone 2 at 4K. With compilation-enabled inference on a single NVIDIA H200 GPU, it reaches 73.07 FPS at 1080p and 28.62 FPS at 4K.

The gap between AI capability deployment and safety assurance is structural and cannot be closed through internal evaluations alone. Analyzing documented harms (2017-2026) across the human lifespan reveals that disparate incidents --- from voice assistants endangering children to AI companions implicated in adolescent suicide and automated care denials for the elderly --- reflect a single failure: evaluating systems against an implicit adult user model. We introduce five developmental moderators structured into a Lifespan Vulnerability Matrix across six life stages. The matrix reveals a nonmonotonic risk distribution: vulnerability peaks at both ends of the lifespan, governance coverage is asymmetrical, and authority to convert model errors into real-world impact is scrutinized least. We then propose a set of lifespan design principles generalizing existing child-protective instruments, and an institution-agnostic, three-tier external evaluation framework --- comprising automated simulation, expert red teaming, and post-deployment monitoring --- indexed directly to matrix risk cells.

Web-connected sensing applications continuously generate multivariate time series with evolving behavioral categories, making Time Series Class-Incremental Learning (TSCIL) essential for adapting to new classes without forgetting previously acquired knowledge. However, preserving diverse and evolving class distributions under strict privacy and memory constraints remains a fundamental challenge. Existing approaches often face a trade-off between retaining sufficient historical knowledge for accurate recognition and minimizing the memory required for continual adaptation. To address these limitations, we propose TRACE, a memory-efficient, replay-free framework that preserves class-specific knowledge through generative modeling rather than sample rehearsal. TRACE employs a compact, frozen temporal-spectral representation and lightweight class-specific autoencoders to capture complex intra-class variations. By independently preserving each class model, TRACE avoids recursive generation and catastrophic forgetting, while enabling direct classification through reconstruction-based scoring without retaining historical samples or large statistical matrices. Extensive experiments across five TSCIL benchmarks demonstrate that TRACE achieves competitive classification accuracy while reducing mean retained memory by approximately 98.26% compared with TS-ACL, the strongest evaluated analytical baseline, and achieving an approximately 33.43% relative improvement in accuracy on GrabMyo.

Imagery captured by unmanned aerial vehicle (UAV) systems increasingly raises privacy concerns, as UAV aerial surveillance routinely captures sensitive objects such as pedestrians and vehicles alongside ordinary scene content. Existing privacy-preserving approaches typically employ class-agnostic saliency detection to indiscriminately obfuscate all visually salient regions, leading to unnecessary information loss and hindering downstream scene analysis. This paper proposes a category-selective encryption framework for privacy-preserving UAV surveillance. The core advance over prior class-agnostic approaches is semantic-level selectivity: the system first recognizes what is present in the scene via an 80-class detector, then encrypts only user-specified sensitive categories while leaving non-sensitive objects and background intact for downstream analytics. The framework supports two complementary encryption schemes: Thumbnail-Preserving Encryption (TPE), which applies adaptive pixelation with a learnable blend parameter and recovers content through a jointly trained Enhanced U-Net, and XOR stream cipher encryption, which offers parameter-free, mathematically exact reversibility. The architectural design, composite loss formulation, and joint training methodology are presented in detail, establishing a foundation for practical privacy-preserving UAV surveillance that balances protection strength and scene utility.

Simulation benchmarks are a useful way to measure progress in robotics; typically they provide success rate as the dominant measure of performance. However, success rate reports only whether a rollout completed the task, hiding by what margin it succeeded, how sensitive the success is to goal parameter changes, requirements it may have violated, and whether success persisted. Signal Temporal Logic (STL) provides a formal language for capturing specifications; it can be used to define requirements, and measure how close a robot behavior is to satisfying or violating them. Using STL monitors in a simulator, however, requires exposing simulator state and mapping it to specification-relevant signals. We present Specola, a tool for evaluating robot behaviors in simulation with respect to STL specifications. Specola provides an interface for writing specifications over simulator signals, and creates and manages the relevant monitors while a policy executes. We provide a Python API and command-line interface for integrating monitoring into existing policy execution loops. We provide an adapter interface for connecting new simulators, and provide adapters for nine simulation environments and benchmarks, including LIBERO, RoboCasa, Isaac Lab, and RoboDojo. We provide case studies that illustrate the nuanced evaluation that Specola enables. Code and adapters are available at https://github.com/specola-eval/specola.

Multi-objective reinforcement learning is a fundamental problem for modern AI, with applications spanning post-training, reasoning models, and robotic control. A central challenge is to learn policies that balance competing objectives, such as helpfulness versus honesty in safety and alignment, or speed versus energy in robotics. The optimal policy can vary sharply with objective-weighting combinations, and the extent of this variation depends crucially on the degree of conflict among objectives, making per-preference policy learning sample-inefficient. In this paper, we introduce objective-conflict coefficients and derive near-optimal regret bounds for online settings using this measure, thereby quantifying the impact of misalignment on learning efficiency. Then, we design routed ensemble policy learning, an algorithm that maintains a small set of basis policies and dynamically routes trajectories to the policy with the highest potential reward. Finally, we consider trajectory selection as another approach to mitigate objective conflict, and design a kernel surrogate trajectory attribution algorithm that estimates how feedback signals from a set of trajectories influence policy learning on other trajectories. The kernel can capture nonlinear interactions such as synergy and antagonism between trajectories. We evaluate our approach across diverse environments and benchmarks: routed ensembles improve average performance by 4.3% over baselines for online environments from Meta-World and MO-Gymnasium, and kernel surrogate attribution improves correlation with leave-one-out analysis over existing baselines by 50%. When applied to trajectory selection, kernel-based attribution yields a 27% improvement in reward over baseline attribution methods.

Multi-objective multi-agent reinforcement learning (MOMARL) addresses agent coordination and objective trade-offs simultaneously while optimising overall system performance. Existing work tends to treat multi-agent cooperation and multi-objective optimisation as largely independent components of a learning problem, which amounts to an exorbitant complexity except in trivial cases. We study here whether the structure provided by multiple objectives can itself be exploited to facilitate cooperation, rather than being treated as an additional optimisation challenge. We formulate MOMARL at both the team and individual levels and introduce objective-wise counterfactual credit assignment to evaluate each agent's contribution across the team objectives. We propose Meta-Gradient Informed Multi-Objective Multi-Agent Reinforcement Learning (MIRL), with static (MIRL-S) and dynamic (MIRL-D) variants. By incorporating meta-gradient information, MIRL enables agents to learn how to prioritise objectives individually while collectively optimising the system-level objective, allowing the multi-objective structure to directly guide cooperation. Experiments on a multi-objective multiple particle environment show that our methods improve Pareto-front quality over baselines, with the dynamic variant achieving the strongest performance.

Demand on blockchain-based marketplaces arrives in bursts and at a high rate. On these platforms the data are completely visible, which makes it feasible to compare the bursty demand against the usual load. The congestion manifests as an increased demand for gas, the primary resource a blockchain sells. Existing studies consider how to price the gas. The demand itself, and what it is made of during bursts, has not been measured. We measure it on Robinhood Chain, a layer-2 blockchain that quickly became a venue for trading tokenized equities and memecoins. We collected the data from every block of its first three active months and more than 50 million Uniswap swaps that occurred during bursty periods. We find that a burst trades about as many pools as a quiet period, mostly different ones, and trades them harder, on top of elevated demand that accumulated for hours. With the bursts aligned on their peaks, the median peak is 2.5 times the usual load. Both the rise and the fall of the demand are described by the sum of two exponential functions with different timescales, a slow one of hours and a fast one of minutes. Against a matched quiet period, a burst has 1.05 times the pools, 1.40 times the trades, and 1.21 times the senders. We use two models to explain the arrival of bursts: externally triggered cascades and self-excitation. The former fits the data relatively better. Inside a burst, the entry of each trader into a pool is followed for about 24 seconds by more entries than the existing traders predict. Whether this is because of imitation or because of a shared signal is left open.

Computing local-level variable importance (VI) is critical for understanding machine learning models. Local permutation-based VI metrics offer a simple yet powerful approach, measuring how much a model's output changes when perturbing its inputs. However, existing perturbation strategies often ignore feature dependence or account for it only through restricted approximations. Perturbing a feature without fully respecting this dependence can generate unrealistic counterfactuals, thereby evaluating the model in regions unsupported by the observed data distribution. We address this gap by introducing a new model-agnostic local VI metric and estimator, Local Understanding of Covariate Importance with conditional Density estimation (LUCID). LUCID perturbs features of interest using plausible values drawn from the full conditional distribution given the remaining observed features. We establish theoretical connections between LUCID and existing VI measures, and provide an estimator with convergence guarantees and uncertainty quantification. Finally, we demonstrate LUCID's stable performance and practical utility across simulated and real-world examples.

Wide-area 6DoF rendering from sparse Ambisonic recordings is commonly addressed by interpolating parametric spatial descriptors rather than the waveforms themselves. This preserves the fidelity of a nearby measured signal and avoids phase artifacts caused by mixing distant arrays. However, existing DirAC-based approaches rely on geometric interpolation weights that do not depend on the acoustic scene observed by the microphones. We propose a lightweight Graph Neural Network (GNN) that refines only this interpolation stage. The known Ambisonic arrays and the query position are represented as a graph whose nodes contain local DirAC features and whose edges encode relative distance and directional relations. The network predicts positive query-conditioned weights over the measured arrays, while the target energy, diffuseness, and direction of arrival are reconstructed by an analytical DirAC layer. Experiments on unseen rooms, unseen speech signals, and unseen array configurations show that the proposed weighting improves energy reconstruction and radiation-pattern similarity with respect to baseline interpolation.

Existing jailbreak attackers fall into two classes: static planners fix their single- or multi-turn adversarial plans in advance, remaining simple but losing sight of the victim's replies, wasting turns and limiting transferability; interactive generators read each victim reply to generate the next-turn attack but require in-loop evaluators for success signals or hints, incurring substantial costs and inflated performance. Bridging this gap, we propose SEMA-Auto, a sighted and efficient multi-turn attacker with autonomy, interacting with the victim model and terminating the session by itself upon success. To stabilize its behaviors and boost its capabilities, SEMA-Auto is trained in two stages. Synthetic bootstrap tuning installs the notion of attack targets and conforming decisions by fine-tuning the base model with bootstrapped trajectories. Decision-supervised reinforcement learning then optimizes two objectives jointly yet separately, reinforcing the attack with higher jailbreak reward while supervising the decision at each turn against an absolute target. Our method achieves state-of-the-art (SOTA) attack success rates (ASR) across multiple datasets, victims, and judges, outperforming all static and interactive baselines, within almost the fewest turns, demonstrating both effectiveness and efficiency. For example, SEMA-Auto performs a $76.0\%$ ASR@1 within an average of $3.0$ turns against DeepSeek-V4-Flash on AdvBench, while the best static baseline is $54.8\%$ with fixed $7$ turns and the best interactive baseline is $43.8\%$ in $3.7$ turns. Against safety-hardened frontier models, GPT-oss-120B, GPT-5.4, Claude-Haiku-4.5, and GPT-6-Astra, our counter-trained attacker achieves an average $59.6\%$ ASR@1 on HarmBench, $13.8\%$ over prior SOTA. Our code is available at: https://github.com/microsoft/SEMA

Clinical audio models increasingly align recordings with participant metadata or metadata-derived text. This supervision is useful but ambiguous: the paired target can contain symptoms, demographics, missingness, and cohort structure, not just disease information. We study this ambiguity by making the audio-metadata pairing itself the intervention. Our Pairing-Controlled Transfer Audit trains matched contrastive trajectories under correct pairing, global shuffling, disease-label-preserving shuffling, and label-and-recorded-sex-preserving shuffling. Each endpoint is then frozen and tested with three readouts: profile correspondence, attribute decodability, and matched disease prediction. On UKCOVID, correct improves retrieval and sex decodability over label-only shuffling, showing that the pairing intervention worked. The sex-probe contrast is largely absorbed by the sex-preserving control, while smaller retrieval differences remain. Disease-transfer contrasts, however, remain uncertain. After UKCOVID discovery fixed the pipeline, CODA TB, Cambridge COVID-19 Sounds, and Coswara were rerun as a locked external panel with the same arms, backbones, matching, readout, calibration, primary contrast, and primary metric. The external panel preserves the descriptive separation without proving a universal null. Successful profile correspondence and improved disease transfer are therefore different empirical claims.

Model selection and hyperparameter tuning pose a central challenge in off-policy evaluation (OPE) because errors of fitted Q-functions and occupancy-ratio estimates cannot be directly evaluated from held-out transitions. Bellman-residual and minimax validation methods often require auxiliary regressions or optimization over separate critic classes, adding computational and tuning costs. We introduce pairwise Bellman validation (PBV), a tuning-free minimax procedure for selecting among finite classes of fitted Q-functions and occupancy-ratio estimators using held-out transitions. PBV constructs critics from pairwise candidate differences, requiring only finitely many empirical moment calculations without auxiliary validation models or separate critic optimization. We establish finite-sample oracle inequalities relative to the best available candidate, even under candidate-class misspecification. Our key insight is that Q-function errors and log ratios between candidate and true occupancy ratios provide oracle critics whose Bellman and adjoint Bellman moments control estimation error. Pairwise candidate differences approximate these unknown critics, while penalties account for comparator approximation error. For Q-function selection, we first develop unweighted PBV, which evaluates pairwise Bellman moments under the offline distribution and admits oracle guarantees under a pairwise coercivity condition relating the offline distribution to the target-policy dynamics. We then develop target-aligned PBV (TA-PBV), which uses marginalized importance weighting to provide error control without pairwise coercivity. For occupancy-ratio selection, we develop adjoint PBV (A-PBV), which evaluates adjoint Bellman moments against log-ratio critics. On Hopper and Deep OPE, unweighted PBV performs competitively, while target-occupancy weighting can improve performance in settings with controlled overlap.