This dissertation investigates how reinforcement learning (RL) methods can be designed to be safe, sample-efficient, and robust. Framed through the unifying perspective of contextual-bandit RL, the work addresses two major application domains - ranking and recommendation, and text-to-image diffusion models. The first part of the thesis develops theory and algorithms for safe deployment in ranking systems. An exposure-based generalisation bound is derived, leading to a counterfactual risk-minimisation objective whose solution is guaranteed not to underperform the logging policy, even with sparse feedback. This guarantee is extended to doubly robust estimators, enabling safety even under adversarial or misspecified user models and offering practitioners explicit control over permissible utility loss. The second part turns to single-action bandits, where various off-policy estimators are unified within a baseline-correction framework. A closed-form optimal baseline is proposed and shown to minimise both evaluation and policy-gradient variance, thereby improving off-policy learning reliability. The final part examines the trade-offs between efficiency and effectiveness in generative RL. A systematic study of PPO and REINFORCE motivates the Leave-One-Out PPO (LOOP) algorithm, which combines multiple diffusion trajectories with a REINFORCE-style baseline inside PPO's clipped objective. LOOP achieves PPO-level sample efficiency while producing generations that align more faithfully with textual attributes.

Large Language Models (LLMs) have been shown to achieve impressive results for many reasoning-based NLP tasks, suggesting a degree of deductive reasoning capability. However, it remains unclear to which extent LLMs, in both informal and autoformalisation methods, are robust on logical deduction tasks. Moreover, while many LLM-based deduction methods have been proposed, a systematic study that analyses the impact of their design components is lacking. Addressing these two challenges, we propose the first study of the robustness of formal and informal LLM-based deductive reasoning methods. We devise a framework with two families of perturbations: adversarial noise and counterfactual statements, which jointly generate seven perturbed datasets. We organize the landscape of LLM reasoners according to their reasoning format, formalisation syntax, and feedback for error recovery. The results show that adversarial noise affects autoformalisation, while counterfactual statements influence all approaches. Detailed feedback does not improve overall accuracy despite reducing syntax errors, pointing to the challenge of LLM-based methods to self-correct effectively.

We consider the XXZ quantum spin chain in its massless, disordered regime at zero field. We derive an exact expression for the two-spinon form-factor of $S^z=1/2\sigma^z$ by taking a limit of the massive XYZ form-factors found by Lashkevich and by Lukyanov and Terras. This result is used to find the two-spinon contribution to the spectral decomposition of the longitudinal structure factor $S^{zz}(k,w)$. We find that this contribution provides an accurate approximation to the full structure factor over a wide range of the anisotropy parameter. The asymptotic behaviour of $S^{zz}(k,w)$ is computed as the upper and lower $w$ thresholds of the two-spinon $(w,k)$ band are approached, and an analysis of the region of validity of this threshold behaviour is performed. Our results reproduce and refine existing threshold behaviour predictions and extend these results to an accurate description throughout the two-spinon continuum.

Despite substantial advancements in aligning large language models (LLMs) with human values, current safety mechanisms remain susceptible to jailbreak attacks. We hypothesize that this vulnerability stems from distributional discrepancies between alignment-oriented prompts and malicious prompts. To investigate this, we introduce LogiBreak, a novel and universal black-box jailbreak method that leverages logical expression translation to circumvent LLM safety systems. By converting harmful natural language prompts into formal logical expressions, LogiBreak exploits the distributional gap between alignment data and logic-based inputs, preserving the underlying semantic intent and readability while evading safety constraints. We evaluate LogiBreak on a multilingual jailbreak dataset spanning three languages, demonstrating its effectiveness across various evaluation settings and linguistic contexts.

The increasing electricity demands of personal computers, communication networks, and data centers contribute to higher atmospheric greenhouse gas emissions, which in turn lead to global warming and climate change. Therefore the energy consumption of code must be minimized. Code can be generated by large language models. We look at the influence of prompt modification on the energy consumption of the code generated. We use three different Python code problems of varying difficulty levels. Prompt modification is done by adding the sentence ``Give me an energy-optimized solution for this problem'' or by using two Python coding best practices. The large language models used are CodeLlama-70b, CodeLlama-70b-Instruct, CodeLlama-70b-Python, DeepSeek-Coder-33b-base, and DeepSeek-Coder-33b-instruct. We find a decrease in energy consumption for a specific combination of prompt optimization, LLM, and Python code problem. However, no single optimization prompt consistently decreases energy consumption for the same LLM across the different Python code problems.

When metals are magnetized, emulsions phase separate, or galaxies cluster, domain walls and patterns form and irremediably coarsen over time. Such coarsening is universally driven by diffusive relaxation toward equilibrium. Here, we discover an inertial counterpart - wave coarsening - in active elastic media, where vibrations emerge and spontaneously grow in wavelength, period, and amplitude, before a globally synchronized state called a time crystal forms. We observe wave coarsening in one- and two-dimensional solids and capture its dynamical scaling. We further arrest the process by breaking momentum conservation and reveal a far-from-equilibrium nonlinear analogue to chiral topological edge modes. Our work unveils the crucial role of symmetries in the formation of time crystals and opens avenues for the control of nonlinear vibrations in active materials.

We determine the late-time dynamics of a generic spin ensemble with inhomogeneous broadening - equivalently, qubits with arbitrary Zeeman splittings - coupled to a dissipative environment with strength decreasing as $1/t$. The approach to the steady state follows a power law, reflecting the interplay between Hamiltonian dynamics and vanishing dissipation. The decay exponents vary non-analytically with the ramp rate, exhibiting a cusp singularity, and $n$-point correlation functions factorize into one- and two-point contributions. Our exact solution anchors a universality class of open quantum systems with explicitly time-dependent dissipation.

Group convolutional neural networks (G-CNNs) have been shown to increase parameter efficiency and model accuracy by incorporating geometric inductive biases. In this work, we investigate the properties of representations learned by regular G-CNNs, and show considerable parameter redundancy in group convolution kernels. This finding motivates further weight-tying by sharing convolution kernels over subgroups. To this end, we introduce convolution kernels that are separable over the subgroup and channel dimensions. In order to obtain equivariance to arbitrary affine Lie groups we provide a continuous parameterisation of separable convolution kernels. We evaluate our approach across several vision datasets, and show that our weight sharing leads to improved performance and computational efficiency. In many settings, separable G-CNNs outperform their non-separable counterpart, while only using a fraction of their training time. In addition, thanks to the increase in computational efficiency, we are able to implement G-CNNs equivariant to the $\mathrm{Sim(2)}$ group; the group of dilations, rotations and translations. $\mathrm{Sim(2)}$-equivariance further improves performance on all tasks considered.

Axions can be copiously produced in localized regions of neutron star magnetospheres where the ambient plasma is unable to efficiently screen the induced electric field. As these axions stream away from the neutron star they can resonantly transition into photons, generating a large broadband contribution to the neutron star's intrinsic radio flux. In this work, we develop a comprehensive end-to-end framework to model this process from the initial production of axions to the final detection of radio photons, and derive constraints on the axion-photon coupling, $g_{a\gamma\gamma}$, using observations of 27 nearby pulsars. We study the modeling uncertainty in the sourced axion spectrum by comparing predictions from 2.5 dimensional particle-in-cell simulations with those derived using a semi-analytic model; these results show remarkable agreement, leading to constraints on the axion-photon coupling that typically differ by a factor of no more than $\sim 2$. The limits presented here are the strongest to date for axion masses $10^{-8} \, {\rm eV} \lesssim m_a \lesssim 10^{-5} \, {\rm eV}$, and crucially do not rely on the assumption that axions are dark matter.

The environmental impact of Artificial Intelligence (AI)-enabled systems is increasing rapidly, and software engineering plays a critical role in developing sustainable solutions. The "Greening AI with Software Engineering" CECAM-Lorentz workshop (no. 1358, 2025) funded by the Centre Européen de Calcul Atomique et Moléculaire and the Lorentz Center, provided an interdisciplinary forum for 29 participants, from practitioners to academics, to share knowledge, ideas, practices, and current results dedicated to advancing green software and AI research. The workshop was held February 3-7, 2025, in Lausanne, Switzerland. Through keynotes, flash talks, and collaborative discussions, participants identified and prioritized key challenges for the field. These included energy assessment and standardization, benchmarking practices, sustainability-aware architectures, runtime adaptation, empirical methodologies, and education. This report presents a research agenda emerging from the workshop, outlining open research directions and practical recommendations to guide the development of environmentally sustainable AI-enabled systems rooted in software engineering principles.

Active filaments are a workhorse for propulsion and actuation across biology, soft robotics and mechanical metamaterials. However, artificial active rods suffer from limited robustness and adaptivity because they rely on external control, or are tethered to a substrate. Here we bypass these constraints by demonstrating that non-reciprocal interactions lead to large-scale unidirectional dynamics in free-standing slender structures. By coupling the bending modes of a buckled beam anti-symmetrically, we transform the multistable dynamics of elastic snap-through into persistent cycles of shape change. In contrast to the critical point underpinning beam buckling, this transition to self-snapping is mediated by a critical exceptional point, at which bending modes simultaneously become unstable and degenerate. Upon environmental perturbation, our active filaments exploit self-snapping for a range of functionality including crawling, digging and walking. Our work advances critical exceptional physics as a guiding principle for programming instabilities into functional active materials.

We argue that the degrees of freedom in a d-dimensional CFT can be re-organized in an insightful way by studying observables on the moduli space of causal diamonds (or equivalently, the space of pairs of timelike separated points). This 2d-dimensional space naturally captures some of the fundamental nonlocality and causal structure inherent in the entanglement of CFT states. For any primary CFT operator, we construct an observable on this space, which is defined by smearing the associated one-point function over causal diamonds. Known examples of such quantities are the entanglement entropy of vacuum excitations and its higher spin generalizations. We show that in holographic CFTs, these observables are given by suitably defined integrals of dual bulk fields over the corresponding Ryu-Takayanagi minimal surfaces. Furthermore, we explain connections to the operator product expansion and the first law of entanglement entropy from this unifying point of view. We demonstrate that for small perturbations of the vacuum, our observables obey linear two-derivative equations of motion on the space of causal diamonds. In two dimensions, the latter is given by a product of two copies of a two-dimensional de Sitter space. For a class of universal states, we show that the entanglement entropy and its spin-three generalization obey nonlinear equations of motion with local interactions on this moduli space, which can be identified with Liouville and Toda equations, respectively. This suggests the possibility of extending the definition of our new observables beyond the linear level more generally and in such a way that they give rise to new dynamically interacting theories on the moduli space of causal diamonds. Various challenges one has to face in order to implement this idea are discussed.

From protein motifs to black holes, topological solitons are pervasive nonlinear excitations that are robust and can be driven by external fields. So far, existing driving mechanisms all accelerate solitons and antisolitons in opposite directions. Here we introduce a local driving mechanism for solitons that accelerates both solitons and antisolitons in the same direction instead: non-reciprocal driving. To realize this mechanism, we construct an active mechanical metamaterial consisting of non-reciprocally coupled oscillators subject to a bistable potential. We find that such nonlinearity coaxes non-reciprocal excitations - so-called non-Hermitian skin waves, which are typically unstable - into robust oneway (anti)solitons. We harness such non-reciprocal topological solitons by constructing an active waveguide capable of transmitting and filtering unidirectional information. Finally, we illustrate this mechanism in another class of metamaterials that displays the breaking of 'supersymmetry' causing only antisolitons to be driven. Our observations and models demonstrate a subtle interplay between non-reciprocity and topological solitons, whereby solitons create their own driving force by locally straining the material. Beyond the scope of our study, non-reciprocal solitons might provide an efficient driving mechanism for robotic locomotion and could emerge in other settings, e.g. quantum mechanics, optics and soft matter.

The ActivityNet Large Scale Activity Recognition Challenge 2017 Summary: results and challenge participants papers.

We derive the asymptotic distribution of the supremum distance of the deconvolution kernel density estimator to its expectation for certain supersmooth deconvolution problems. It turns out that the asymptotics are essentially different from the corresponding results for ordinary smooth deconvolution.

A well-known method to prepare ground states of fermionic many-body hamiltonians is adiabatic state preparation, in which an easy to prepare state is time-evolved towards an approximate ground state under a specific time-dependent hamiltonian. However, which path to take in the evolution is often unclear, and a direct linear interpolation, which is the most common method, may not be optimal. In this work, we explore new types of adiabatic paths based on the spectral decomposition of the two-body projection of the residual hamiltonian (the difference between the final and initial hamiltonian). The decomposition defines a set of hamiltonian terms which may be adiabatically interpolated in a piecewise or combined fashion. We demonstrate the usefulness of partially piecewise interpolation through examples involving Fermi-Hubbard models where, due to symmetries, level crossings occur in direct (fully combined) interpolation. We show that this specific deviation from a direct path appropriately breaks the relevant symmetries, thus avoiding level crossings and enabling an adiabatic passage. On the other hand, we show that a fully piecewise scheme, which interpolates every hamiltonian term separately, exhibits a worst-case complexity of $O(L^6/\Delta^3)$ as compared to $O(L^4/\Delta^3)$ for direct interpolation, in terms of the number of one-body modes $L$ and the minimal gap $\Delta$ along the path. This suboptimality result suggests that only those terms which break necessary symmetries should be taken into account for piecewise interpolation, while the rest is treated with direct interpolation.

In many scenarios -- when we bite food or during a crash -- fracture is inevitable. Finding solutions to steer fracture to mitigate its impact or turn it into a purposeful functionality, is therefore crucial. Strategies using composites, changes in chemical composition or crystal orientation, have proven to be very efficient, but the crack path control remains limited and has not been achieved in load-bearing structures. Here, we introduce fracture metamaterials consisting of slender elements whose bending enables large elastic deformation as fracture propagates. This interplay between bending and fracture enables tunable energy dissipation and the design of on-demand crack paths of arbitrary complexity. To this end, we use topology optimisation to create unit cells with anisotropic fracture energy, which we then tile up to realize fracture metamaterials with uniform density that we 3D-print. The thin ligaments that constitute the unit cells confer them a strikingly distinct response in tension and shear, and we show that by controlling the orientation and layout of the unit cells the sequential progress of the crack can be controlled, making the fracture path arbitrarily tortuous. This tortuosity increases the energy dissipation of the metamaterial without changing its stiffness. Using bespoke arrangements of unit cells, metamaterials can have on-demand fracture paths of arbitrary complexity. Our findings bring a new perspective on inelastic deformations in mechanical metamaterials, with potential applications in areas as diverse as the food industry, structural design, and for shock and impact damping.