arXiv subjects

Niccolò Caselli

Publications and source records attributed to Niccolò Caselli.

10 recordsLinked to original sources

arXiv2026

Mind the Gap: Promises and Pitfalls of Hierarchical Planning in LeWorldModel

We investigate whether temporal hierarchy can improve LeWorldModel on long-horizon goal-conditioned control. We introduce Hi-LeWM, an extension that freezes the pretrained low-level LeWM and adds high-level planning over latent subgoals. We evaluate Hi-LeWM on PushT and Cube across increasing goal offsets. Hierarchy does not automatically improve performance: at short horizons, the best configuration uses a one-step high-level horizon, while longer horizons reveal a mismatch between the learned high-level action space and the inference-time search distribution. Experiments with true future latent subgoals show that the frozen low-level controller can execute well-aligned intermediate targets, indicating that high-level subgoal generation is the main bottleneck. Unconstrained search can select latent macro-actions that appear favorable under the learned model but produce poor control targets. Constraining search around macro-actions encoded from training trajectories, with appropriate subgoal execution timing, recovers useful hierarchical regimes, improving over flat LeWM by +11.3 percentage points at medium-range horizons and +14.7 percentage points at the longest PushT horizon. Overall, temporal abstraction can benefit compact frozen LeWM, but only when high-level search remains compatible with the low-level controller

cs.RO↗

arXiv2026

CAPRA: Scaling Feedback on Software Architecture Deliverables with a Multi-Agent LLM System

Automated assessment in software engineering education has advanced significantly for code grading and essay scoring. However, reviewing software architecture deliverables, which requires analyzing structural completeness and requirements traceability, has not yet been fully automated. Applying Large Language Models (LLMs) to this task requires robust architectures to ensure technical feedback is accurate and reliable for students. This paper presents CAPRA (Configurable Architecture Proficiency Report Assessment), a multi-agent LLM system that analyzes software architecture deliverables to generate personalized, template-compliant LaTeX feedback. As a core design choice, CAPRA coordinates multiple specialized agents and employs a Python-based microservice for multi-modal document extraction, utilizing PyMuPDF and vision-enabled LLMs (specifically gpt-4o) to parse text and UML diagrams. To ensure educational reliability and mitigate hallucinations, CAPRA introduces a deterministic Evidence Anchoring step using fuzzy matching via normalized Levenshtein distance, along with a ConsistencyManager agent that cross-verifies, deduplicates, and merges findings. System performance is assessed using a structured eight-criterion binary evaluation taxonomy covering: (i) extraction completeness, (ii) feature validation, (iii) issue grounding and severity detection, (iv) recommendation specificity and traceability, and (v) template and tone compliance. A preliminary empirical evaluation on 10 student reports shows that CAPRA satisfied 88.8% of the evaluated criteria under a strict two-rater aggregation rule, achieved moderate inter-rater agreement with human evaluators (kappa = 0.582), and processed each report in slightly over 4 minutes. While these results support the viability of LLM-supported architectural feedback, human oversight remains essential for subjective assessment dimensions.

cs.SE↗

arXiv2026

Do Video Foundation Models Understand Intuitive Physics? A Layerwise Probing Analysis

We study whether pretrained video foundation models encode intuitive-physics information in their frozen representations, and how this information varies across model families, layers, and probe types. Using frozen-feature probing on IntPhys2 and Minimal Video Pairs (MVP), we compare predictive joint-embedding models (V-JEPA), masked reconstruction models (VideoMAE), and a diffusion-based video generator (LTX-Video). V-JEPA achieves the strongest overall results across benchmarks, especially with probes that model temporal dynamics, while VideoMAE remains competitive and LTX-Video recovers weaker but non-trivial signal. Layerwise analyses show that physics-relevant information is weakest in early layers and becomes most accessible at intermediate-to-late depth, and temporal controls show that disrupting frame order substantially reduces performance, especially on MVP. Together, these results suggest that intuitive-physics knowledge emerges reliably in pretrained video representations, but its accessibility depends strongly on pretraining paradigm, representational depth, and readout mechanism.

cs.CV↗

arXiv2025

Active Force Dynamics in Red Blood Cells Under Non-Invasive Optical Tweezers

Red blood cells (RBCs) sustain mechanical stresses associated with microcirculatory flow through ATP-driven plasma membrane flickering. This is an active phenomenon driven by motor proteins that regulate interactions between the spectrin cytoskeleton and the lipid bilayer; it is manifested in RBC shape fluctuations reflecting the cell's mechanical and metabolic state. Yet, direct quantification of the forces and energetic costs underlying this non-equilibrium behavior remains challenging due to the invasiveness of existing techniques. Here, a minimally invasive method that combines bead-free, low-power optical tweezers with high-speed video microscopy was employed to track local membrane forces and displacements in single RBCs during the same time window. This independent dual-channel measurement enabled the construction of a mechano-dynamic phase space for RBCs under different chemical treatments, that allowed for differentiating between metabolic and structural states based on their fluctuation-force signatures. Quantification of mechanical work during flickering demonstrated that membrane softening enhanced fluctuations while elevating energy dissipation. The proposed optical tweezers methodology provides a robust framework for mapping the active mechanics of living cells, enabling precise probing of cellular physiology and detection of biomechanical dysfunction in diseases.

physics.bio-ph↗

arXiv2024

Collective lattice excitations in the dynamic route for melting hydrodynamic 2D-crystals

Surface stiffnesses engender steady patterns of Faraday waves (FWs), so called hydrodynamic crystals as correspond to ordered wave lattices made of discrete subharmonics under monochromatic driving. Mastering rules are both inertia-imposed parametric resonance for frequency-halving together with rigidity-driven nonlinearity for wavefield self-focusing. They harness the discretization needed for coherent FW-packets to localize in space and time. Collective lattice excitations are observed as dispersionless propagating dislocations that lead periodic modulations arising from explicit symmetry breaking. In a field theory perspective, a halving genesis for the collective distorting modes is revealed as the natural pathway for hydrodynamic crystal melting.

nlin.PS↗

arXiv2023

Optical control of spatially localized red blood cell activity by holographic tweezing

Red blood cells possess unique biomechanical ability to squeeze through capillaries smaller than their size to enable gas and ion exchange. A key signature of their active biomechanics is the out-of-equilibrium fluctuation of the plasma membrane, also known as flickering motion. This active flickering is driven by motor proteins that connect the forces between the spectrin skeleton and the lipid bilayer. However, studying flickering motions in living red blood cells is challenging without altering their physical properties. Here, we implemented a holographic optical tweezer that sculpted a laser beam to create a force field distributed directly along the membrane equatorial contour. We show heterogeneous membrane flickering activity driven by membrane kickers in free-standing cells. Then we inhibited the active kickers by optical forces under minimal invasion, thus benchmarking the active motion against thermal fluctuations. Our work paves the way for optical control of biophysical forces, providing touchless strategies for mechanotransduction in living cells.

physics.bio-ph↗

arXiv2021

Electrically driven random lasing from a modified Fabry-Perot laser diode

Random lasers (RLs) are intriguing devices with promising applications as light sources for imaging, sensing, super resolution spectral analysis or complex networks engineering. RLs can be obtained from optically pumped dyes, optical fibers and crystals, or electrically pumped semiconductor heterostructures. Semicon-ductor RLs are usually fabricated by introducing scattering defects into the active layer, adding a degree of complexity to the fabrication process and losing the ease of realization potentially offered by disordered structures. Ready availability of electrically pumped RLs, avoiding costly fabrication approach, would boost the use of these devices in research and applications. Here, we realize an incoherent semiconductor RL by simply processing the output mirror of an off-the-shelf Fabry-Perot laser diode via controlled laser ablation. Optical feedback provided by the intact back mirror and the ablated front mirror results in multi-mode ran-dom lasing with low spatial coherence and speckled output emission profile.

physics.optics↗

arXiv2020

End-coupled random lasers: a basis for artificial neural networks

Light interference in strongly disordered photonic media can generate lasers where random modes are amplified in unpredictable way. The ease of fabrication, along with their low coherence caused by multiple small-linewidth peaks, made random lasers (RL) emerging, efficient, speckle-free light sources and a means to achieve spectral super-resolution. With potential to become a mature and accessible technology, their complex system's nature furnishes endless opportunities to unveil fundamental physics, since they can act as elements of optical network architectures. To date no experimental studies have analyzed the optical interaction between independent resonators in networks of RLs. Realizing RLs with a pumped strip joining two rough mirrors, we experimentally investigate and numerically simulate the emergence of networks when, by sharing scattering centers, RLs become coupled. We prove that the emission of a single RL can be manipulated by the action of others in the network, giving rise to substantial peak rearrangements and energy redistribution, fingerprint of mode coupling. Our findings, involving a few coupled RLs, firmly set the basis for the study of full-grown photonic networks. Oddly, both their deep understanding and their deployment as hardware clearly point in the direction of a novel disruptive technology: artificial intelligence on photonic random neural networks.

physics.optics↗

arXiv2020

Moulding hydrodynamic 2D-crystals upon parametric Faraday waves in shear-functionalized water surfaces

Faraday waves (FWs), or surface waves oscillating at half of the natural frequency when a liquid is vertically vibrated, are archetypes of ordering transitions on liquid surfaces. The existence of unbounded FW-patterns sustained upon bulk frictional stresses has been evidenced in highly viscous fluids. However, the role of surface rigidity has not been investigated so far. Here, we demonstrate that dynamically frozen FWs that we call 2D-hydrodynamic crystals do appear as ordered patterns of nonlinear surface modes in water surfaces functionalized with soluble (bio)surfactants endowing in-plane shear stiffness. The strong phase coherence in conjunction with the increased surface rigidity bear the FW-ordering transition, upon which the hydrodynamic crystals were reversibly molded by parametric control of their degree of order. Crystal symmetry and unit cell size were tuned depending on the FW-dispersion regime. The hydrodynamic crystals here discovered could be exploited in touchless strategies of soft matter scaffolding. Particularly, the surface-directed synthesis of structured materials based on colloids or polymers and cell culture patterns for tissue engineering could be ameliorated under external control of FW-coherence

cond-mat.soft↗

arXiv2015

Tailoring the photon hopping by nearest and next-nearest-neighbour interaction in photonic arrays

Arrays of photonic cavities are relevant structures for developing large-scale photonic integrated circuits and for investigating basic quantum electrodynamics phenomena, due to the photon hopping between interacting nanoresonators. Here, we investigate, by means of scanning near-field spectroscopy, numerical calculations and an analytical model, the role of different neighboring interactions that give rise to delocalized supermodes in different photonic crystal array configurations. The systems under investigation consist of three nominally identical two-dimensional photonic crystal nanocavities on membrane aligned along the two symmetry axes of the triangular photonic crystal lattice. We find that the nearest and next-nearest-neighbour coupling terms can be of the same relevance. In this case, a non-intuitive picture describes the resonant modes, and the photon hopping between adjacent nano-resonators is strongly affected. Our findings prove that exotic configurations and even post-fabrication engineering of coupled nanoresonators could directly tailor the mode spatial distribution and the group velocity in coupled resonator optical waveguides.

physics.optics↗