Existing benchmarks for computational materials discovery primarily evaluate static predictive tasks or isolated computational sub-tasks. While valuable, these evaluations neglect the inherently iterative and adaptive nature of scientific discovery. We introduce MAterials Discovery Environments (MADE), a novel framework for benchmarking end-to-end autonomous materials discovery pipelines. MADE simulates closed-loop discovery campaigns in which an agent or algorithm proposes, evaluates, and refines candidate materials under a constrained oracle budget, capturing the sequential and resource-limited nature of real discovery workflows. We formalize discovery as a search for thermodynamically stable compounds relative to a given convex hull, and evaluate efficacy and efficiency via comparison to baseline algorithms. The framework is flexible; users can compose discovery agents from interchangeable components such as generative models, filters, and planners, enabling the study of arbitrary workflows ranging from fixed pipelines to fully agentic systems with tool use and adaptive decision making. We demonstrate this by conducting systematic experiments across a family of systems, enabling ablation of components in discovery pipelines, and comparison of how methods scale with system complexity.

Mosquito Species Classification (MSC) is crucial for vector surveillance and disease control. The collection of mosquito bioacoustic data is often limited by mosquito activity seasons and fieldwork. Mosquito recordings across regions, habitats, and laboratories often show non-biological variations from the recording environment, which we refer to as domain features. This study finds that models directly trained on audio recordings with domain features tend to rely on domain information rather than the species' acoustic cues for identification, resulting in illusory good performance while actually performing poor cross-domain generalization. To this end, we propose a Domain-Robust Bioacoustic Learning (DR-BioL) framework that combines contrastive learning with distribution alignment. Contrastive learning aims to promote cohesion within the same species and mitigate inter-domain discrepancies, and species-conditional distribution alignment further enhances cross-domain species representation. Experiments on a multi-domain mosquito bioacoustic dataset from diverse environments show that the DR-BioL improves the accuracy and robustness of baselines, highlighting its potential for reliable cross-domain MSC in the real world.

Mosquito-borne diseases affect more than one billion people each year and cause close to one million deaths. Traditional surveillance methods rely on traps and manual identification that are slow, labor-intensive, and difficult to scale. Audio-based mosquito monitoring offers a non-destructive, lower-cost, and more scalable complement to trap-based surveillance, but reliable species classification remains difficult under real-world recording conditions. Mosquito flight tones are narrow-band, often low in signal-to-noise ratio, and easily masked by background noise, and recordings for several epidemiologically relevant species remain limited, creating pronounced class imbalance. Variation across devices, environments, and collection protocols further increases the difficulty of robust classification. Such variation can cause models to rely on domain-specific recording artefacts rather than species-relevant acoustic cues, which makes transfer to new acquisition settings difficult. The BioDCASE 2026 Cross-Domain Mosquito Species Classification (CD-MSC) challenge is designed around this deployment problem by evaluating performance on both seen and unseen domains. This paper presents the official baseline system and evaluation pipeline as a simple, fully reproducible reference for the CD-MSC challenge task. The baseline uses log-mel features and a multitemporal resolution convolutional neural network (MTRCNN) with species and auxiliary domain outputs, together with complete training and test scripts. The baseline system performs strongly on seen domains but degrades markedly on unseen domains, showing that cross-domain generalisation, rather than within-domain recognition, is the central challenge for practical mosquito species classification from multi-source bioacoustic recordings.

Bayesian synergistic filter (BSF), a novel generic filter for information fused model rectification and state-space estimation with uncertainty quantification, is proposed. Model discrepancy between the actual system and the computational model is a challenging issue for reliable state-space estimation. Due to the model discrepancy, the state-space estimation can be seriously distorted and misleading. To tackle this challenge, the proposed BSF evaluates the model discrepancy by fully exploiting the subtle information between the measurements and the corresponding predictions. By compensating for the evaluated model discrepancy, the computational model can be rectified to minimize its deviation from the actual system. Therefore, the computational model and the model rectification are synergized to develop a reliable state-space estimation. Moreover, the proposed BSF fuses the model rectification and the noise parameter estimation such that the stationarity constraint in conventional state-space filters can be released. Consequently, the proposed BSF achieves reliable state-space estimation for general linear/nonlinear systems under stationary/nonstationary situations. Furthermore, by taking the benefit of Bayesian inference, the estimation uncertainties of all estimates can be quantified accordingly. To demonstrate the efficacy and applicability of the proposed BSF, two simulated examples with various model discrepancies and stationary/nonstationary scenarios are discussed. Moreover, the performance of a conventional state-space filter and ablation tests are presented for comparison. Besides, a case study of a 600-meter-high television tower subjected to nonstationary earthquake excitation is presented.

The advancement of large language models (LLMs) has accelerated the development of autonomous financial trading systems. While mainstream approaches deploy multi-agent systems mimicking analyst and manager roles, they often rely on abstract instructions that overlook the intricacies of real-world workflows, which can lead to degraded inference performance and less transparent decision-making. Therefore, we propose a multi-agent LLM trading framework that explicitly decomposes investment analysis into fine-grained tasks, rather than providing coarse-grained instructions. We evaluate the proposed framework using Japanese stock data, including prices, financial statements, news, and macro information, under a leakage-controlled backtesting setting. Experimental results show that fine-grained task decomposition significantly improves risk-adjusted returns compared to conventional coarse-grained designs. Crucially, further analysis of intermediate agent outputs suggests that alignment between analytical outputs and downstream decision preferences is a critical driver of system performance. Moreover, we conduct standard portfolio optimization, exploiting low correlation with the stock index and the variance of each system's output. This approach achieves superior performance. These findings contribute to the design of agent structure and task configuration when applying LLM agents to trading systems in practical settings.

Model merging combines fine-tuned checkpoints into a single multi-task model without retraining. Existing methods - such as task arithmetic, model soups, TIES, and DARE - are computationally efficient and empirically successful, but rely on heuristic design choices and lack formal optimality guarantees. We show that merging can be formulated as a convex quadratic programme over residual updates, yielding weights that minimise a squared-output calibration objective using calibration inputs and fine-tuned model outputs, and subsuming existing methods as special cases. Our framework yields a closed-form diagnostic - the fraction of residual energy captured by a chosen basis - that predicts downstream merge quality using only the calibration set. Empirically, the QP matches or outperforms existing methods in the single-layer setting, and we characterise when the optimal basis provides significant gains over the cheaper diagonal QP. We extend to multi-layer merging via a sequential layer-wise algorithm and demonstrate consistent gains across language and vision benchmarks.

The design and operation of tidal stream energy farms will require the accurate prediction of tidal currents from field data. In the present paper we compare the performance of a physics based flow model, the traditional harmonic analysis method, and a newly developed code RTide. RTide is based on the Response Method proposed by Munk and Cartwright in the 1960s and uses machine learning to overcome the key disadvantages of their original approach. We use field data from the Meygen site in the Inner Sound of the Pentland Firth. We find RTide predictions based solely on the field data outperform both the flow model predictions and the currents predicted by harmonic analysis. Interestingly, feeding the physics-based flow model into RTide as an additional input does not outperform the predictions based on the field data alone. However, RTide can still be used as a correction tool to physics-based flow models.

The volume and complexity of scientific literature are expanding rapidly, making it increasingly difficult to extract and synthesise information across studies. This chal-lenge is particularly evident in the biological sciences, where data span from molecules to ecosystems and linking across hierarchical levels is inherently challenging. Large Language Model (LLM) pipelines offer a scalable solution, but most lack modular-ity, transparency, and mechanisms for human oversight. We present MetaBeeAI, an open-source, modular pipeline that uses LLMs to extract structured information from scientific papers for systematic review and meta-analysis. The system includes an intuitive interface that displays model outputs alongside the source text, allowing users to inspect, correct, and iteratively improve performance. MetaBeeAI produces an auditable, machine-readable record of prompts, configuration settings, and expert annotations, supporting reliable replication and continual refinement. We evaluated the pipeline on 924 research papers, extracting information on bee species, pesticides, exposure methodologies, and other environmental factors. Results highlight the value of expert-in-the-loop validation for prompt optimisation and show that MetaBeeAI can handle heterogeneous experimental designs and biological contexts. MetaBeeAI provides a general framework for structured knowledge extraction, enabling scalable, transparent, and reproducible evidence synthesis, and new approaches to accelerate discovery in the life sciences using AI. ### Competing Interest Statement The authors have declared no competing interest. Schmidt Sciences, https://ror.org/044fk6795 The Kavli Foundation, https://ror.org/00kztt736 AIBIO-UK Rhodes Trust, https://ror.org/04v48nr57 UK Research and Innovation, https://ror.org/001aqnf71, MR/T021691/1 European Science Foundation Swedish Research Council, VR 2021-03194, 2018-01466 Stiftung Vinetum, https://ror.org/05xs0ct17

The urgent pursuit of net-zero emissions presents a critical challenge for modern societies, necessitating a speedup of transformative shifts across sectors to mitigate climate change. Predicting trends and drivers in the integration of energy technologies is essential to addressing this challenge, as it informs policy decisions, strategic investments, and the deployment of innovative solutions crucial for transitioning to a sustainable energy future. Despite the importance of accurate forecasting, current methods remain limited, especially in leveraging the vast, unlabelled energy literature available. However, with the advent of large language models (LLMs), the ability to interpret and extract insights from extensive textual data has significantly advanced. Sentiment analysis, in particular, has just emerged as a vital tool for detecting scientific opinions from the energy literature, which can be harnessed to forecast energy trends. This study introduces a novel multi-agent framework, EnergyEval, to evaluate the sentiment and factuality of the energy literature. The core novelty of the multi-agent framework is found to be the use of heterogeneous energy-specialised roles with different LLMs. This investigation, using both multiple persona agents and different LLMs, provides a bespoke collaboration mechanism for multi-agent debate (MAD). In addition, we believe our approach can extend across the energy industry, where deep application of MAD is yet to be exploited. We apply EnergyEval to the case of UK offshore wind literature, assessing its predictive performance. Our findings indicate that the sentiment predicted by the EnergyEval effectively aligns with observed trends in increasing the installed capacity and reductions in Levelised Cost of Energy (LCOE). It also helps us to identify key drivers in offshore wind development. The advantage of employing a multi-agent LLM debate team allows us to achieve competitive accuracy compared to single-LLM-based methods, while significantly reducing computational costs. Overall, the results highlight the potential of EnergyEval as a robust tool for forecasting technology developments in the pursuit of net-zero emissions.

Data is often generated in streams, with new observations arriving over time. A key challenge for learning models from data streams is capturing relevant information while keeping computational costs manageable. We explore intelligent data subsampling for offline learning, and argue for an information-theoretic method centred on reducing uncertainty in downstream predictions of interest. Empirically, we demonstrate that this prediction-oriented approach performs better than a previously proposed information-theoretic technique on two widely studied problems. At the same time, we highlight that reliably achieving strong performance in practice requires careful model design.

Malaria is one of the deadliest vector borne diseases affecting sub-Saharan Africa. A suite of systems are being used to monitor and manage malaria risk and disease incidence, with an increasing focus on technological interventions that allow private citizens to remotely record and upload data. However, data collected by citizen scientists must be standardised and consistent if it is to be used for scientific analysis. Studies that aim to improve data collection quality and quantity have often included incentivisation, providing citizen scientists with monetary or other benefits for their participation in data collection. We tested whether monetary incentives enhance participation and data collection in a study trialling an acoustic mosquito sensor. Working with the community in two health areas in the Democratic Republic of Congo, we measured data collection participation, completeness, and community responses. Our results showed mixed responses to the incentive, with more participants interested in the social status and monetary value of the technology used than the monetary incentive itself. The effect of incentives on data collection varied over the course of the trial, increasing participation in the start of the trial but with no effect in the latter half of the trial. Feedback from participants showed that opinions on technology, research objectives, and incentives varied between communities, and was associated with differences in data collection quantity and quality, suggesting that differences in community interest in data collection and the incentives may be more important than the incentive value itself. These results suggest that though there is an initial benefit, extrinsic motivations do not override differences in intrinsic motivations over time, and enhanced communication and dialogue with participants may improve citizen science participation and attitudes.

Following rapid ice thinning in the mid-Holocene, Pope Glacier (adjacent to Thwaites Glacier in the Amundsen Sea sector) was at least 30-35 m thinner than present for at least 3 kyr in the mid- to Late Holocene. The timing of the end of this ice lowstand and subsequent rethickening of ice to near its present configuration is poorly constrained. We present five paired 10Be and 26Al cosmogenic nuclide exposure ages that provide constraints on the timing of this ice sheet readvance. The ages are sourced from samples collected from a moraine

BACKGROUND:Current malaria surveillance methods are considered too expensive to scale-up within limited-resource settings; hence, new technologies and approaches are necessary to maximize the collection of data and ultimately design new malaria control tools. Effective mosquito surveillance can be enhanced through the utilization of digital technologies and the engagement of citizens in real-time data collection. This study used the HumBug acoustic sensor with the MozzWear app to detect and identify host-seeking mosquitoes based on their flight sounds, with citizens receiving airtime incentives for recording and uploading sounds. METHODS:A randomized controlled trial was used to assess the role of incentives to encourage the local community to record and upload mosquito sounds using the MozzWear application. Participants were randomized into two groups: (1) a control group, in which no incentive was provided; and (2) an incentive group, in which airtime credit was provided to participants. Both groups were provided with HumBug smartphones running the MozzWear app plus adapted mosquito bed nets ('HumBug Nets') to hold the phones during recording and were asked to record and upload mosquito flight tone data once per week for a period of four months. The intervention group was rewarded with an airtime incentive every week after the data were uploaded. At the end of the study, an experience survey was administered to participants in both groups to assess their experience participating in this study. RESULTS:The overall results indicate that the control group performed well in terms of the number of nights spent recording and uploading data compared to the incentive group. The level of intrinsic and extrinsic motivation differs between demographic variables. Their feedback suggested that fighting against malaria was more important and was everyone's concern in rural Tanzania. In addition, the participants expressed their interest in being involved in future research related to mosquito surveillance and the fight against malaria. CONCLUSION:Citizens can play a valuable role in scientific research; even without giving them incentives, they can still participate in the study. By participating in mosquito surveillance and malaria prevention studies, community members can make significant contributions to addressing mosquito-borne diseases and improving health outcomes.

The NASA Surface Water Ocean Topography mission (SWOT), launched on 16 December 2022, will provide the highest spatial and temporal altimetric measurements of coastal oceans to date. The mission is ideally suited to studying mesoscale and submesoscale processes and is expected to enhance our understanding of coastal tides greatly. Although improved tidal analysis and prediction is, of course, useful for studying tides, arguably more important is the accurate removal of tidal variability from SWOT observations. This is a consequence of the fact that the tidal signal often dominates other sub-mesoscale processes which are of high interest to SWOT researchers. While SWOT presents unprecedented spatial resolution, the temporal sparsity renders the applications of conventional tidal analysis methods difficult in the early stages of the mission. Despite significant improvements in global and regional barotropic tidal models in the past few decades, the complexity of coastal and estuarine tides as well as the relatively limited in-situ measurements available for assimilation can lead to significant errors when used for tidal corrections. Further complications are introduced by the uncertainty in mean sea surface (MSS) estimates from gridded MSS Products. These errors can account for large percentages of the global Sea Level Anomaly error and grow significantly over rough bathymetry. As such, the accurate assessment of the uncertainty for the released data products and corrections derived from primary SWOT data are critical to the early success of SWOT science teams. We develop a fully Bayesian variant of tidal harmonic analysis to achieve tidal super-resolution and MSS correction for early-stage (

On September 16th, 2023, an anomalous 10.88 mHz seismic signal was observed globally, persisting for 9 days. One month later an identical signal appeared, lasting for another week. Several studies have theorized that these signals were produced by seiches which formed after two landslide-generated mega-tsunamis in an East Greenland fjord. This theory is supported by seismic inversions, and analytical and numerical modeling, but no direct observations have been made. Here, we present primary observations of this phenomenon using data from the Surface Water Ocean Topography mission. By ruling out other oceanographic processes, we validate the seiche theory of previous authors and independently estimate its initial amplitude at 7.9 m using Bayesian machine learning and seismic data. This study demonstrates the value of satellite altimetry for studying fast oceanic processes and extreme events, while also highlighting the need for specialized methods to address the altimetric data's limitations, namely temporal sparsity. These data and approaches will help in understanding future unseen extremes driven by climate change.

Tides pose significant operational and engineering challenges and are critical drivers of many natural processes. Accurate tidal predictions are important for modeling these phenomena. Conventionally, tidal prediction is carried out using harmonic analysis, the accuracy of which degrades when non-stationary and nontidal forcing are present. While Munk and Cartwright’s response method avoids the assumptions that give rise to this degradation, the difficulty of defining realistic interactions between inputs has inhibited automated applications. Here, we develop a non-parametric framework for tidal analysis and prediction of sea levels under compound forcing. The approach embeds a class of neural networks capable of representing any arbitrary Volterra series—the mathematical basis of the response method—within the classic method. The new ML Response Framework overcomes the automation challenges imposed by the original method and can directly infer high-order nonlinear interactions. This makes the inclusion of meteorological and other non-tidal forcing straightforward. Furthermore, we show that by accounting for this nonstationarity explicitly, a better astronomical tidal estimate is obtained. A method is devised to obtain physical insights from the learned model, illustrating how it can be used to study the interaction and modulation of astronomical tides by external forcing. By taking a nonparametric approach, our framework makes the study of phenomena that heretofore could not be accounted for straightforward. We provide several case studies, including the analysis and prediction of tide-surge interaction, riverine tides, and nuisance flooding. These applications, and more, can be replicated using only three lines of code with the open-source Python package (RTide).

Recent work has reported that AI classifiers trained on audio recordings can accurately predict severe acute respiratory syndrome coronavirus 2 (SARSCoV2) infection status. Here, we undertake a large scale study of audio-based deep learning classifiers, as part of the UK governments pandemic response. We collect and analyse a dataset of audio recordings from 67,842 individuals with linked metadata, including reverse transcription polymerase chain reaction (PCR) test outcomes, of whom 23,514 tested positive for SARS CoV 2. Subjects were recruited via the UK governments National Health Service Test-and-Trace programme and the REal-time Assessment of Community Transmission (REACT) randomised surveillance survey. In an unadjusted analysis of our dataset AI classifiers predict SARS-CoV-2 infection status with high accuracy (Receiver Operating Characteristic Area Under the Curve (ROCAUC) 0.846 [0.838, 0.854]) consistent with the findings of previous studies. However, after matching on measured confounders, such as age, gender, and self reported symptoms, our classifiers performance is much weaker (ROC-AUC 0.619 [0.594, 0.644]). Upon quantifying the utility of audio based classifiers in practical settings, we find them to be outperformed by simple predictive scores based on user reported symptoms.

From early in the coronavirus disease 2019 (COVID-19) pandemic, there was interest in using machine learning methods to predict COVID-19 infection status based on vocal audio signals, for example, cough recordings. However, early studies had limitations in terms of data collection and of how the performances of the proposed predictive models were assessed. This article describes how these limitations have been overcome in a study carried out by the Turing-RSS Health Data Laboratory and the UK Health Security Agency. As part of the study, the UK Health Security Agency collected a dataset of acoustic recordings, SARS-CoV-2 infection status and extensive study participant meta-data. This allowed us to rigorously assess state-of-the-art machine learning techniques to predict SARS-CoV-2 infection status based on vocal audio signals. The lessons learned from this project should inform future studies on statistical evaluation methods to assess the performance of machine learning techniques for public health tasks.

We analyse and explain the increased generalisation performance of iterate averaging using a Gaussian process perturbation model between the true and batch risk surface on the high dimensional quadratic. We derive three phenomena \latestEdits{from our theoretical results:} (1) The importance of combining iterate averaging (IA) with large learning rates and regularisation for improved regularisation. (2) Justification for less frequent averaging. (3) That we expect adaptive gradient methods to work equally well, or better, with iterate averaging than their non-adaptive counterparts. Inspired by these results\latestEdits{, together with} empirical investigations of the importance of appropriate regularisation for the solution diversity of the iterates, we propose two adaptive algorithms with iterate averaging. These give significantly better results compared to stochastic gradient descent (SGD), require less tuning and do not require early stopping or validation set monitoring. We showcase the efficacy of our approach on the CIFAR-10/100, ImageNet and Penn Treebank datasets on a variety of modern and classical network architectures.