In previous articles, we proposed explaining the dark matter (DM) component, without invoking exotic matter, by utilizing the second component of General Relativity (GR) responsible for the Lense-Thirring effect. While the first component is the " g " Newtonian field, we designate this second component as the " k " gravitic field. In GR, the k-field originates from mass currents; however, our previous studies indicate that such mass currents cannot account for the magnitude of the k-field required to explain the entirety of DM. Given the similarity between the k-field and the magnetic field in electromagnetism (EM), we propose explaining its unexpectedly high value through a quantum approach to gravity, drawing an analogy with magnetic fields in EM which can only be explained by spin, a concept from quantum mechanics. We apply these concepts to explain the DM of the CMB. This makes it possible to obtain the order of magnitude of these fundamental concepts. Although this work is exploratory, these values could be used to test theories of Quantum Gravity and the concepts could also allow us to define a linear approximation of Quantum gravity inspired by the linearization of GR. This explanation could also constitute a step towards justifying the existence of a uniform gravitic field 0 k (pro- posed in previous articles) capable of explaining the DM of current astrophysical structures.

Tool-augmented language models ground their responses in external tool outputs, improving reliability beyond parametric knowledge alone.However, even when the correct tool is called and returns accurate information, models may still contradict such outputs, omit critical details, or generate unsupported claims.Existing hallucination benchmarks for toolaugmented dialogue primarily focus on the toolcalling stage, evaluating whether the correct tool and arguments are selected, while overlooking errors that arise during post-tool response generation.To address this gap, we introduce ToolHACE, a span-level benchmark for hallucination detection in tool-augmented LLM responses, covering five fine-grained hallucination categories.Through extensive experiments with modern hallucination detectors, we show that post-tool-call hallucinations are widespread and remain challenging to detect.Our results further demonstrate that effective detection requires training on specialized tool-augmented corpora, while transfer from tool-agnostic hallucination datasets performs poorly in this setting.These findings highlight the need for dedicated benchmarks and models for reliable hallucination detection in toolaugmented language systems.

LLMs increasingly serve as general-purpose information systems, including for factual queries about public statistics.In high-stakes domains such as migration, where data accuracy shapes both policy and public attitudes toward at-risk populations, models may answer queries about social groups unequally.We evaluate this concern empirically using GPT-4 Turbo, GPT-4o, GPT-5 chat, and GPT-5.1 on migration-related education statistics.Each model receives zero-shot queries on educational-attainment distributions for national and foreign citizens aged 25-54 across 34 European countries from 2020-2023.We compare outputs against official Eurostat data using RMSE, total variation distance (TVD), cosine similarity, distributional analysis, and response-coverage rates.Across models, results show a consistent gap between plausibility and accuracy: all models preserve broad distributional structure (cosine similarity > 0.93), while deviating substantially from ground-truth values (RMSE 10-14 and TVD 12-16 percentage points).We characterize this combination of structural plausibility and numerical error as an output-level analog of the illusion of competence.GPT-4 models also produce explicit non-answers for 30-78% of responses, with refusals (explicit non-answers) up to 3.4 times as frequent for foreigners than for nationals.GPT-5 models eliminate refusals but return hallucinated values.These findings show that pluralistic evaluation must distinguish legitimate human variation from unsupported model variation and assess whether factual information remains comparably available and reliable across social groups.

Code retrieval is vital to modern software engineering as it boosts reuse and speeds up debugging. However, current benchmarks primarily emphasize functional relevance while neglecting code quality. To address this gap, we introduce CoQuIR, the first large-scale, multilingual benchmark specifically designed to evaluate quality-aware code retrieval across four critical dimensions: correctness, efficiency, security, and maintainability. CoQuIR includes fine-grained quality annotations over 42,725 queries and 134,907 code snippets in 11 programming languages. Evaluating 23 retrievers (both open-source and proprietary) shows that even state-of-the-art models often fail to separate buggy or insecure code from robust counterparts. We further investigate methods for explicitly training retrievers to recognize code quality, demonstrating that quality-aware metrics can be improved without loss of semantic relevance; downstream code generation benefits from these gains. CoQuIR underscores the importance of embedding quality signals into retrieval systems as a crucial component for more trustworthy developer tools.

Recent advancements in large language models (LLMs) have underscored their vulnerability to safety alignment jailbreaks, particularly when subjected to downstream fine-tuning. However, existing mitigation strategies primarily focus on reactively addressing jailbreak incidents after safety guardrails have been compromised, removing harmful gradients during fine-tuning, or continuously reinforcing safety alignment throughout fine-tuning. As such, they tend to overlook a critical upstream factor: the role of the original safety-alignment data. This paper therefore investigates the degradation of safety guardrails through the lens of representation similarity between upstream alignment datasets and downstream fine-tuning tasks. Our experiments demonstrate that high similarity between these datasets significantly weakens safety guardrails, making models more susceptible to jailbreaks. Conversely, low similarity between these two types of datasets yields substantially more robust models and thus reduces harmfulness score by up to 10.33%. By highlighting the importance of upstream dataset design in the building of durable safety guardrails and reducing real-world vulnerability to jailbreak attacks, these findings offer actionable insights for fine-tuning service providers to prioritize upstream models with low jailbreak risk.

This thesis explores the role of carbon in climate-neutral energy systems. Climate mitigation is often framed as the decarbonization of the energy system, but in reality, in a climate-neutral system, carbon would still be present and circulate through remaining fossil fuels, biomass, CO2, synthetic fuels, chemicals, and negative emission technologies. The interactions between these components lead to complex mechanisms, calling for a systemic vision of carbon flows in the energy system.

To address this challenge, this work develops the Energy-Carbon Cycle Framework, which formalises the coupling between carbon sources and sinks under a neutrality constraint. It shows that residual fossil use is a possible system choice only with sufficient availability of carbon sinks. This concept is then translated into practice within the EnergyScope TD energy system model through a method that tracks the carbon flows between the different energy system technologies.

The proposed method is applied to the French energy system under climate neutrality constraints. This future climate-neutral energy system is compared with the national French strategy and other reference prospective documents. The results propose a clear merit order of lignocellulosic biomass use, in which sacrificing biomass to store carbon to reach climate neutrality is prioritised.

Building on this foundation, the thesis explores the specific case of the aviation sector. It is shown that e-bio-fuels are the most optimal pathway to produce sustainable aviation fuels (SAF), although this optimality is subject to uncertainty, which can favour the e-fuel pathway. The production cost of SAF, the total system cost, and the energy required for their production all depend strongly on the selectivity of the Fischer-Tropsch process towards jet fuel.

This work highlights that domestic production of SAF can reduce the need for flexible gas-based electricity production. These results are evaluated for different uncertainties surrounding the energy system, but the impact of interconnection and the recent downward trend in renewable energy CAPEX significantly reduce the initial values, bringing them to almost no effect.

Energy and carbon exchange between countries forces the expansion of the scope of this work to a European energy system, via the EnergyScope Multi-Cells model. To reduce the computational load of such an expansion, spatial and temporal aggregation methods are proposed and validated, leading to a reduction of the model's solve time by approximately 98 % without major change in energy system behaviour.

This new version of the model allows for uncertainty analysis of the European energy system. The robust decision-making method is used to identify carbon management strategies that are robust to uncertainties across multiple metrics, including total cost and life cycle impact assessment. This method particularly underlines that achieving European climate neutrality requires 339â 1018 MtCO2/yr of geological sequestration.

Overall, this work demonstrates that carbon remains a central component of climate-neutral energy systems. Rather than being eliminated, carbon resources must be understood from a systemic perspective, balancing two main roles: fulfilling the demand for carbonaceous sustainable fuels and enabling the system to reach climate neutrality by offsetting carbon sources with carbon sinks.

The emerging technology of wearable robots is gradually transforming our daily life. Up to 2026, many commericallized exoskeleton products have flourished, including those from companies ReWalk, HAL, ULS Robotics, Kenqing Technologies, ready on market to improve wearer's mobility and endurance. Nearly all the market-available exoskeletons are motor-driven, based on bulky external components, rigid transmission elements, or complex control architectures, which decreases wearing willingness. Developing soft wearable robots can well address some of the challenges of hard exoskeletons, improving wearing comfort, unobtrusiveness, and social ac-ceptance. However, to realize soft wearable robots critically depends on soft, thin, compact, and lightweight actuators to deliver sufficient torque (10 ~ 100 Nm) and displacement (5 ~ 30 cm) while maintaining energy-efficiency and reliability. This thesis investigates soft actuators in textile format (textile actuators) to enable unobtrusive actuation on body. This thesis explores the design, modeling, fabrication, experimental characterization, and application scenarios of fabric actuators based on shape memory alloys (SMAs) and electrostatic clutches (ESClutches). This thesis has three contributions to soft wearable robotics. Firstly, by encapsulating ESClutches and commercially available SMA springs into woven fabrics, we developed an energy-efficient fabric actuator (weighing 55 grams) that can lift 40 N mass and consumes milliwatts power for statically holding the mass. Secondly, using mechanics-informed principles, we developed X-Crossing architecture to enable SMA fabrics of 4.5 grams lifting 1 kg mass with 50% contraction, outperforming published SMA fabrics, including knitted and knotted SMA fabric actuators. Thirdly, we developed variable-stiffness mechanics model for X-Crossing SMA fabric actuator. This model predicts the actuator performances and add understandings to the actuator characteristics, including the peak in its performance curves and fundamental limits on actuator contraction. In this thesis, we also demonstrate the applications of developed fabric actuators. The clutchable fabric actuator mounted on the upper limb of a mannequin lifted 2.6 kilograms of mass, reducing 70 % energy consumption by using ESClutches compared to that without using ESClutches. Using this actuator, we demonstrate a waistband with controllable tightness as exoskeleton anchoring for force transmission to body, and using the ESClutches reduces 80% of energy consumption in defined tasks. We also demonstrate the X-Crossing architectured SMA fabric actuator can be well controlled to bend a mannequin elbow to target angles, and this fabric actuator can be fabricated as wrist sleeve to generate compression up to 13 kPa, and then maintain a residual compression pressure of 5 kPa with zero energy cost.

Low-density lipoproteins (LDL) are naturally occurring, monodisperse nanoparticles composed of proteins and lipids with an average diameter in the range of 20 nm. Their biological functions such as cholesterol trafficking and their roles in various diseases have been extensively studied. However, their properties as nanoparticles are seldom discussed. In this thesis, we investigate the formation of a protein corona on LDL, establish methods for its characterization, and explore the use of LDL-protein complex in cancer detection, biomarker discovery and disease monitoring. First, we show that LDL isolated from human serum or plasma carry a protein corona. Polyethylene glycol (PEG)-based precipitation followed by size-exclusion chromatography (SEC) preserves exchangeable proteins associated with LDL, while density-gradient ultracentrifugation largely removes the soft corona and yields LDL whose protein composition is dominated by ApoB-100. The two isolation strategies produce particles with distinct hydrodynamic size, homogeneity, and protein composition. Furthermore, incubation studies with purified proteins show negligible interaction of LDL with human serum albumin and apotransferrin, while immunoglobulin G exhibits only weak association, indicating that corona formation is selective rather than a consequence of indiscriminate adsorption of abundant plasma proteins. Building on this discovery, we focus on the study of LDL together with their protein corona, that we call protein-rich lipoproteins (ProLips), specifically on their potential in cancer detection. The tool we use is a proteomic comparison between samples from cancer patients and sample from healthy donors. Compared with whole-serum proteomics, ProLips proteomics consistently reveals substantially more differentially expressed proteins and provides clearer separation between healthy individuals and patients with melanoma, colon carcinoma, and liver cancer. The reduced contribution of highly abundant plasma proteins improves the detection of disease-associated proteins. Finally, six proteins as a melanoma-associated biomarker set are identified and validated. At last, we evaluate whether ProLips proteomics can monitor disease progression and treatment response. In a melanoma rat model, ProLips protein profiles distinguish different stages of tumor development. We show that in colorectal cancer patients sampled before, during, and after surgery, temporal changes in ProLips composition reflect surgery-associated remodeling, while persistent deviations after surgery are associated with clinical relapse. In liver cancer patients receiving immunotherapy, ProLips proteomics captures individualized treatment-associated changes, and both TMT- and DIA-based analyses reveal consistent biological trends despite differences in individual proteins identified.

Understanding turbulence in magnetically confined fusion plasmas is a major element in the reliable design and operation of fusion power plants. Turbulence produces interacting fluctuations ranging from the Larmor radius to the size of the machine, which are notoriously difficult to model. In particular, despite recent advances, the interaction of microturbulence, i.e. turbulence on the scale of the Larmor radius, with macroturbulence, on the scale of the machine size, is not well understood. In order to describe the microturbulence, usually a gyrokinetic ordering is imposed, where the amplitude of the turbulent fluctuations is assumed to be small and separated from the background ($\delta$-f approach). On the other hand, in order to describe the macroturbulence, a drift-kinetic ordering is usually imposed, where the turbulent fluctuations are assumed to have a large spatial extent and large amplitude, with no separation between equilibrium and fluctuating quantities (full-f approach). In this thesis, we combine both orderings and self-consistently derive a comprehensive full-f drift-kinetic and $\delta$-f gyrokinetic model. The standard two-fluid drift-reduced Braginskii model and the $\delta$-f gyrokinetic model are retrieved in the appropriate limits. Then, we use an orthogonal Hermite-Laguerre basis expansion to efficiently include collisions and obtain a numerically efficient model.

Simulations are carried out, first, for a higher-order drift-kinetic model in a linear plasma device showing clear limitations of this model. Simulations using the comprehensive full-f drift-kinetic and $\delta$-f gyrokinetic model are then carried out in the same geometry. A hybrid spectral and finite difference numerical scheme is implemented and the simulations show that the limitations seen in the extended drift-kinetic model are removed. Microturbulence is not present at the large ion-ion collision frequency, however, when collisions are suppressed, the gyrokinetic fields are seen to affect the turbulence.

Stepping up in complexity, we carry out simulations of the comprehensive full-f drift-kinetic and $\delta$-f gyrokinetic model in a tokamak geometry. We use two different levels of resistivity corresponding to increasing the input power. At low power, the simulations share several similarities with the simulations in the linear plasma device. At high power, the formation of a transport barrier at the plasma edge through an increase in the radial electric field is observed while the gyrokinetic fields are dominating the turbulence in the core of the device through universal modes.

Finally, we focus on the boundary conditions of the model. In the vicinity of solid walls, a Debye sheath forms where the quasi-neutrality assumption is violated. Therefore, any global simulation must impose boundary conditions at the sheath entrance. We consistently derive sheath boundary conditions in the case of perpendicular incidence of the magnetic field with the wall. Simulations with these boundary conditions are carried out in a linear plasma device geometry. The sheath boundary conditions are shown to have an impact on the particle balance and equilibrium state of the plasma.

Barrel cortex has a characteristic somatotopic map, where each whisker maps to a single barrel. Each cortical column therefore has a known principal whisker input, so a functional response can be studied against structure identified in advance. Individual barrels are visible in the CCFv3 average template, but the parcellation stops at area and layer. Anatomically detailed modelling of the full region had been done in rat, and no model of mouse barrel cortex reached this level of detail. In this thesis, I identified the features that distinguish barrel cortex, and supplied them to a model built in atlas space with the barrel columns annotated.

Firstly, I segmented 33 barrels in CCFv3, and used an area-preserving flatmap to annotate the barrel columns representing the posteromedial barrel subfield. The resulting columns follow the curvature of the cortex, span full depth and never overlap. The annotations are released with the flatmap as a tool for data integration and analysis.

Then, I used this enhanced atlas space and a pipeline validated in rat non-barrel somatosensory cortex to build a detailed model. The model contains 33 barrel columns, approximately 230,000 neurons and around 357 million synapses, drawn from 60 morphological and 15 electrical types. Every inhibitory neuron additionally carries a molecular class, so the model can be compared against mouse measurements directly. The connectome is built from anatomy alone, with no rule imposing how likely one cell type is to connect to another. I extended the pipeline with morphological synthesis, generalised electrical models and a new set of thalamocortical projections, with VPM confined to individual barrels and POm spread across the field.

In barrel cortex, layer 4 dendritic trees grow preferentially towards, and remain within, the barrel that carries their principal whisker input. Dendritic polarity has been described in individual cell types, but not measured across them. To address this, I curated dendritic reconstructions from mouse barrel cortex and applied the same descriptors to every cell, so classes can be ranked against one another. Polarity is class-dependent and concentrated in layer 4, in spiny stellate cells, non-Martinotti somatostatin neurons and parvalbumin neurons. I used the result to build two models differing in a single property of basal dendritic synthesis: one polarised, one symmetric. Finally, I ran in silico experiments on a 3x3 grid of barrel columns and the associated septum, with C2 as the principal column throughout. I simulated three experimental contexts: the quiet and whisking brain states, and active touch of the C2 whisker. Membrane potential and spikes were recorded from every neuron, with the afferent connectivity of each known. The model reproduces patterns reported in mouse: the laminar sequence of the touch response, the recruitment order of the molecular classes, and their dependence on brain state.

We decomposed the spiking response using cyclicity analysis and retained two components. The two modes have distinct spatial profiles: the first is focused on the principal column, while the second crosses the septum and the neighbouring barrels. The same decomposition applied across the two circuits divides the signal into a common and a contrast plane. The contrast plane isolates the unique response contributions of each circuit. Layer 4 dendritic polarisation sharpens the touch response in space and advances it in time.