Merck Serono (Italy)

Current literature suggests that alignment faking is an emergent property of large language models. We present the first empirical evidence that a small instruction-tuned model, specifically LLaMA 3 8B, can also exhibit alignment faking. We further show that prompt-only interventions, including deontological moral framing and scratchpad reasoning, significantly reduce this behavior without modifying model internals. This challenges the assumption that prompt-based interventions are trivial and that deceptive alignment requires scale. We introduce a taxonomy distinguishing shallow deception, shaped by context and suppressible through prompting, from deep deception, which reflects persistent, goal-driven misalignment. Our findings refine the understanding of deception in language models and underscore the need for deceptive alignment evaluations across model sizes and deployment settings.

In automotive electronic control systems, control logic is executed based on the vehicle's driving state, which is typically measured using onboard sensors. However, due to limitations in sensor accuracy and cost, many critical state parameters required for vehicle-level control must be estimated rather than directly measured. To address the challenge of accurate state estimation under complex and dynamic driving conditions, this paper proposes a vehicle state estimation framework based on temporal feature modeling. The method is built around Long Short-Term Memory (LSTM) neural networks and incorporates a bidirectional recurrent structure along with a sliding window mechanism. It constructs a dynamic estimation model for key state parameters–including longitudinal velocity, lateral velocity, pitch angle, and roll angle–using historical sequences of multi-source sensor data. Simulation results under various driving conditions–including acceleration/deceleration, slopes, and low-adhesion surfaces–demonstrate that the proposed method can stably estimate essential vehicle states, providing reliable inputs for vehicle stability control algorithms. This study shows that deep learning methods based on temporal feature extraction can effectively overcome the limitations of traditional physical models in adapting to complex driving scenarios.

This paper proposes a novel time-varying high-gain feedback control method for active suspension systems, aiming at achieving accurate and rapid regulation of vehicle body height and attitude, thus significantly enhancing ride comfort and vehicle stability. Firstly, multiple uncertainties, including sprung mass and suspension parameters, are comprehensively considered in system modeling. Subsequently, based on the principle of separating uncertain parameters from their nominal values, the dynamic control model of a half-car active suspension system is established, and the upper bounds of associated uncertainties are accurately analyzed. On this basis, time-varying high-gain feedback control laws are proposed for vehicle body vertical displacement and pitch angle, respectively. Finally, comparative simulation studies with both passive suspension and adaptive backstepping control are conducted, demonstrating that the proposed method provides superior dynamic response performance and higher steady-state tracking accuracy.

Although the role of cellular immunity in checkpoint immunotherapy (CPI) for cancer is well established1,2, the effect of antibody-mediated humoral immunity is comparably underexplored. Here we used rapid extracellular antigen profiling3 to map the autoantibody reactome within a cohort of 374 patients with cancer treated with CPIs and 131 healthy control participants for autoantibodies to 6,172 extracellular and secreted proteins (the 'exoproteome'). Globally, patients with cancer treated with CPIs had diverse autoreactivities that were elevated relative to control individuals but changed minimally with treatment. Autoantibody signatures in patients treated with CPI strikingly distinguished them from healthy individuals. Although associations of specific autoantibodies with immune-related adverse events were sparse, we detected numerous individual autoantibodies that were associated with greatly altered odds ratios for response to therapy. These included autoantibodies to immunomodulatory proteins, such as cytokines, growth factors and immunoreceptors, as well as tumour surface proteins. Functional evaluation of several autoantibody responses indicated that they neutralized the activity of their target proteins, which included type I interferons (IFN-I), IL-6, OSM, TL1A, and BMPR1A and BMPR2. Modelling the effects of autoantibodies to IFN-I and TL1A in preclinical mouse tumour models resulted in enhanced CPI efficacy, consistent with their effects in patients. In conclusion, these findings indicate that autoantibodies to the exoproteome modify CPI responses and highlight therapeutically actionable pathways that can be exploited to augment immunotherapy.

PurposeThe paper aims to investigate the biological role of microRNAs secreted by preimplantation embryo into the blastocoel fluid and to detect a distinctive molecular signature for identifying embryos with the highest implantation potential.MethodsWe carried on a multicenter retrospective study involving five European IVF centers. We collected 112 blastocoel fluid samples from embryos on day 5 post-fertilization, cultured individually, along with data on blastocyst grade and embryo transfer outcomes. Using a custom TLDA Array, we compared the expression levels of 89 miRNAs between 33 fluids from high-quality implanted embryos and 30 fluids from high-quality not-implanted embryos. Expression differences were assessed using SAM and t-test. Additionally, correlation and function enrichment analysis and network construction were conducted to identify the biological roles of deregulated microRNAs.ResultsWe identified six up-regulated microRNAs in the blastocoel fluid from implanted embryos, significantly and positively correlated across all samples (r ≥ 0.7; P ≤ 0.05). They could take part in pluripotency circuits, regulating and being regulated by transcription factors associated with stemness, cell growth, and embryo development. The ROC curve analysis confirmed the potential of these miRNAs as implantation classifiers.ConclusionThe six miRNAs up-regulated in blastocoel fluid from implanted embryos may represent a functional molecular signature for evaluating blastocyst quality and identifying the most competent embryos. Their evaluation associated with non-invasive preimplantation genetic testing, integrating epigenetic and genomic analyses, could enhance implantation grade and allow for identification of the euploid embryo not able to implant.

The assessment and control of potential mutagenic impurities (PMIs) within pharmaceutical products are managed in accordance with the ICH M7 guideline “Assessment and control of DNA reactive (mutagenic) impurities in pharmaceuticals to limit potential carcinogenic risk”. This guidance highlights four control options that can be used to give assurance of control of PMIs to below a level of toxicological concern for the intended use. These control options range from testing to confirm that impurity levels within the active pharmaceutical ingredient or product are below an acceptable limit (ICH M7 Option 1 Control) to a control strategy that relies on process controls and scientific principles (e.g., purging) to demonstrate impurity presence to below a level of concern in lieu of analytical testing (option 4). While ICH M7 control option 4 is an established approach to justify that levels of a potential mutagenic impurity are below an acceptable limit, there have been health authority challenges that the use of ICH M7 control option 4 rationales is not appropriate for N-nitrosamines without included confirmatory analytical testing data to confirm absence. The reasons behind this lack of acceptance for ICH M7 control option 4 alone may include (i) a higher perceived potency for nitrosamines over other mutagenic impurities as they alert as part of the cohort of concern and (ii) inappropriate application of purge rationales such that, in some instances, confirmatory testing data highlighted higher levels for the impurity than had been predicted to be present. Through the inclusion of industry relevant case studies, this publication outlines that, while some nitrosamines may require control to lower levels than the ICH M7 threshold of toxicological concern, the concept of the ICH M7 option 4 control is scientifically justified when the properties for the nitrosamine are considered and an appropriate conservative purge rationale is applied.

Abstract. Hydrological models are fundamental tools for the characterization and management of karst systems. We propose an updated version of KarstMod, software dedicated to lumped-parameter rainfall–discharge modelling of karst aquifers. KarstMod provides a modular, user-friendly modelling environment for educational, research, and operational purposes. It also includes numerical tools for time series analysis, model evaluation, and sensitivity analysis. The modularity of the platform facilitates common operations related to lumped-parameter rainfall–discharge modelling, such as (i) setup and parameter estimation of a relevant model structure and (ii) evaluation of internal consistency, parameter sensitivity, and hydrograph characteristics. The updated version now includes (i) external routines to better consider the input data and their related uncertainties, i.e. evapotranspiration and solid precipitation; (ii) enlargement of multi-objective calibration possibilities, allowing more flexibility in terms of objective functions and observation type; and (iii) additional tools for model performance evaluation, including further performance criteria and tools for model error representation.