Generative machine learning (ML) models hold great promise for accelerating materials discovery through the inverse design of inorganic crystals, enabling an unprecedented exploration of chemical space. Yet, the lack of standardized evaluation frameworks makes it challenging to evaluate, compare, and further develop these ML models meaningfully. In this work, we introduce LeMat-GenBench, a unified benchmark for generative models of crystalline materials, supported by a set of evaluation metrics designed to better inform model development and downstream applications. We release both an open-source evaluation suite and a public leaderboard on Hugging Face, and benchmark 12 recent generative models. Results reveal that an increase in stability is accompanied by lower novelty and diversity on average, with no model excelling across all dimensions. Altogether, LeMat-GenBench establishes a reproducible and extensible foundation for fair model comparison and aims to guide the development of more reliable, discovery-oriented generative models for crystalline materials.
Optimizing numerical systems and mechanism design is crucial for enhancing player experience in Massively Multiplayer Online (MMO) games. Traditional optimization approaches rely on large-scale online experiments or parameter tuning over predefined statistical models, which are costly, time-consuming, and may disrupt player experience. Although simplified offline simulation systems are often adopted as alternatives, their limited fidelity prevents agents from accurately mimicking real player reasoning and reactions to interventions. To address these limitations, we propose a generative agent-based MMO simulation system empowered by Large Language Models (LLMs). By applying Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) on large-scale real player behavioral data, we adapt LLMs from general priors to game-specific domains, enabling realistic and interpretable player decision-making. In parallel, a data-driven environment model trained on real gameplay logs reconstructs dynamic in-game systems. Experiments demonstrate strong consistency with real-world player behaviors and plausible causal responses under interventions, providing a reliable, interpretable, and cost-efficient framework for data-driven numerical design optimization.
Machine learning interatomic potentials (MLIPs) are commonly trained and evaluated by their ability to reproduce first-principles energies and forces. Practical value, however, is often determined by downstream processes such as structural relaxation. Much of MLIP literature is motivated by the expectation that improving force accuracy will produce correspondingly better downstream behavior. Although this expectation is natural, it may not hold in practice due to model inaccuracies. Here, we demonstrate that force accuracy and relaxation accuracy can be driven in opposite directions. Starting from pretrained MLIPs, we differentiate through relaxation trajectories and optimize the the final structures. This intervention produces increases in force-prediction errors while substantially reducing structural errors. Importantly, this represents a trade-off between two aspects of physical fidelity: the learned forces become less faithful to first-principles forces, while the resulting relaxed structures become more faithful to first-principles predictions. Rather than observing a discrepancy between local quality and downstream performance, we demonstrate that this discrepancy can be induced through optimization. Our findings suggest that MLIPs need not be optimized as approximations of first-principles forces, but can instead learn task-specific vector fields that guide relaxation trajectories toward more accurate final structures. By propagating supervision through complete relaxation trajectories, our framework captures dependencies between successive predictions that are absent when structure--force pairs are treated independently. More broadly, these results demonstrate that conventional force metrics can provide an incomplete assessment of downstream performance and highlights the importance of evaluating scientific machine-learning models within computational workflows.
A common recipe to improve diffusion models at test-time so that samples score highly against a user-specified reward is to introduce the gradient of the reward into the dynamics of the diffusion itself. This procedure is often ill posed, as user-specified rewards are usually only well defined on the data distribution at the end of generation. While common workarounds to this problem are to use a denoiser to estimate what a sample would have been at the end of generation, we propose a simple solution to this problem by working directly with a flow map. By exploiting a relationship between the flow map and velocity field governing the instantaneous transport, we construct an algorithm, Flow Map Trajectory Tilting (FMTT), which provably performs better ascent on the reward than standard test-time methods involving the gradient of the reward. The approach can be used to either perform exact sampling via importance weighting or principled search that identifies local maximizers of the reward-tilted distribution. We demonstrate the efficacy of our approach against other look-ahead techniques, and show how the flow map enables engagement with complicated reward functions that make possible new forms of image editing, e.g. by interfacing with vision language models.
Monocular 3D face reconstruction estimates the parameters of a 3D morphable face model (3DMM) from a single image, providing an explicit geometric representation for facial expression analysis. Owing to the scarcity of paired 2D images and 3D facial meshes, most methods are trained in a self-supervised manner by rendering the predicted face and minimizing image-level reconstruction losses, such as photometric and landmark errors, against the input image. However, these losses enforce visual alignment but do not ensure that the predicted parameters preserve affective information. Recent methods add image-level emotion consistency loss, but it can be satisfied by exaggerating facial deformations. To address these issues, we propose FIELDS (Face reconstruction with accurate Inference of Expression using Learning with Direct Supervision), which learns affect-informative expression representations in the FLAME parameter space. Specifically, FIELDS augments 2D self-supervised reconstruction with scan-derived expression anchoring and a parameter-space emotion loss. During training, scan-fitted targets directly constrain expression-related FLAME parameters to discourage exaggerated deformations, while a jointly trained emotion head uses discrete emotion and continuous valence--arousal annotations to make the predicted parameters informative for affect analysis. Experiments show that FIELDS improves affective utility from the predicted parameters in both within- and cross-dataset settings. Notably, it also achieves lower reconstruction errors overall and across most facial regions than the comparison baselines under both rigid and similarity alignment.
Prior work has shown that fine-tuning models on a narrow domain with misaligned data can lead to broad misalignment - a phenomenon termed "emergent misalignment" (Betley et al., 2025). While all tested models were susceptible to emergent misalignment, some models showed more resistance than others. Specifically the Qwen-2.5 family proved to be relatively resistant, while GPT-4o exhibited the strongest misalignment. In this paper we evaluate if current-generation open-weights models exhibit similar resistance to the Qwen-2.5 family and measure misalignment robustness over a range of model architectures and scales. We replicate the effect across nine modern open-weights models (Gemma 3 and Qwen 3 families, 1B-32B parameters). Models fine-tuned on insecure code generation show a 0.74% misalignment rate (compared to 0.07% for base models), matching the lower end of prior open-model results but dramatically lower than GPT-4o's 20%. We identify a critical format-dependent vulnerability: requiring JSON output doubles misalignment rates compared to natural language prompts (0.96% vs 0.42%), an effect concentrated in the Gemma 3 family (1.43% vs 0.22%) and not significant in Qwen 3 (0.72% vs 0.59%). This suggests that structural constraints may bypass safety training by reducing the model's 'degrees of freedom' to refuse. These findings confirm emergent misalignment as a reproducible phenomenon in modern open-weights models, with rates substantially lower than observed in proprietary systems.
Infrared small target detection remains challenging due to limited feature representation and severe background interference, resulting in sub-optimal performance. While recent CLIP-inspired methods attempt to leverage textual guidance for detection, they are hindered by inaccurate text descriptions and reliance on manual annotations. To overcome these limitations, we propose DGSPNet, an end-to-end language prompt-driven framework. Our approach integrates dual-granularity semantic prompts: coarse-grained textual priors (e.g., 'infrared image', 'small target') and fine-grained personalized semantic descriptions derived through visual-to-textual mapping within the image space. This design not only facilitates learning fine-grained semantic information but also can inherently leverage language prompts during inference without relying on any annotation requirements. By fully leveraging the precision and conciseness of text descriptions, we further introduce a text-guide channel attention (TGCA) mechanism and text-guide spatial attention (TGSA) mechanism that enhances the model's sensitivity to potential targets across both low- and high-level feature spaces. Extensive experiments demonstrate that our method significantly improves detection accuracy and achieves state-of-the-art performance on three benchmark datasets.
Few-Shot Semantic Segmentation (FSS) models achieve strong performance in segmenting novel classes with minimal labeled examples, yet their decision-making processes remain largely opaque. Explainable AI has matured on standard computer vision tasks, but not on FSS, where an explanation would tell a practitioner which support pixels the prediction rested on and which examples to provide next. We argue that matching-based FSS models are interpretable by design: their query-support similarity scores, once normalized and processed, attribute the prediction to individual support pixels. Our approach, Affinity Explainer (AffEx), makes this intrinsic interpretability explicit by extracting attribution maps directly from matching scores at multiple feature levels, without requiring gradients or external perturbations. We extend the standard causal metrics for evaluating explanations to the FSS domain and propose additional metrics to better capture the practical utility of explanations in few-shot scenarios. Experiments on two FSS benchmarks, covering seven architectures that range from classical matching-based models to recent pipelines built on vision foundation models (DINOv2/v3, SAM, SAM2) and a prototype-based head, show that AffEx outperforms adapted standard attribution methods at nearly every evaluation point, and that the gain comes from the matching mechanism itself. On pipelines whose segmentation stage transforms the matching output, the same evaluation doubles as a diagnostic of where the decision is made, and on the prototype-based head AffEx subsumes the prototype-style read-out as a special case. Qualitatively, the attribution maps expose how each architecture resolves the support set, and we show how they guide model diagnosis. Code is available at https://github.com/pasqualedem/AffinityExplainer.
Background: Antimicrobial decision support requires both clinical knowledge and patient-specific reasoning. We developed Knowledge and Reasoning Augmented Learning (KRAL) to improve language model performance on antimicrobial decision-making tasks. Methods: In this retrospective technical study, a 27-billion-parameter student model underwent guideline-grounded supervised fine-tuning followed by Multi-turn Agentic GRPO using clinical reasoning trajectories. Six configurations were evaluated on 123 clinician-validated clinical antimicrobial treatment cases from Peking Union Medical College Hospital (PUMCH) and 100 JAMA Clinical Challenge cases. A language model judge scored responses, and paired tests compared configurations. Component ablations assessed benchmark performance and training stability. Results: On the PUMCH benchmark, KRAL achieved the highest composite score (0.895 on a 0-1 scale), outperforming all five controls (all P<0.001). On JAMA, KRAL scored 0.600 and exceeded the base model with and without retrieval (P = 0.0048 and 0.0499, respectively). Reinforcement learning ablation experiments demonstrate that the reward smoothing can significantly improve the performance and training stability of vanilla GRPO. Conclusions: KRAL improved performance on a single-center clinical antimicrobial benchmark and publicly available JAMA clinical challenge cases by combining guideline-grounded learning with Multi-turn Agentic GRPO. Further external clinical validation is needed to evaluate the practical utility and effectiveness of KRAL.
We introduce GLOBE, a nonlinear neural surrogate for boundary-driven, homogeneous, weakly nonlinear PDEs that draws inductive bias from boundary-element methods and equivariant ML, representing solutions as superpositions of learnable Green's-function-like kernels evaluated from boundary faces to targets. The architecture is translation-, rotation-, and parity-equivariant; discretization-invariant; and units-invariant via rigorous nondimensionalization. On AirFRANS (steady incompressible RANS over NACA airfoils), a compact ~117k-parameter GLOBE model has lower mean-squared error than published values for the strongest reported ML surrogates on both the interpolation and low-data splits, in single-run comparisons under the released AirFRANS normalization. The model evaluates fields at arbitrary points, handles non-watertight meshes, and generalizes under Reynolds number and angle-of-attack distribution shifts. We further introduce a hierarchical Barnes-Hut acceleration that exploits the kernel's far-field decay to reduce evaluation cost from quadratic to near-linear in the number of boundary elements, with a tunable opening angle that recovers exact dense evaluation in the zero-angle limit. Quantitative validation focuses on 2D AirFRANS; on 3D DrivAerML cars, GLOBE predicts surface and volume fields on the full meshes, and we compare it with a matched Transolver baseline. Together, these results show that boundary-element-inspired inductive biases can substantially improve the accuracy of ML surrogates for boundary-driven PDEs while scaling to large 3D meshes.
Cameras are a core sensing modality in modern intelligent transportation systems (ITS), providing rich visual information on road-user activities. Multi-Camera Vehicle Tracking (MCVT) uses this data to reconstruct vehicle trajectories across camera networks, supporting applications such as traffic flow prediction and optimisation. However, most existing MCVT studies emphasise tracking accuracy while paying limited attention to real-time performance and scalability, both essential for real-world and city-scale deployment. To address this gap, we propose Edge-Assisted, Scalable and Efficient MCVT (EASE-MCVT), a distributed edge--server framework designed for real-time throughput and scalable operation. On the edge side, each camera stream is processed through object detection, single-camera tracking, geo-mapping and feature extraction, while only lightweight metadata, including vehicle locations and appearance features, is sent to the central server for cross-camera association. To improve both tracking accuracy and system efficiency, EASE-MCVT is optimised from algorithmic and system perspectives. Algorithmically, it introduces a dynamic workload scheme for tracklet-level feature extraction, a server-side re-match module to reconnect fragmented tracklets, and a self-supervised camera link model that learns spatio-temporal constraints to accelerate and stabilise cross-camera association. Systemically, it integrates production-oriented data engineering components to standardise deployment and data exchange for large-scale operation. To the best of our knowledge, EASE-MCVT is the first MCVT framework explicitly designed to address both real-time performance and scalability in a distributed edge--server setting. Experiments on the RoundaboutHD and CityFlow datasets demonstrate real-time throughput with competitive tracking accuracy, paving the way for city-wide real-time traffic management.
Pervasive sensor deployments have generated significant interest in IoT data analysis across fields such as healthcare, transportation and urban planning. At the same time, the rise of Large Language Models (LLMs) offer a new tool for exploring IoT data through their expressive, natural language interface. Unfortunately, natural language querying of IoT sensor streams is impractical today: raw telemetry exhausts LLM context windows in fractions of a second at scale, incurs prohibitive per-token costs, and produces hallucinated responses. Existing systems work around these limits with lossy text summarization or unconstrained code synthesis, producing expensive, unverified computations. Multi-step queries compound this problem: logical errors in intermediate operations can derail the final answer. We present Flash-Fusion, an agentic framework for natural-language querying of tabular IoT data. Flash-Fusion identifies which available data fields and supported derivations can answer the user's query, constructs a plan using typed, domain-specific operators, and validates its structure and schema compatibility before executing. Because constructing plans still requires costly LLM reasoning, Flash-Fusion caches validated operator sequences for reuse across compatible queries. A smaller LLM adapts a cached plan's parameters to each query, cutting latency and cost. We evaluate Flash-Fusion across human activity recognition, ECG monitoring, and vehicular telematics, covering direct lookups, multi-step reasoning, predictive tasks, and out-of-scope requests. Flash-Fusion achieves 92.2% query accuracy versus 76.1% for ReAct, the strongest baseline, while completing queries 7x faster and consuming 5x fewer tokens. Flash-Fusion enables users across disciplines to explore IoT data without manually constructing queries.
Cardiovascular disease (CVD) is a leading cause of mortality worldwide. Electrocardiograms (ECGs) are the most widely used non-invasive tool for cardiac assessment, yet large, well-annotated ECG corpora are scarce due to cost, privacy, and workflow constraints. Generating ECGs can be beneficial for the mechanistic understanding of cardiac electrical activity, enable the construction of large, heterogeneous, and unbiased datasets, and facilitate privacy-preserving data sharing. Generating realistic ECG signals from clinical context is important yet underexplored. Recent work has leveraged diffusion models for text-to-ECG generation, but two challenges remain: (i) existing methods often overlook the physiological simulator knowledge of cardiac activity; and (ii) they ignore broader, experience-based clinical knowledge grounded in real-world practice. To address these gaps, we propose SE-Diff, a novel physiological simulator and experience enhanced diffusion model for comprehensive ECG generation. SE-Diff integrates a lightweight ordinary differential equation (ODE)-based ECG simulator into the diffusion process via a beat decoder and simulator-consistent constraints, injecting mechanistic priors that promote physiologically plausible waveforms. In parallel, we design an LLM-powered experience retrieval-augmented strategy to inject clinical knowledge, providing more guidance for ECG generation. Extensive experiments on real-world ECG datasets demonstrate that SE-Diff improves both signal fidelity and text-ECG semantic alignment over baselines, proving its superiority for text-to-ECG generation. We further show that the simulator-based and experience-based knowledge also benefit downstream ECG classification.
As the complexity of Software Engineering (SE) tasks continues to escalate, Multi-Agent Systems (MASs) have emerged as a focal point of research and practice due to their autonomy and scalability. Furthermore, through leveraging the reasoning and planning capabilities of Large Language Models (LLMs), the application of LLM-based MASs in the field of SE is garnering increasing attention. However, there is no dedicated study that systematically explores the design of LLM-based MASs, including the Quality Attributes (QAs) on which designers mainly focus, the design patterns used by designers, and the rationale guiding the design of LLM-based MASs for SE tasks. To this end, we conducted a study to identify the QAs that LLM-based MASs for SE tasks focus on, the design patterns used in the MASs, and the design rationale for the MASs. We collected 106 papers on LLM-based MASs for SE tasks as the source. Our study shows that: (1) Code Generation is the most common SE task solved by LLM-based MASs among ten identified SE tasks, (2) Functional Suitability is the QA on which designers of LLM-based MASs pay the most attention, (3) Role-Based Cooperation is the design pattern most frequently employed among 16 patterns used to construct LLM-based MASs, and (4) Improving the Quality of Generated Code is the most common rationale behind the design of LLM-based MASs. Building on the study results, we presented the implications for the design of LLM-based MASs to support SE tasks.
Sketching compresses a large dataset into a smaller summary, allowing regression on the sketch to approximate the full-data objective. A sketch released under differential privacy can be reused with any non-private regression solver, any number of times, without further privacy loss. We study private sparse sketches for $\ell_2$ and $\ell_1$ regression. We first prove lower bounds on the noise required by common sparse sketches for both problems when the sketching matrix is public. Since regression on a sketch can be solved without revealing the sketching matrix, we construct sketches whose randomness is kept secret. Hiding how individual records enter the sketch provides an additional source of privacy, allowing less noise for the same privacy guarantee. We prove differential privacy guarantees for both constructions, and experiments on real datasets show improved utility over public-sketch baselines.
Large language models (LLMs) are increasingly studied as behavioral subjects (e.g., synthetic participants or users) and decision systems (e.g., LLM-as-a-judge), yet it remains unclear how observed cognitive biases relate to changes in model output probabilities. Anchoring bias, a classic human judgment bias, provides a useful test case. Prior research largely relies on generated responses, leaving distributional effects and input-field contributions less explored. This paper advances the study of anchoring in LLMs through three contributions: (1) a log-probability-based behavioral analysis examining how numerical anchors shift distributions over candidate outputs, using alternative questions and anchor pairs alongside a positive control to reduce reliance on potentially memorized stimuli; (2) exact Banzhaf-value attribution over structured prompt fields to quantify anchor contributions to log-probabilities; and (3) an Anchoring Bias Sensitivity Score (ABSS) integrating behavioral and attributional evidence across six open-source models. Results reveal predominantly positive behavioral anchoring effects, varying substantially across models. Gemma-2B, Phi-2, and Llama-2-7B achieve the highest ABSS rankings, whereas other models exhibit weaker behavioral-attributional alignment despite sometimes showing substantial behavioral shifts. Attributional effects also vary across prompts, indicating that behavioral anchoring does not necessarily imply stronger anchor-field contributions. These findings raise questions about LLM reliability as synthetic participants and potential vulnerabilities to numerical cues in automated evaluation and decision-making. More broadly, the framework connects behavioral science with explainable AI, particularly interpretability, and AI safety, providing a reproducible approach to investigating cognitive-bias-like effects in LLMs.
AI researchers have long focused on poker-like games as a testbed for environments characterized by multi-player dynamics, imperfect information, and reasoning under uncertainty. While recent breakthroughs have matched elite human play at no-limit Texas hold'em, the multi-player dynamics are subdued: most hands converge quickly with only two players engaged through multiple rounds of bidding. In this paper, we present Solly, the first AI agent to achieve elite human play in reduced-format Liar's Poker, a game characterized by extensive multi-player engagement. We trained Solly using self-play with a model-free, actor-critic, deep reinforcement learning algorithm. Solly played at an elite human level as measured by win rate (won over 50% of hands) and equity (money won) in heads-up and multi-player Liar's Poker. Solly also outperformed large language models (LLMs), including those with reasoning abilities, on the same metrics. Solly developed novel bidding strategies, randomized play effectively, and was not easily exploitable by world-class human players. In one extension, when Solly's policies were provided to LLMs as a domain-specific guide for inference, LLM performance improved and token costs decreased. This result suggests that low-cost RL agents can be used to boost LLM efficiency and reasoning capabilities.
We introduce a novel framework that transforms the resource-intensive (adversarial) prompt optimization problem into an \emph{efficient, amortized inference task}. Our core insight is that pretrained, non-autoregressive generative LLMs, such as Diffusion LLMs, which model the joint distribution over prompt-response pairs, can serve as powerful surrogates for prompt search. This approach enables direct conditional generation of prompts, effectively replacing costly discrete optimization with a small number of parallelizable samples. We provide a probabilistic analysis demonstrating that under mild fidelity assumptions, only a few conditional samples are required to recover high-reward (harmful) prompts. Empirically, our method substantially outperforms prior attacks in both success rate and compute cost, producing low-perplexity, diverse jailbreaks that transfer to a wide range of black-box target models, including robustly trained and proprietary LLMs. Beyond adversarial prompting, our framework opens new directions for red teaming, automated prompt optimization, and leveraging emerging Flow- and Diffusion-based LLMs.
PAC-Bayes is a popular and efficient framework for obtaining generalization guarantees in situations involving uncountable hypothesis spaces. Unfortunately, in its classical formulation, it only provides guarantees on the expected risk of a randomly sampled hypothesis. This requires stochastic predictions at test time, making PAC-Bayes unusable in many practical situations where a single deterministic hypothesis must be deployed. We propose a unified framework for extracting guarantees for a single hypothesis from PAC-Bayesian guarantees. We present a general derandomization framework, and derive from it a numerical bound for majority votes and for three different probability distributions over the posterior space. We study the empirical behaviour of our approach, showing that it consistently outperforms popular baselines in terms of generalization bounds for single classifiers (by up to a factor of 2), but sometimes requires a sacrifice in accuracy.
We introduce OpenMarkX, an extensible platform for watermarking open-source model and agent outputs under standard inference. It takes models, compatible methods, private keys, and evaluation settings, and produces watermarked weights, detection results, and quality and robustness reports. Its 16 text-watermark configurations are complemented by agent, vision-language, image, and physics-field workflows. Our proposed method, OpenMark, addresses learnability and robustness through joint model--policy learning and perturbation-aware training. On several open-source LLMs, OpenMark achieves 0.92 TPR at 1% FPR while reducing perplexity by 20.5% on average relative to the strongest prior method, and remains more detectable after such modifications, e.g., retaining 0.57 TPR at 5% FPR after 0.5-ratio model merging, compared with at most 0.34 for prior methods.