arXiv subjects

Andrea Morandi

Publications and source records attributed to Andrea Morandi.

At least 19 recordsLinked to original sources

arXiv2026

Spectral Retrieval: Multi-Scale Sinc Convolution over Token Embeddings for Localized Retrieval in LLM Multi-Agent Systems

[Abridged] - Spectral Retrieval is a plug-in re-ranking stage that interpolates between per-token MaxSim and mean-pool retrieval through a multi-scale sinc convolution over token embeddings. In standard dense retrieval each document is one mean-pooled vector; when relevance localises into a short subspan, the signal averages into noise. Spectral Retrieval reuses per-token embeddings from a late-interaction index and convolves them with a normalised sinc kernel at multiple scales. At L=1 the kernel acts as the identity, recovering per-token MaxSim; as L grows it approaches a uniform filter, recovering mean pooling. The maximum cosine over positions and scales yields a score provably no less informative than either endpoint. On a controlled synthetic benchmark with 1,000 documents and planted single-position spikes, mean-pool retrieval sits at chance (Recall@10 ~ 0.02) regardless of spike strength, while Spectral Retrieval reaches Recall@10 = 1.0 once the planted cosine exceeds the corpus-level token noise floor. On LIMIT-small with a frozen all-mpnet-base-v2 encoder, Spectral Retrieval lifts Recall@10 from 0.33 to 0.90, MRR from 0.22 to 0.79, and strict Success@10 from 0.12 to 0.84, without retraining. The method fits naturally into multi-agent LLM systems, where each agent benefits from a tighter, role-specific retrieval window over a shared corpus.

cs.IR↗

arXiv2026

Sequential Consensus for Multi-Agent LLM Debates: A Wald-SPRT compute governor with calibration-based failure detection

Multi-agent LLM debate improves factuality and reasoning, but most recipes pick a fixed round count, over-spending on easy items and under-spending on hard ones. We adapt Wald's Sequential Probability Ratio Test (SPRT) as a plug-in compute governor for LLM debates. After each round, an LLM judge emits a [0,1] consensus score on the latest agent positions; a Wald monitor accumulates the log-likelihood ratio of "useful convergence" vs "not yet useful" under a Beta likelihood family, and stops when either boundary is crossed or returns a capped best-effort outcome at R_max. Under i.i.d. assumptions the rule inherits SPRT type-I/type-II error guarantees; in deployment the calibration itself is the more important object, since it estimates whether the judge score actually separates useful from unhelpful convergence in a given domain. We evaluate two tracks: (i) a Monte-Carlo study under calibrated Beta models characterising working curves, error rates, capping behaviour, and sensitivity; and (ii) a real-LLM evaluation on 200 attempted MMLU and 200 attempted GSM8K items with three heterogeneous agents (gpt-5, claude-opus-4-6, gemini-2.5-pro) and a claude-opus-4-6 judge, using disjoint 40-item calibration subsets. On GSM8K the rule stops in 1.01 average rounds (4.06 LLM calls) at 97.0% accuracy vs 99.0% for fixed-5 debate at 15 calls: a 3.7x call reduction at -2pp accuracy. On MMLU the calibrated KL collapses to about 0 and the rule caps on 99.5% of items at 2.1x cost. The takeaway is not that SPRT makes debate more accurate, but that a classical sequential test serves as a cheap compute-control and failure-detection layer for multi-agent LLM systems.

cs.LG↗

arXiv2026

Scale-Equivariant Generative Forecasting: Weight-Tied Dilated Convolutions, Wavelet Scattering Inputs, and Spectral-Consistency Training for Self-Similar Time Series

Many natural and engineered time series -- equity returns, climate anomalies, turbulent velocities, neural recordings, packet-level network traffic -- are approximately self-similar: their horizon-$T$ distribution is tied to the horizon-$1$ distribution by one scaling exponent $H$. Standard deep generative sequence models (transformers, dilated TCNs, the WaveNet family) ignore this. Their receptive fields are wide, but kernel parameters live independently at every dilation level, yielding a multi-scale architecture, not a scale-equivariant one. We make three contributions. First, we give a precise definition of discrete scale equivariance for 1D causal networks and prove that dyadic dilation commutes (up to boundary effects) with any dilated-convolution stack whose kernel weights are shared across levels. Tying the kernel shrinks the convolutional parameter budget by an $L$-fold factor (where $L$ is depth) and hard-wires self-similarity in as an inductive bias. Second, we wrap this Scale-Equivariant WaveNet (SE-WaveNet) backbone in three components that carry the same prior: a one-level Daubechies-4 wavelet input, a Hurst-FiLM block exposing the local scaling exponent, and a spectral-consistency training term targeting the $|f|^{-(2H+1)}$ power-law spectrum. The head is a conditional normalising flow, chosen to preserve equivariance. Third, on 30 years of S&P 500 daily log-returns, SE-WaveNet samples reproduce the empirical scaling-collapse diagnostic on the Allan-Variance top-25 universe (median $\mathcal{C}^\star = 0.020$), while a vanilla WaveNet at matched capacity does not ($\geq 0.06$). NLL, KS-calibration, and tail energy distance tie or beat the baseline, with $L\times$ fewer convolutional parameters.

cs.LG↗

arXiv2026

RTLC -- Research, Teach-to-Learn, Critique: A three-stage prompting paradigm inspired by the Feynman Learning Technique that lifts LLM-as-judge accuracy on JudgeBench with no fine-tuning

LLM-as-a-judge is now the default measurement instrument for open-ended generation, but on the public JudgeBench benchmark even strong instruction-tuned judges barely scrape past random on objective-correctness pairwise items. We introduce RTLC, a three-stage prompting recipe -- Research, Teach-to-Learn, Critique -- that promotes a single black-box LLM into an ensemble-of-thought judge with no fine-tuning, retrieval, or external tools. Stage 1 wraps the input in a fixed pedagogical scaffold porting the Feynman Learning Technique (study $\to$ teach $\to$ find gaps $\to$ simplify) into LLM prompting. Stage 2 draws N=10 independent candidate verdicts at temperature 0.4. Stage 3 acts as its own critic, cross-comparing the candidate set against the original question to emit one critiqued verdict at temperature 0. On JudgeBench-GPT (350 hard pairwise items), Claude 3.7 Sonnet's pairwise accuracy climbs from 64.6% (single-shot vanilla prompt) to 78.6% (RTLC critique-of-10) -- an absolute 14.0-percentage-point gain. RTLC also beats N=10 self-consistency majority voting (77.7%) and a zero-shot first candidate (74.0%). A clean three-step ablation attributes +9.4 pp to the Teach-to-Learn scaffold, +3.7 pp to N=10 marginalisation, and +0.9 pp to explicit critique. We discuss the cost-accuracy frontier (RTLC sits above self-consistency at every working point), the error-budget breakdown across the four JudgeBench categories (knowledge, reasoning, math, coding), and how RTLC composes orthogonally with post-hoc judge-score calibration, with the two interventions compounding multiplicatively in practice.

cs.CL↗

arXiv2026

Correcting Selection Bias in Sparse User Feedback for Large Language Model Quality Estimation: A Multi-Agent Hierarchical Bayesian Approach

[Abridged] Production LLM deployments receive feedback from a non-random fraction of users: thumbs sit mostly in the tails of the satisfaction distribution, and a naive average over them can land 40-50 percentage points away from true system quality. We treat this as a topic- and sentiment- stratified selection-bias problem and propose a three-agent hierarchical Bayesian pipeline that does not require ground-truth labels on individual interactions. A Topic Clustering Agent partitions the stream via UMAP + HDBSCAN over text embeddings; a Bias Modeling Agent fits a two-stage hierarchical Beta-Binomial under NUTS, inferring per-topic selection rates $s_c$ and quality $q_c$ with partial pooling; a Synthesis Agent reweights $q_c$ by true topic prevalence $\hat\pi_c = n_c/N$ to report a bias-corrected aggregate posterior $\bar Q = \sum_c \hat\pi_c q_c$ with credible interval, plus drift signals for online recalibration. Validation uses UltraFeedback (N=10,232 retained interactions, $C=18$ clusters, $Q^\star=0.6249$) with simulated topic- and sentiment-dependent selection biases. We compare five Bayesian variants against Naive and IPW baselines. A mild prior on the feedback channel (typical positive-feedback rate and negative-to-positive ratio, both readable from any production dashboard without labels) keeps Hierarchical-Informed within 4-13 pp of $Q^\star$ as the bias ratio sweeps from 1:1 to 30:1, with 95% credible intervals covering $Q^\star$ in 50/50 random-seed replicates at $\kappa_{\max}=10$. Without channel-side priors, every weak-prior variant misses $Q^\star$ by 22-33 pp: the per-cluster sufficient statistics admit a one-parameter family of equally good fits, and the prior on the bias channel (not on latent quality) is what breaks the degeneracy.

cs.CL↗

arXiv2026

Two Ways to De-Bias an LLM-as-a-Judge: A Continuous-Score Comparison of Hierarchical Bayesian Calibration and Neural-ODE Score Transport

[Abridged] Using a Large Language Model (LLM) as an automatic rater (LLM-as-a-judge) is cheap but potentially biased: some judges run lenient, others strict, the middle of the scale gets compressed, and verbose answers may be over-rewarded. A common remedy is post-hoc calibration: leave the cheap judge in place and, on a modest set of paired anchors, fit a transformation from raw judge scores to an estimate of the human rating. We compare two correctors that take opposing views on how this mapping should be modeled: a parametric, small-anchor hierarchical Bayesian linear correction with per-score uncertainty, and a non-parametric Neural-ODE (FFJORD) score-transport flow. Both are run head-to-head on UltraFeedback fine-grained_score (1700 paired examples, 200 held out), with calibration split into three operational sub-questions: population-mean recovery, per-item accuracy, and distributional-shape match. The headline result is that the choice between methods is primarily a data-budget question. Both correctors close the raw $+0.71$-point mean offset to within $\pm 0.08$ of the GPT-4 reference, at 100 and at 1500 anchors. Past that, the methods swap roles. With 100 anchors, the linear corrector reconstructs the human-score distribution roughly twice as well by KL divergence (0.031 vs. 0.058) and ties the flow on MAE. With 1500 anchors the flow wins on every metric (MAE 0.320 vs. 0.359, Pearson 0.922 vs. 0.896, KL 0.026 vs. 0.037). The Bayesian linear corrector saturates well below 1500 anchors: residual $\tanh$-shaped non-linearity is, by construction, structure a linear correction cannot fit. The flow keeps improving as labels grow. We translate these findings into an explicit decision rule for production deployments.

cs.CL↗

arXiv2024

Mesoscopic light transport in nonlinear disordered media

Nonlinear disordered media uniquely combine multiple scattering and second-harmonic generation. Here, we investigate the statistical properties of the nonlinear light generated within such media. We report super-Rayleigh statistics of the second-harmonic speckle intensity, and demonstrate that it is caused by the mesoscopic correlations arising in extreme scattering conditions. The measured conductance is the lowest ever observed in an isotropically scattering 3D medium, with applications in broadband second-harmonic generation, wavefront shaping in nonlinear disordered media, and photonic computing.

physics.optics↗

arXiv2023

Large-scale photonic computing with nonlinear disordered media

Neural networks find widespread use in scientific and technological applications, yet their implementations in conventional computers have encountered bottlenecks due to ever-expanding computational needs. Photonic neuromorphic hardware, which manipulates information and represents data continuously in the optical domain, is one of the promising platforms with potential advantages of massive parallelism, ultralow latency, and reduced energy consumption. While linear photonic neural networks are within reach, photonic computing with large-scale optical nonlinear nodes remains largely unexplored. Here, we demonstrate a large-scale, high-performance nonlinear photonic neural system based on a disordered polycrystalline slab composed of lithium niobate nanocrystals. Mediated by random quasi-phase-matching and multiple scattering, linear and nonlinear optical speckle features are generated as the interplay between the simultaneous linear random scattering and the second-harmonic generation, defining a complex neural network in which the second-order nonlinearity acts as internal nonlinear activation functions. Benchmarked against linear random projection, such nonlinear mapping embedded with rich physical computational operations shows improved performance across a large collection of machine learning tasks in image classification, regression, and graph classification with varying complexity. Demonstrating up to 27,648 input and 3,500 nonlinear output nodes, the combination of optical nonlinearity and random scattering serves as a scalable computing engine for diverse applications.

physics.optics↗

arXiv2023

Graph model for multiple scattering in lithium niobate on insulator integrated photonic networks

We present a graph-based model for multiple scattering of light in integrated lithium niobate on insulator (LNOI) networks, which describes an open network of single-mode integrated waveguides with tunable scattering at the network nodes. We first validate the model at small scale with experimental LNOI resonator devices and show consistent agreement between simulated and measured spectral data. Then, the model is used to demonstrate a novel platform for on-chip multiple scattering in large-scale optical networks up to few hundred nodes, with tunable scattering behaviour and tailored disorder. Combining our simple graph-based model with material properties of LNOI, this platform creates new opportunities to control randomness in large optical networks.

physics.optics↗

arXiv2021

Broadband photon pair generation from a single lithium niobate microcube

Nonclassical light sources are highly sought after as they are an integral part of quantum communication and quantum computation devices. Typical sources rely on bulk crystals that are not compact and have limited bandwidth due to phase-matching conditions. In this work, we demonstrate the generation of photon pairs from a free-standing lithium niobate microcube at the telecommunication wavelength through the spontaneous parametric down-conversion process. The maximum photon pair generation rate obtained from a single microcube with the size of ~4 microns is ~80 Hz, resulting in an efficiency of ~1.2 GHz/Wm per unit volume, which is an order of magnitude higher than the efficiency of photon-pair generation in bulky nonlinear crystals. The microcubes are synthesized through a solvothermal method, offering the possibility for scalable devices via bottom-up assembly. Our work constitutes an important step forward in the realization of compact nonclassical light sources with broadband tunability for various applications in quantum communication, quantum computing, and quantum metrology.

quant-ph↗

arXiv2020

Modeling of Random Quasi-Phase-Matching in Birefringent Disordered Media

We provide a vectorial model to simulate second-harmonic generation (SHG) in birefringent, transparent media with an arbitrary configuration of non-linear ($χ^{(2)}$) crystalline grains. We apply this model on disordered assemblies of LiNbO$_3$ and BaTiO$_3$ grains to identify the influence of the birefringence on the random quasi-phase-matching process. We show that in monodispersed assemblies, the birefringence relaxes the grain-size dependence of the SHG efficiency. In polydispersed assemblies with sufficiently large grains, we find that the birefringence introduces an SHG efficiency enhancement of up to 54% compared to isotropic reference crystals, which is grain size independent. This enhancement increases linearly with the grain size, if the birefringent grains can be phase matched. These two different scaling behaviours are used in Kurtz and Perry's powder-technique to identify the phase-matchability of a material. We show on the example of LiNbO$_3$ and ADP that this technique cannot be applied when the grains get smaller than the coherence length, because the SHG scaling with the grain size becomes material specific.

physics.optics↗

arXiv2020

Broadband Mie-driven random quasi-phase-matching

High-quality crystals without inversion symmetry are the conventional platform to achieve optical frequency conversion via three wave-mixing. In bulk crystals, efficient wave-mixing relies on phase-matching configurations, while at the micro- and nano-scale it requires resonant mechanisms that enhance the nonlinear light-matter interaction. These strategies commonly result in wavelength-specific performances and narrowband applications. Disordered photonic materials, made up of a random assembly of optical nonlinear crystals, enable a broadband tunability in the random quasi-phase-matching (RQPM) regime and do not require high-quality materials. Here, we combine resonances and disorder by implementing RQPM in Mie-resonant spheres of a few microns realized by the bottom-up assembly of barium titanate nano-crystals. The measured second harmonic generation (SHG) reveals a combination of broadband and resonant wave mixing, in which Mie resonances drive and enhance the SHG, while the disorder keeps the phase-matching conditions relaxed. This new phase-matching regime can be described by a random walk in the SHG complex plane whose step lengths depend on the local field enhancement within the micro-sphere. Our nano-crystals assemblies provide new opportunities for tailored phase-matching at the micro-scale, beyond the coherence length of the bulk crystal. They can be adapted to achieve frequency conversion from the near-ultraviolet to the infrared ranges, they are low-cost and scalable to large surface areas.

physics.optics↗

arXiv2019

Local tuning of WS2 photoluminescence using polymeric micro-actuators in a monolithic van der Waals heterostructure

The control of the local strain profile in 2D materials offers an invaluable tool for tailoring the electronic and photonic properties of solid-state devices. In this paper, we demonstrate a local engineering of the exciton photoluminescence (PL) energy of monolayer tungsten disulfide (WS2) by means of strain. We apply a local uniaxial stress to WS2 by exploiting electron-beam patterned and actuated polymeric micrometric artificial muscles (MAMs), which we implement onto monolithic synthetic WS2/graphene heterostructures. We show that MAMs are able to induce an in-plane stress to the top WS2 layer of the van der Waals heterostructure and that the latter can slide on the graphene underneath with negligible friction. As a proof of concept for the local strain-induced PL shift experiments, we exploit a two-MAM configuration in order to apply uniaxial tensile stress on well-defined micrometric regions of WS2. Remarkably, our architecture does not require the adoption of fragile suspended microstructures. We observe a spatial modulation of the excitonic PL energy of the WS2 monolayers under stress, which agrees with the expected strain profile and attains a maximum redshift of about 40 meV at the maximum strain intensity point. After the actuation, a time-dependent PL blueshift is observed in agreement with the viscoelastic properties of the polymeric MAMs. Our approach enables inducing local and arbitrary deformation profiles and circumvents some key limitations and technical challenges of alternative strain engineering methods requiring the 2D material transfer and production of suspended membranes.

cond-mat.mes-hall↗

arXiv2018

Modified Gravity (MOG) and its test on galaxy clusters

The MOdified Gravity (MOG) theory of J. Moffat assumes a massive vector particle which causes a repulsive contribution to the tensor gravitation. For the galaxy cluster A1689 new data for the X-ray gas and the strong lensing properties are presented. Fits to MOG are possible by adjusting the galaxy density profile. However, this appears to work as an effective dark matter component, posing a serious problem for MOG. New gas and strong lensing data for the cluster A1835 support these conclusions and point at a tendency of the gas-alone to overestimate the lensing effects in MOG theory.

astro-ph.CO↗

arXiv2017

The galaxy group NGC2563

We present a $Chandra$ study of the hot intragroup medium (hIGM) of the galaxy group NCG2563. The $Chandra$ mosaic observations, with a total exposure time of ~430 ks, allow the gas density to be detected beyond $R_{200}$ and the gas temperature out to 0.75 $R_{200}$. This represents the first observational measurement of the physical properties of a poor groups beyond $R_{500}$. By capitalizing on the exquisite spatial resolution of $Chandra$ that is capable to remove unrelated emission from point sources and substructures, we are able to radially constrain the inhomogeneities of gas ("clumpiness"), gas fraction, temperature and entropy distribution. Although there is some uncertainty in the measurements, we find evidences of gas clumping in the virialization region, with clumping factor of about 2 - 3 at $R_{200}$. The gas clumping-corrected gas fraction is significantly lower than the cosmological baryon budget. These results may indicate a larger impact of the gas inhomogeneities with respect to the prediction from hydrodynamic numerical simulations, and we discuss possible explanations for our findings.

astro-ph.CO↗

arXiv2016

A Strong Merger Shock in Abell 665

Deep (103 ks) \chandra\ observations of Abell 665 have revealed rich structures in this merging galaxy cluster, including a strong shock and two cold fronts. The newly discovered shock has a Mach number of $M$ = 3.0 $\pm$ 0.6, propagating in front of a cold disrupted cloud. This makes Abell~665 the second cluster where a strong merger shock of $M \approx$ 3 has been detected, after the Bullet cluster. The shock velocity from jump conditions is consistent with (2.7 $\pm$ 0.7) $\times$ 10$^3$ km sec$^{-1}$. The new data also reveal a prominent southern cold front, with potentially heated gas ahead of it. Abell 665 also hosts a giant radio halo. There is a hint of diffuse radio emission extending to the shock at the north, which needs to be examined with better radio data. This new strong shock provides a great opportunity to study the re-acceleration model with the X-ray and radio data combined.

astro-ph.HE↗

arXiv2016

Probing dark energy via galaxy cluster outskirts

We present a Bayesian approach to combine $Planck$ data and the X-ray physical properties of the intracluster medium in the virialization region of a sample of 320 galaxy clusters ($0.056 3$ keV) observed with $Chandra$. We exploited the high-level of similarity of the emission measure in the cluster outskirts as cosmology proxy. The cosmological parameters are thus constrained assuming that the emission measure profiles at different redshift are weakly self-similar, that is their shape is universal, explicitly allowing for temperature and redshift dependence of the gas fraction. This cosmological test, in combination with $Planck$+SNIa data, allows us to put a tight constraint on the dark energy models. For a constant-$w$ model, we have $w=-1.010\pm0.030$ and $Ω_m=0.311\pm0.014$, while for a time-evolving equation of state of dark energy $w(z)$ we have $Ω_m=0.308\pm 0.017$, $w_0=-0.993\pm0.046$ and $w_a=-0.123\pm0.400$. Constraints on the cosmology are further improved by adding priors on the gas fraction evolution from hydrodynamic simulations. Current data favour the cosmological constant with $w\equiv-1$, with no evidence for dynamic dark energy. We checked that our method is robust towards different sources of systematics, including background modelling, outlier measurements, selection effects, inhomogeneities of the gas distribution and cosmic filaments. We also provided for the first time constraints on which definition of cluster boundary radius is more tenable, namely based on a fixed overdensity with respect to the critical density of the Universe. This novel cosmological test has the capacity to provide a generational leap forward in our understanding of the equation of state of dark energy.

astro-ph.CO↗

arXiv2016

Shocking Features in the Merging Galaxy Cluster RXJ0334.2-0111

We present a 66 ksec $\textit{Chandra}$ X-ray observation of the galaxy cluster RXJ0334.2-0111. This deep observation revealed a unique bow shock system associated with a wide angle tail (WAT) radio galaxy and several intriguing substructures. The temperature across the bow shock jumps by a factor of $\sim$ 1.5 (from 4.1 keV to 6.2 keV), and is consistent with the Mach number $M = 1.6_{-0.3}^{+0.5}$. A second inner surface brightness edge is a cold front that marks the border between infalling subcluster cool core and the ICM of the main cluster. The temperature across the cold front increases from $1.3_{-0.8}^{+0.3}$ keV to $6.2_{-0.6}^{+0.6}$ keV. We find an overpressurized region $\sim$ 250 kpc east of the cold front that is named "the eastern extension (EE)". The EE may be a part of the third subcluster in the ongoing merger. We also find a tail shaped feature that originates near the bow shock and may extend up to a distance of $\sim$ 1 Mpc. This feature is also likely overpressurized. The luminous FR-I radio galaxy, 3C89, appears to be the cD galaxy of the infalling subcluster. We estimated 3C89's jet power from jet bending and the possible interaction between the X-ray gas and the radio lobes. A comparison between the shock stand-off distance and the Mach number for all known shock front/cold front combinations suggests that the core is continuously shrinking in size by stripping.

astro-ph.HE↗