arXiv subjects

Michael Hinczewski

Publications and source records attributed to Michael Hinczewski.

At least 19 recordsLinked to original sources

arXiv2026

Test-Time Scaling in Reasoning LLMs: Inference Regimes, Evaluation, and Reproducibility

Large language models can solve harder reasoning problems with more inference-time compute. The term "test-time scaling," however, covers several inference algorithms: extending deliberation along one trajectory, sampling completed candidates and aggregating them by voting or verification, and searching over partial states. These algorithms differ in statistical structure, compute requirements, and failure modes. Treating them as interchangeable under a scalar "budget," or reporting accuracy without specifying the inference protocol, makes results difficult to compare across studies. We study test-time scaling along three axes. First, we formalize it as budgeted inference over the implicit prefix tree of an autoregressive model and distinguish single-trajectory sequential scaling, leaf-level scaling with terminal reduction, and prefix-level scaling. Second, we treat the full inference system as the evaluated object and separate end-to-end performance from candidate-bank diagnostics. We introduce an evaluation profile whose coordinates and simple functionals recover or bound common repeated-sampling metrics, and require compute accounting and uncertainty estimates that match the protocol. Third, we distinguish exact replay from distributional reproducibility and state the requirements for each. We also organize open-weight reasoning models by model-side and interface mechanisms. Our empirical study covers broad knowledge, symbolic reasoning, and competition mathematics, and we publicly release 1,403,520 sampled model attempts. The project website is available at https://mohsenhariri.github.io/scorio/tts. The released datasets are Trace (https://huggingface.co/datasets/harimo/scorio-trace), Lite (https://huggingface.co/datasets/harimo/scorio-lite), Math (https://huggingface.co/buckets/harimo/scorio-math), and SuperGPQA (https://huggingface.co/buckets/harimo/scorio-gpqa).

cs.LG↗

arXiv2026

Physical Approaches to Metabolic Scaling in Living Systems

Living systems continuously transform matter and energy through the chemical processes that constitute their metabolism. The overall metabolic rate of an organism correlates positively with its body mass, however both the exact scaling behavior and possible explanations for this behavior have been under intense debate for two centuries. This review synthesizes empirical findings and theoretical frameworks on the energetics of living systems from an interdisciplinary perspective, with a focus on physical concepts. A general thermodynamic framework to study metabolism is laid out, together with a coarse-grained description of metabolic biochemistry. The rich history of experimental work in this field is surveyed, revealing a variety of metabolic scaling patterns at different levels of biological organization, from individual cells to whole populations. Several biophysical models proposed to explain the sublinear scaling of metabolic rate with body mass are summarized. Though the traditional focus has been on adult organisms, the review also highlights recent advances that probe metabolism during development. Improvements in experimental techniques, extensive datasets, and a host of open questions, suggest the field will continue gaining momentum in the near term. The review concludes with an outlook for this future progress: an interdisciplinary approach to elucidate metabolic scaling across different developmental stages and organism sizes.

cond-mat.soft↗

arXiv2026

Scorio.jl: A Julia package for ranking stochastic responses

Scorio.jl is a Julia package for evaluating and ranking systems from repeated responses to shared tasks. It provides a common tensor-based interface for direct score-based, pairwise, psychometric, voting, graph, and listwise methods, so the same benchmark can be analyzed under multiple ranking assumptions. We describe the package design, position it relative to existing Julia tools, and report pilot experiments on synthetic rank recovery, stability under limited trials, and runtime scaling.

cs.MS↗

arXiv2026

Ranking Reasoning LLMs under Test-Time Scaling

Test-time scaling evaluates reasoning LLMs by sampling multiple outputs per prompt, but ranking models in this regime remains underexplored. We formalize dense benchmark ranking under test-time scaling and introduce Scorio, a library that implements statistical ranking methods such as paired-comparison models, item response theory (IRT) models, voting rules, and graph- and spectral-based methods. Across $20$ reasoning models on four Olympiad-style math benchmarks (AIME'24, AIME'25, HMMT'25, and BrUMO'25; up to $N=80$ trials), most full-trial rankings agree closely with the Bayesian gold standard $\mathrm{Bayes}_{\mathcal{U}}@80$ (mean Kendall's $\tau_b = 0.93$--$0.95$), and $19$--$34$ methods recover exactly the same ordering. In the single-trial regime, the best methods reach $\tau_b \approx 0.86$. Using greedy decoding as an empirical prior ($\mathrm{Bayes}_{\mathbf{R}_0}@N$) reduces variance at $N=1$ by $16$--$52\%$, but can bias rankings when greedy and stochastic sampling disagree. These results identify reliable ranking methods for both high- and low-budget test-time scaling. We release Scorio as an open-source library at https://github.com/mohsenhariri/scorio.

cs.LG↗

arXiv2026

Observation of Light-Driven Levitation Near Epsilon-Near-Zero Surfaces

Optical manipulation of micro- and nanoparticles near surfaces is fundamental for applications in sensing and microfluidics, yet controlling particle-surface interactions remains challenging. Here we experimentally investigate light-induced forces on dielectric particles near epsilon-near-zero (ENZ) metamaterial surfaces using photonic force microscopy. By illuminating trapped particles with tunable visible light, we observe a wavelength-dependent repulsive force unique to ENZ surfaces, contrasting with the attractive forces near dielectric or metallic substrates. This repulsion peaks near the ENZ frequency and may be attributed to combined optical ENZ effects and thermophoretic forces. Our findings demonstrate that ENZ metamaterials can induce stable levitation of particles via light-driven forces, offering a novel mechanism for contactless manipulation in microfluidic environments. This work advances understanding of light-matter interactions at ENZ interfaces and suggests potential for ENZ-based optical control of micro- and nanoscale objects, with potential applications in micro- and nanofluidic environments.

physics.optics↗

arXiv2025

Thermodynamic Performance Limits for Score-Based Diffusion Models

We establish a fundamental connection between score-based diffusion models and non-equilibrium thermodynamics by deriving performance limits based on entropy rates. Our main theoretical contribution is a lower bound on the negative log-likelihood of the data that relates model performance to entropy rates of diffusion processes. We numerically validate this bound on a synthetic dataset and investigate its tightness. By building a bridge to entropy rates - system, intrinsic, and exchange entropy - we provide new insights into the thermodynamic operation of these models, drawing parallels to Maxwell's demon and implications for thermodynamic computing hardware. Our framework connects generative modeling performance to fundamental physical principles through stochastic thermodynamics.

cs.LG↗

arXiv2025

Don't Pass@k: A Bayesian Framework for Large Language Model Evaluation

Pass$@k$ is widely used to report the reasoning performance of LLMs, but it often produces unstable and potentially misleading rankings, especially when the number of trials (samples) is limited and computational resources are constrained. We present a principled Bayesian evaluation framework that replaces Pass$@k$ and average accuracy over $N$ trials (avg$@N$) with posterior estimates of a model's underlying success probability and credible intervals, yielding stable rankings and a transparent decision rule for differences. Evaluation outcomes are modeled as categorical (not just 0/1) with a Dirichlet prior, giving closed-form expressions for the posterior mean and uncertainty of any weighted rubric and enabling the use of prior evidence when appropriate. Theoretically, under a uniform prior, the Bayesian posterior mean is order-equivalent to average accuracy (Pass$@1$), explaining its empirical robustness while adding principled uncertainty. Empirically, in simulations with known ground-truth success rates and on AIME'24/'25, HMMT'25, and BrUMO'25, the posterior-based procedure achieves faster convergence and greater rank stability than Pass$@k$ and recent variants, enabling reliable comparisons at far smaller sample counts. The framework clarifies when observed gaps are statistically meaningful (non-overlapping credible intervals) versus noise, and it naturally extends to graded, rubric-based evaluations. Together, these results recommend replacing Pass$@k$ for LLM evaluation and ranking with a posterior-based, compute-efficient protocol that unifies binary and non-binary evaluation while making uncertainty explicit. Source code is available at https://github.com/mohsenhariri/scorio

cs.AI↗

arXiv2025

PATCH: a deep learning method to assess heterogeneity of artistic practice in historical paintings

The history of art has seen significant shifts in the manner in which artworks are created, making understanding of creative processes a central question in technical art history. In the Renaissance and Early Modern period, paintings were largely produced by master painters directing workshops of apprentices who often contributed to projects. The masters varied significantly in artistic and managerial styles, meaning different combinations of artists and implements might be seen both between masters and within workshops or even individual canvases. Information on how different workshops were managed and the processes by which artworks were created remains elusive. Machine learning methods have potential to unearth new information about artists' creative processes by extending the analysis of brushwork to a microscopic scale. Analysis of workshop paintings, however, presents a challenge in that documentation of the artists and materials involved is sparse, meaning external examples are not available to train networks to recognize their contributions. Here we present a novel machine learning approach we call pairwise assignment training for classifying heterogeneity (PATCH) that is capable of identifying individual artistic practice regimes with no external training data, or "ground truth." The method achieves unsupervised results by supervised means, and outperforms both simple statistical procedures and unsupervised machine learning methods. We apply this method to two historical paintings by the Spanish Renaissance master, El Greco: The Baptism of Christ and Christ on the Cross with Landscape, and our findings regarding the former potentially challenge previous work that has assigned the painting to workshop members. Further, the results of our analyses create a measure of heterogeneity of artistic practice that can be used to characterize artworks across time and space.

cs.CV↗

arXiv2024

Asymmetric Interactions Shape Survival During Population Range Expansions

An organism that is newly introduced into an existing population has a survival probability that is dependent on both the population density of its environment and the competition it experiences with the members of that population. Expanding populations naturally form regions of high and low density, and simultaneously experience ecological interactions both internally and at the boundary of their range. For this reason, systems of expanding populations are ideal for studying the combination of density and ecological effects. Conservation ecologists have been studying the ability of an invasive species to establish for some time, attributing success to both ecological and spatial factors. Similar behaviors have been observed in spatially structured cell populations, such as those found in cancerous tumors and bacterial biofilms. In these scenarios, novel organisms may be the introduction of a new mutation or bacterial species with some form of drug resistance, leading to the possibility of treatment failure. In order to gain insight into the relationship between population density and ecological interactions, we study an expanding population of interacting wild-type cells and mutant cells. We simulate these interactions in time and study the spatially dependent probability for a mutant to survive or to take over the front of the population wave (gene surfing). Additionally, we develop a mathematical model that describes this survival probability and find agreement when the payoff for the mutant is positive (corresponding to cooperation, exploitation, or commensalism). By knowing the types of interactions, our model provides insight into the spatial distribution of survival probability. Conversely, given a spatial distribution of survival probabilities, our model provides insight into the types of interactions that were involved to generate it.

q-bio.PE↗

arXiv2024

Extremal Kerr Black Hole Dark Matter from Hawking Evaporation

The Hawking process results in a monotonic decrease of the black hole mass, but a biased random walk of the black hole angular momentum. We demonstrate that this stochastic process leads to a significant fraction of primordial black holes becoming extremal Kerr black holes (EKBHs) of one to a few Planck masses regardless of their initial mass. For these EKBHs, the probability of ever absorbing a photon or other particle from the cosmic environment is small, even in the cores of galaxies. Assuming that EKBHs are stable, they behave as cold dark matter, and can comprise all of the dark matter if they are formed with the correct initial abundance.

gr-qc↗

arXiv2023

Machine learning in and out of equilibrium

The algorithms used to train neural networks, like stochastic gradient descent (SGD), have close parallels to natural processes that navigate a high-dimensional parameter space -- for example protein folding or evolution. Our study uses a Fokker-Planck approach, adapted from statistical physics, to explore these parallels in a single, unified framework. We focus in particular on the stationary state of the system in the long-time limit, which in conventional SGD is out of equilibrium, exhibiting persistent currents in the space of network parameters. As in its physical analogues, the current is associated with an entropy production rate for any given training trajectory. The stationary distribution of these rates obeys the integral and detailed fluctuation theorems -- nonequilibrium generalizations of the second law of thermodynamics. We validate these relations in two numerical examples, a nonlinear regression network and MNIST digit classification. While the fluctuation theorems are universal, there are other aspects of the stationary state that are highly sensitive to the training details. Surprisingly, the effective loss landscape and diffusion matrix that determine the shape of the stationary distribution vary depending on the simple choice of minibatching done with or without replacement. We can take advantage of this nonequilibrium sensitivity to engineer an equilibrium stationary state for a particular application: sampling from a posterior distribution of network weights in Bayesian machine learning. We propose a new variation of stochastic gradient Langevin dynamics (SGLD) that harnesses without replacement minibatching. In an example system where the posterior is exactly known, this SGWORLD algorithm outperforms SGLD, converging to the posterior orders of magnitude faster as a function of the learning rate.

cs.LG↗

arXiv2022

Fano Resonant Optical coatings platform for Full Gamut and High Purity Structural Colors

Structural coloring is a photostable and environmentally friendly coloring approach that harnesses optical interference and nanophotonic resonances to obtain colors with a range of applications including steganography, d\'ecor, data storage, and anticounterfeiting measures. We show that optical coatings exhibiting the photonic Fano Resonance is an ideal platform for structural coloring -- it provides full color access, high color purity, high brightness, controlled iridescence, and scalable manufacturability. We show that an additional oxide film deposited on Fano resonant optical coatings (FROCs) increases the color purity (up to 97%) and color gamut coverage range (> 99% coverage of the sRGB and Adobe color spaces). For coloring applications that do not require high spatial resolution, FROCs provide a significant advantage over existing structural coloring schemes.

physics.optics↗

arXiv2021

Shortcuts in stochastic systems and control of biophysical processes

The biochemical reaction networks that regulate living systems are all stochastic to varying degrees. The resulting randomness affects biological outcomes at multiple scales, from the functional states of single proteins in a cell to the evolutionary trajectory of whole populations. Controlling how the distribution of these outcomes changes over time -- via external interventions like time-varying concentrations of chemical species -- is a complex challenge. In this work, we show how counterdiabatic (CD) driving, first developed to control quantum systems, provides a versatile tool for steering biological processes. We develop a practical graph-theoretic framework for CD driving in discrete-state continuous-time Markov networks. Though CD driving is limited to target trajectories that are instantaneous stationary states, we show how to generalize the approach to allow for non-stationary targets and local control -- where only a subset of system states are targeted. The latter is particularly useful for biological implementations where there may be only a small number of available external control knobs, insufficient for global control. We derive simple graphical criteria for when local versus global control is possible. Finally, we illustrate the formalism with global control of a genetic regulatory switch and local control in chaperone-assisted protein folding. The derived control protocols in the chaperone system closely resemble natural control strategies seen in experimental measurements of heat shock response in yeast and E. coli.

cond-mat.stat-mech↗

arXiv2021

General Inverse Design of Thin-Film Metamaterials With Convolutional Neural Networks

The design of metamaterials which support unique optical responses is the basis for most thin-film nanophotonics applications. In practice this inverse design problem can be difficult to solve systematically due to the large design parameter space associated with general multi-layered systems. We apply convolutional neural networks, a subset of deep machine learning, as a tool to solve this inverse design problem for metamaterials composed of stacks of thin films. We demonstrate the remarkable ability of neural networks to probe the large global design space (up to $10^{12}$ possible parameter combinations) and resolve all relationships between metamaterial structure and corresponding ellipsometric and reflectance / transmittance spectra. The applicability of the approach is further expanded to include the inverse design of synthetic engineered spectra in general design scenarios. Furthermore, this approach is compared with traditional optimization methods. We find an increase in the relative optimization efficiency of the networks with increasing total layer number, revealing the advantage of the machine learning approach in many-layered systems where traditional methods become impractical.

physics.comp-ph↗

arXiv2020

Low force unfolding of a single-domain protein by parallel pathways

Deviations from linearity in the dependence of the logarithm of protein unfolding rates, $\log k_u(f)$, as a function of mechanical force, $f$, measurable in single molecule experiments, can arise for many reasons. In particular, upward curvature in $\log k_u(f)$ as a function of $f$ implies that the underlying energy landscape must be multidimensional with the possibility that unfolding ensues by parallel pathways. Here, simulations using the SOP-SC model of a wild type $β$-sandwich protein and several mutants, with immunoglobulin folds, show upward curvature in the unfolding kinetics. There are substantial changes in the structures of the transition state ensembles as force is increased, signaling a switch in the unfolding pathways. Our results, when combined with previous theoretical and experimental studies, show that parallel unfolding of structurally unrelated single domain proteins can be determined from the dependence of $\log k_u(f)$ as a function of force (or $\log k_u[C]$ where $[C]$ is the denaturant concentration).

cond-mat.soft↗

arXiv2020

Controlling the speed and trajectory of evolution with counterdiabatic driving

The pace and unpredictability of evolution are critically relevant in a variety of modern challenges: combating drug resistance in pathogens and cancer, understanding how species respond to environmental perturbations like climate change, and developing artificial selection approaches for agriculture. Great progress has been made in quantitative modeling of evolution using fitness landscapes, allowing a degree of prediction for future evolutionary histories. Yet fine-grained control of the speed and the distributions of these trajectories remains elusive. We propose an approach to achieve this using ideas originally developed in a completely different context: counterdiabatic driving to control the behavior of quantum states for applications like quantum computing and manipulating ultra-cold atoms. Implementing these ideas for the first time in a biological context, we show how a set of external control parameters (i.e. varying drug concentrations / types, temperature, nutrients) can guide the probability distribution of genotypes in a population along a specified path and time interval. This level of control, allowing empirical optimization of evolutionary speed and trajectories, has myriad potential applications, from enhancing adaptive therapies for diseases, to the development of thermotolerant crops in preparation for climate change, to accelerating bioengineering methods built on evolutionary models, like directed evolution of biomolecules.

cond-mat.stat-mech↗

arXiv2020

Designer Bloch Plasmon Polariton Dispersion in Hyperbolic Meta-Gratings

Hyperbolic metamaterials (HMMs) represent a novel class of fascinating anisotropic plasmonic materials, supporting highly confined propagating plasmon polaritons in addition to surface plasmon polaritons. However, it is very challenging to tailor and excite these modes at optical frequencies by prism coupling because of the intrinsic difficulties in engineering non-traditional optical properties with artificial nanostructures and the unavailability of high refractive index prisms for matching the momentum between the incident light and the guided modes. Here, we report the mechanism of excitation of high-k Bloch-like Plasmon Polariton (BPPs) modes with ultrasmall modal volume using a meta-grating, which is a combined structure of a metallic diffraction grating and a type II HMM. We show how a 1D plasmonic grating without any mode in the infrared spectral range, if coupled to a HMM supporting high-k modes, can efficiently enable the excitation of these modes via coupling to far-field radiation. Our theoretical predictions are confirmed by reflection measurements as a function of angle of incidence and excitation wavelength. We introduce design principles to achieve a full control of high-k modes in meta-gratings, thus enabling a better understanding of light-matter interaction in this type of hybrid meta-structures. The proposed spectral response engineering is expected to find potential applications in bio-chemical sensors, integrated optics and optical sub-wavelength imaging.

physics.optics↗

arXiv2019

Theoretical Perspectives on Biological Machines

Many biological functions are executed by molecular machines, which consume energy and convert it into mechanical work. Biological machines have evolved to transport cargo, facilitate folding of proteins and RNA, remodel chromatin and replicate DNA. A common aspect of these machines is that their functions are driven out of equilibrium. It is a challenge to provide a general framework for understanding the functions of biological machines, such as molecular motors, molecular chaperones, and helicases. Using these machines as prototypical examples, we describe a few general theoretical methods providing insights into their functions. Although the theories rely on coarse-graining of these complex systems they have proven useful in not only accounting for many in vitro experiments but also addressing questions such as how the trade-off between precision, energetic costs and optimal performances are balanced. However, many complexities associated with biological machines will require one to go beyond current theoretical methods. Simple point mutations in the enzyme could drastically alter functions, making the motors bi-directional or result in unexpected diseases or dramatically restrict the capacity of molecular chaperones to help proteins fold. These examples are reminders that while the search for principles of generality in biology is intellectually stimulating, one also ought to keep in mind that molecular details must be accounted for to develop a deeper understanding of processes driven by biological machines. Going beyond generic descriptions of in vitro behavior to making genuine understanding of in vivo functions will likely remain a major challenge for some time to come. The combination of careful experiments and the use of physical principles will be useful in elucidating the rules governing the workings of biological machines.

physics.bio-ph↗