Formal Languages and Automata
Papers
Fresh from arXiv, typeset properly. Highlighted words trade on the market.
All of arXiv
Thursday 27 August
We study a family of language operations based on insertion, shuffle, and crossover and investigate the undecidability of adjacent equality together with finite convergence and associated spectrum questions. Insertion and shuffle operations on formal languages arise in formal language theory, models of concurrency, and biologically inspired computation. This paper studies a different question from the usual closure problem, specifically whether an increasing sequence of languages generated by repeated insertion, or by increasing the permitted degree of bounded shuffle, reaches an instance of adjacent equality after finitely many stages. We show that several such adjacent equality questions are undecidable. In particular, reaching such an adjacent equality event is undecidable for each of the following: iterated insertion of a regular language into a context-free language; bounded shuffle of a regular language with a context-free language as the bound increases; and the corresponding self-insertion and self-bounded-shuffle hierarchies for context-free languages. The new reductions proceed directly from the undecidability of context-free-language universality, using separator-delimited block constructions and, for self-operations, an absorbing regular language of guard violations. Earlier trace-based proofs relied on mortality and uniform halting. More generally, we investigate finite-stage equality and stabilization (persistent equality) in hierarchies generated by insertion and bounded shuffle. In addition to giving substantially simpler proofs of earlier undecidability results, we obtain general criteria for one-step equality, develop new reductions for self-insertion, and identify several open problems, including structural questions concerning insertion depth and degree whose resolution determines whether adjacent equality necessarily implies permanent stabilization.
Quantitative Methods
Towards Efficient Auto-Interpretability of Boltz-1 Sparse Autoencoders
Sparse autoencoders (SAEs) could turn AI models in Biology and other fields into engines of scientific discovery by explaining the superhuman capabilities of those models. We postulate that a latent is useful if we know three things: whether it is coherent, whether it can be described, and whether that description has predictive power. These questions are routinely conflated. We assemble them into a single pipeline and report some practical innovations. First, cross-seed dictionary stability prioritises which latents are worth spending resources to investigate. Second, an intruder-detection task asks whether a latent's activating examples share a recognizable pattern. Third, a separate pass proposes a candidate biological description which we convert into falsifiable predictions which can be tested in silico. Deployed on the Boltz-1 Pairformer trunk, stability prioritisation finds interpretable latents using about 4.4 times fewer latent evaluations each, and at 5.2 times lower measured LLM cost, while recovering over half of them. However, we notice that in our setup the LLM judge is heavily biased towards consistent residue identity and a simple filter that selects for features with high amino-acid identity precision outperforms the stability thresholding. This simpler method has the added benefit of requiring no extra SAE seeds but it remains uncertain if it reflects a genuinely useful heuristic or an limitation of the LLM judge.
Read the full abstract
Computer Vision
What Can Low Resource Languages Learn From Each Other?
Despite the rapid advancement of Vision-Language Models (VLMs), their linguistic reach remains largely confined to high-resource languages, leaving the majority of the world's 7,000+ living languages on the wrong side of a growing digital divide. This disparity is especially pronounced in Optical Character Recognition (OCR), where low-resource scripts lack the massive datasets required for traditional scaling laws. We investigate OCR adaptation in extreme data-scarce regimes (<10K real and <250K synthetic images), demonstrating that conventional fine-tuning strategies often reach a performance ceiling. Our key finding reveals a structural inefficiency in language-specific adaptation: while higher layers of specialized models diverge to capture unique script nuances, the lower layers learn redundant, highly similar features. Motivated by this observation, we propose PSMC (Pre-train, Specialize, Merge, and Co-train), a data-efficient framework that capitalizes on a cross-script "transfer effect". Our approach first derives language-specific experts from a high-resource base model, then employs task arithmetic to fuse these experts into a unified, high-performance multilingual back- bone. Extensive evaluation across 10 Indian scripts (supporting 20+ languages) shows that PSMC achieves a ~2% average improvement in Word Recognition Rate (WRR) over individual specialist models without increasing parameter count. Our results indicate that joint training in the merged latent space facilitates a constructive knowledge transfer that benefits all constituent scripts, providing a scalable pathway for inclusive VLM development. Source code and datasets will be released post publication.
Read the full abstract
Combinatorics
Countable Graphs with Finite Path-width: Characterisation and Universality
We study path-width and the closely related parameter line-width in countably infinite graphs. Our first result characterises the graphs of finite path-width: they are the graphs that do not have infinitely many vertices of infinite degree, do not have infinitely many pairwise disjoint infinite paths, and contain no subdivision of some finite tree of maximum degree 3. We then investigate universality under the subgraph relation for graphs of bounded path-width or line-width. In particular, we prove that there exists a universal graph with line-width $\mathcal{O}(k^2)$ for the class of graphs with line-width at most $k$. In contrast, we show that no graph of finite path-width is universal for the class of locally finite graphs with path-width $1$. Finally, we show that for each $k\geq 2$, every universal graph for the class of graphs with path-width at most $k$ has line-width at least $k + 1$.
Read the full abstract
Biological Physics
Noise-robust navigation from an adaptive run-and-tumble policy
How do organisms navigate when the signals guiding them are noisy? Variance adaptation, the rescaling of sensitivity to noise, is common in sensory systems, but its role in navigation is unexplored. We introduce a minimal active Brownian particle whose run-and-tumble policy follows from an optimality principle. Variance adaptation emerges as part of this policy. Adaptation keeps chemotactic drift finite as noise grows, while a non-adaptive particle's collapses exponentially. Adaptation also carries a cost, degrading performance in quiet environments and requiring a tuned adaptation sensitivity.
Read the full abstract
Machine Learning
The Calls are Coming from Inside the Model: Investigating Probe-based Detection of Tool-Calling Errors in LLMs
The hidden states of large language models (LLMs) are known to capture rich information relating to model knowledge and behavior that can be hard to extract from examination of input and output alone. As LLM-based systems increasingly interface with the external world, one area of concern is detecting incorrect or improper use of tools. Motivated by this, we study the effectiveness of using linear probes to detect incorrect tool-calls, measuring probe efficacy across 18 tool-calling LLMs evaluated on the Berkeley Function Calling Leaderboard. Overall, we find that probing is an effective means to catch a range of different tool-calling errors, including errors arising from using an argument that has the wrong value but the correct type, which might not be recorded by standard logging frameworks. Important factors in success include model size, probing layer, and model post-training type. We also show that probes are capable of generalizing to novel types of errors, which is critical in real world deployments.
Read the full abstract
Numerical Analysis
High-order robust basis-update & Galerkin integrators for dynamical low-rank approximation
Dynamical low-rank approximation has become a widely used numerical method in diverse disciplines. Its main idea is to represent the matrix or tensor-valued solution to a time-dependent differential equation as a low-rank factorization. The evolution of the factorization leads to highly stiff dynamics and requires the derivation of novel time integration methods that are not prone to this stiffness. A promising family of integrators are basis-update & Galerkin (BUG) integrators as they enable implicit time integration and structure–preservation. However, current BUG integrators are limited to second–order accuracy while general-order BUG integrators are designed as projections of explicit time integration methods, thus severely limiting their use. In this work, we propose general-order BUG integrators that do not rely on an explicit time integration scheme while requiring a smaller number of basis functions to achieve high-order accuracy. We prove a general order error bound for the proposed augmented and parallel BUG integrators and demonstrate their behaviour for a series of stiff and non-stiff numerical benchmarks in which they significantly outperform previous BUG versions.
Read the full abstract
Functional Analysis
The role of parameter Jacobians in the stability of network outputs
In the framework of network dynamics, learning models, and neural tangent kernels (NTK), we show that the corresponding linearized dynamics leads naturally to a semigroup formulation. More precisely, in our analysis of input/output models, the time-dynamics is presented via special semigroups of linear operators on Hilbert spaces, together with an associated class of semigroup perturbations. In this context, we then present new and explicit a priori perturbation-bound results: for the fixed-kernel linearization constructions arising in the NTK setting, we prove norm-bounds on the corresponding semigroup perturbations, in the form of explicit finite-time perturbation estimates. We further present refinements on prescribed task spaces, Cesàro-averaged (ergodic) comparisons estimates, and versions in which the lower spectral edge assumption is replaced by a spectral-distribution condition. We also extend the comparison to nonautonomous NTK evolutions through piecewise-frozen approximations, record a corresponding discrete Euler specialization, and offer worked examples in order to illustrate our perturbation-bound estimates.
Read the full abstract
Materials Science
Electronic structure, magnetic interactions, and magnonics of 2D trichloride materials
Using advanced density functional theory, this study demonstrates the robust thermodynamic, structural, and dynamical stability of pristine, 3d-doped (Ti, Cr), and 4f-doped (Ce) transition metal trichlorides (MCl3). While standard generalized gradient approximation (GGA) incorrectly predicts metallic behavior, hybrid functional calculations successfully capture their semiconducting nature, yielding accurate band gaps of 2.54 eV (VCl3), 4.01 eV (CrCl3), and 2.61 eV (TiCl3). These materials exhibit intrinsic ferromagnetism, with CrCl3 displaying overlapping topological features in both its magnon and phonon dispersions along the high-symmetry K direction, which induces magnon-phonon coupling. Incorporating 3d dopants into VCl3 successfully tunes the bandgap (1.32 eV for Ti and 2.74 eV for Cr) without destroying ferromagnetism, while 4f Ce-doping in CrCl3 introduces localized states below the Fermi level that yield strong nearest-neighbor exchange coupling. Ultimately, this electronic and magnetic tunability highlights the potential of these 2D trichlorides in advanced spintronic functionalities.
Read the full abstract
Information Retrieval
NormasTCU — A Brazilian Portuguese IR Dataset and an Evaluation of LLM-as-a-Judge for Relevance Assessment
Portuguese Information Retrieval (IR) lacks public datasets, and relevance assessment for specialized collections remains costly. While Large Language Models (LLMs) increasingly support relevance assessment, their reliability in non-English specialized domains remains unclear. We introduce NormasTCU (https://huggingface.co/datasets/LeandroRibeiro/NormasTCU), a Brazilian Portuguese IR dataset with 14,469 legal documents, 46 queries, and 3,048 human judgments over 812 query-document pairs. Using NormasTCU, we evaluated LLM-as-a-judge for relevance assessment by prompting three models with two prompt techniques to grade these pairs. We then compared the rankings of 15 IR systems derived from LLM-generated and human reference qrels. LLMs consistently showed a positive scoring bias (mean absolute error: 0.46–0.66 on a 0-2 scale). Furthermore, pair-level agreement with human judgments achieved only fair to moderate levels, with Cohen's kappa ranging from 0.32 to 0.53. Despite this bias, LLM-generated judgments often yielded highly similar system rankings for nDCG@10 and MRR (observed Kendall's tau greater than or equal 0.90, although the bootstrap confidence intervals did not always remain above this threshold), but were less reliable for P@10 and R@10. Notably, LLM-based rankings were sometimes more strongly correlated with the reference ranking than individual human annotations were. As a practical implication, our results suggest that LLMs could effectively support scalable relevance assessment in specialized Portuguese corpora when evaluated using nDCG or MRR (rank-aware metrics), but they should be avoided when relying on precision or recall.
Read the full abstract
Number Theory
An Erdős-Wintner theorem for second-order linear recurrent bases
Let $a,b$ be integers with $1\le b\le a$, and let \[ G_0=1,\qquad G_1=a+1,\qquad G_{n+2}=aG_{n+1}+bG_n. \] For real-valued functions which are additive in the greedy $G$-digits, we prove a necessary-and-sufficient criterion for the existence of a limiting distribution. The criterion consists of a first-order drift series and a quadratic digit-energy series, and it recovers the Zeckendorf theorem when $a=b=1$. The main difficulty is necessity: a one-step transfer matrix detects all non-maximal digit values, while the maximal digit becomes visible only in a two-step product. A weighted Euclidean norm symmetrizes the untwisted companion matrix and makes both contractions occur at the true Perron scale. Sufficiency follows from a two-dimensional Perron product lemma with square-summable transverse perturbations and a convergent, not necessarily absolutely convergent, Perron drift. The limiting characteristic function admits a scalar infinite-product representation in a neighbourhood of the origin and a global matrix-product representation at arbitrary frequencies.
Read the full abstract
High Energy Astrophysical Phenomena
The Impact of Errors in the Shape Function of Rotating Neutron Stars in the Oblate Schwarzschild Approximation
Measurements of thermal X-ray flux emitted by neutron stars can constrain their radii and the equation of state of cold dense matter. Accurate flux modeling of rapidly rotating neutron stars requires an approximation for their oblate surface shape, usually given by a shape function. We examine the accuracy of five shape functions by comparing them against numerical relativistic stellar surfaces for seven representative equations of state over a broad range of compactness and spin. We characterize their errors in the surface radius and their derivative with respect to colatitude, test the extent of quasi-universality, and propagate these errors into the differential solid angle d{\Omega}, a geometric factor underlying the observed flux, under the oblate Schwarzschild approximation. We find that the shape functions produce distinct error patterns associated with their forms and parameter ranges. These errors remain structured in compactness, spin, and colatitude. Simple polynomial corrections substantially reduce the radius errors of several shape functions, with most corrected errors within approximately 0.5%. We show that errors in d{\Omega} may be understood as coming from an area contribution, which scales with twice the relative radius error, and a projection contribution controlled primarily by the error in the surface derivative. The projection contribution grows towards the limb, where small surface errors can produce much larger local errors in d{\Omega}. We provide a set of guidelines for the construction of improved quasi-universal shape functions that can be used in future radius estimation efforts. Our recommendations include that both the shape and its derivative be modeled accurately, that more accurate predictions of the polar radius be developed, and that the shape function be calibrated over the same compactness-spin range that is to be tested.
Read the full abstract
Systems and Control
Explicit Separators for Consecutive Levels of Parrilo's Sum-of-Squares Hierarchy over the Copositive Cone
Parrilo's cones $\Kc{n}{r}$ form a nested sequence of semidefinite-representable inner approximations of the copositive cone $\COP_n$. For $n=5$ their union is all of $\COP_5$, yet no single level attains it, and whether consecutive levels actually differ had remained open beyond the classical first step. No explicit matrix in $\Kc{n}{t}\setminus\Kc{n}{t-1}$ had, to our knowledge, been published for any $t\ge2$ and $n\ge5$. We settle the first three cases. Explicit rational matrices, obtained from diagonal scalings of the Horn matrix shifted along a positive interior direction, lie in $\Kc{5}{2}\setminus\Kc{5}{1}$, in $\Kc{5}{3}\setminus\Kc{5}{2}$, and in $\Kc{5}{4}\setminus\Kc{5}{3}$, giving three consecutive strict inclusions $\Kc{5}{1}\subsetneq\Kc{5}{2}\subsetneq\Kc{5}{3}\subsetneq\Kc{5}{4}$. Each is certified by an exact rational Gram matrix and an exact rational dual moment functional, re-verified by a standalone program in integer arithmetic. The separations are robust. One fixed certificate pair covers an interval of shifts of width exceeding $3\cdot10^{-3}$, and $\Kc{5}{2}\setminus\Kc{5}{1}$ has nonempty interior. Combining a scaling theorem of Dickinson, Dür, Gijben and Hildebrand with the completeness theorem of Schweighofer and Vargas shows further that strict adjacent inclusions recur at arbitrarily large levels. All separators were located by one threshold device: the least shift $\eps_r(M)$ carrying $M$ into $\Kc{5}{r}$ along an interior direction is nonincreasing in $r$, and each strict drop between levels marks a window of separators.
Read the full abstract
Dynamical Systems
Below-threshold Bistability and Implementation Lag in a Simplex Model of Radical Vote-Share Dynamics
We study a nonautonomous compartmental model on the probability simplex for radical vote-share dynamics under delayed policy implementation. The model distinguishes a target regime from an effective regime that adjusts with finite speed and includes a persistent alienation reservoir that can later be mobilised by radical actors. In a symmetric benchmark, persistent alienation destroys the global one-threshold picture: below the local instability threshold, a radical-free centrist–alienated equilibrium may coexist with a stable positive equilibrium, separated by a saddle branch created in a saddle-node bifurcation. By hyperbolic persistence, this coexistence survives near the symmetric baseline, while numerical exploration identifies a wider bistable region under asymmetric perturbations. Delayed implementation further separates target and effective threshold passage, producing a parameter-space lag and an additional state-space lag before a visible response in radical support. The analysis combines a Perron–Frobenius threshold for the frozen radical-free equilibrium, a geometric characterisation of positive equilibria, local adiabatic tracking under slow uniformly subcritical drift, and explicit delay bounds for transversal threshold passage. Under initial matching, the parameter-space delay scales as \(O(\kappa_\theta^{-1})\); in a monotone ramp example the bound is nearly attained, whereas the state-space lag is substantially larger and less sensitive to implementation speed. Thus threshold restoration is preventive rather than curative: once the trajectory enters the basin of the radicalised attractor, returning below the local threshold need not restore the radical-free regime.
Read the full abstract
Numerical Analysis
The singular Zienkiewicz tetrahedron: Definition and Integration
We extend the two-dimensional singular Zienkiewicz element to three dimensions, leading to a novel $H^2$-conforming finite element on tetrahedral meshes based on rational shape functions. Besides the finite element construction and its conformity analysis, we develop an exact iterative integration procedure for the associated class of rational functions. The resulting formulae allow for the exact integration of the basis functions, their derivatives, and the products occurring in finite element assembly.
Read the full abstract
Quantum Physics
Error-Adaptive Quasi-local Decoding of the Toric Code
Topological quantum memories need decoders that are both reliable and scalable, but these goals compete: globally informed decoders are accurate near threshold yet expensive, while strictly local rules are fast but can miss long-range structure. Motivated by recent recoverability and mixed-state viewpoints, we make this tradeoff operational for the dephased toric code through a decoder-level local recoverability diagnostic. We compare global MWPM corrections to quasi-local corrections inside a target region and define a matching ratio $R_{\mathrm{match}}$, with mismatch $\epsilon_{\mathrm{match}}=1-R_{\mathrm{match}}$. Across geometry families, $\epsilon_{\mathrm{match}}$ shows strong buffer-controlled suppression and is well organized by a two-geometry scaling form. For scaled families, especially $a=b=d/8$, $R_{\mathrm{match}}$ exhibits a clear crossing and finite-size collapse near $p\!\sim\!0.09$, consistent with a growing recoverability scale near the decoding transition. We use this scaling to formulate an adaptive buffer-selection rule and to identify a distance-scaled initialization for a composite quasi-local RG decoder on the full torus. In the tuned family $a_0=b_0=d/8$, the resulting logical-failure curves show an apparent finite-size crossing at $p\simeq0.09$–$0.10$, below the conventional MWPM threshold scale. We focus on dephasing noise with perfect syndrome measurements to cleanly isolate the underlying behavior.
Read the full abstract
Signal Processing
Data-Aided Asynchronous OFDM Integrated Sensing and Communications: A Mean-Field Variational Bayes Approach
Integrated sensing and communication (ISAC) is regarded as a key technology for sixth-generation wireless networks, allowing sensing and communication operations to jointly utilize the same spectrum and hardware infrastructure. However, in practical uplink ISAC systems, timing offset (TO) and carrier-frequency offset (CFO) introduce phase distortions across subcarriers and OFDM symbols, which can severely degrade both data detection and sensing-parameter estimation. In this paper, we propose a data-aided variational Bayesian (VB) framework for asynchronous uplink OFDM-ISAC systems. Specifically, the received signal is modeled as a sparse multipath superposition, where the transmitted data symbols, complex path gains, spatial frequencies, delay-Doppler parameters, and synchronization parameters are jointly inferred. To enable tractable inference, we develop a mean-field VB algorithm in which von Mises distributions are used for gridless updates of the angular and delay-Doppler phase parameters, while a Gamma-Gaussian prior is adopted to promote path sparsity. A key feature of the proposed framework is its data-aided sensing capability: after initial pilot-based estimation, the detected data symbols are exploited as additional observations to refine the channel and sensing parameters. This substantially increases the effective sensing resources without requiring extra pilot overhead. The simulation results demonstrate that the proposed approach achieves superior performance compared with SAGE, SBL, AB2FM, and pilot-only VB baselines in terms of symbol error rate, channel reconstruction accuracy, path-parameter estimation, TO/CFO estimation, and 3D localization accuracy. The results also demonstrate that ignoring TO and CFO leads to severe sensing degradation, highlighting the importance of synchronization-aware and data-aided receiver design for ISAC systems.
Read the full abstract
Algebraic Geometry
Singularity Categories of Simple Singularities in Positive Characteristic
We study the singularity categories of simple singularities of the same dimension over an algebraically closed field of positive characteristic, and show that, as in characteristic zero, these categories are not equivalent as triangulated categories unless the underlying singularities are analytically isomorphic. In contrast to the characteristic zero case, simple singularities in positive characteristic cannot be distinguished solely from the Auslander-Reiten quivers of their singularity categories. To address this, we extend to positive characteristics a theorem by Hua and Keller, which asserts that the 0th Hochschild cohomology of the dg singularity category of an isolated hypersurface singularity in characteristic zero is isomorphic to the Tyurina algebra of the defining polynomial. Furthermore as an application, we determine the condition for the singularity category of a rational double point (i.e., a simple singularity of dimension two) to be standard. We prove that such a category is standard if and only if the defining polynomial is weighted homogeneous.
Read the full abstract
Operator Algebras
Non-unitarizable representations of compact and discrete quantum groups
We study the similarity problem for representations of compact and discrete quantum groups. We first prove that every non-degenerate contractive representation of a compact or discrete quantum group is automatically unitary. Then, for compact quantum groups, we develop an interpolation method for constructing explicit non-unitarizable non-degenerate representations with norms arbitrarily close to $1$. In particular, this applies to all non-Kac compact quantum groups whose dual has subexponential growth, including the Drinfeld–Jimbo quantum groups $G_q$, as well as to the non-Kac free unitary quantum groups $U_F^+$ with $F\in\operatorname{GL}_2(\mathbb{C})$. On the discrete quantum group side, we establish a quantum analogue of the classical lifting principle for non-unitarizable uniformly bounded representations. Combining this lifting principle with recent results on maximal Kac quantum subgroups, we construct explicit non-unitarizable non-degenerate representations with norms arbitrarily close to $1$ for all unitary free quantum groups $\mathbb FU_F$ and all non-amenable orthogonal free quantum groups $\mathbb FO_F$.
Read the full abstract
Physics and Society
Don't look now: paying for documented need punishes discovery in US lead-pipe funding
The United States allots USD 15 billion for lead service line replacement in proportion to documented lead, the largest of a growing class of environmental-indicator transfers. Reconstructed exactly, the rule prices looking below zero. Resolving an unknown pipe lowers a state's expected allotment in three of the eight jurisdictions we can measure above the statutory floor, and raises it in none. Below it the allotment cannot move. Systems filing nothing are scored lead-free. We prove no fixed formula on self-reported inventories can pay for need without punishing its measurement, so we made an audited one. The unique additive repair credits unknowns at a rate no filing can move, and is strategy-proof. Discovery is repriced from a median loss of USD 48.63 per line per year to zero. Allocation swings fall from 16.5% of the pool to 0.8%, for an audit costing 1.23% of annual funds. The formula's tilt toward vulnerable systems survives.
Read the full abstract
Computer Vision
ABCD: Alpha-Composited Block Coordinate Descent: Constant-VRAM Training for Large Radiance Fields
We present ABCD (Alpha-Composited Block Coordinate Descent), an out-of-core training framework for alpha-composited radiance fields, instantiated here for 3D Gaussian Splatting. Our method reformulates training as block coordinate descent over spatial partitions: only one block of parameters is active at a time, while all others are frozen. By exploiting the associativity of alpha blending, these inactive regions can be pre-rendered and collapsed into foreground and background RGBA images. As a result, for fixed partition size and image resolution, peak VRAM becomes O(1) with respect to total scene extent, rather than growing with full scene size. This enables GPUs with limited memory to train scenes that would otherwise not fit in core. In experiments, our method closely preserves the reconstruction quality of 3DGS, with less than 5% PSNR degradation, while ABCD with compositing ablated suffers roughly 40% degradation. Our code can be found at https://github.com/shiukaheng/abcd
Read the full abstract
Statistical Finance
What survives honest evaluation? Leakage-safe, search-aware assessment of LLM-driven trading strategy discovery
Large language models (LLMs) are increasingly used to discover trading strategies, and much of the resulting literature shares a methodological weakness: many candidate strategies are generated, the best is reported, and neither look-ahead bias nor the intensity of the search behind the reported result is corrected for. We present a strategy-discovery system that makes both corrections structural rather than procedural. First, the agent can only act through registry-validated tools whose feature space excludes look-ahead by construction; we show that this guardrail is not redundant with statistical correction: a deliberately leaky oracle posting a Sharpe ratio of 35 survives Deflated Sharpe and probability-of-backtest-overfitting testing completely. Second, the system records every strategy evaluation its search performs and deflates all reported performance by that trial count, tracing how the best in-sample Sharpe ratio climbs with each trial while the deflation threshold, driven by the agent's own search, climbs faster. Across a 453-stock point-in-time US equity universe and a 39-ETF multi-asset universe with realistic transaction, impact, and borrow costs, honest evaluation certifies passive benchmarks (out-of-sample confidence intervals excluding zero), rejects every LLM-discovered strategy (across two frontier models, search budgets up to one hundred candidates, and five repeated runs), catching selection luck, predicted rank degradation, and out-of-sample collapse through complementary instruments, and evaluates a human trader's production rule system under identical instruments. The framework formalizes why pre-registered hypotheses earn lower evidential bars than brute search, and quantifies the sample sizes that credible certification of moderate edges actually requires.
Read the full abstract
Optics
Deterministic control over launching efficiency of higher-order hyperbolic phonon polaritons
Hyperbolic materials, which exhibit an extreme form of birefringence enabling the volume confinement and frequency-dependent propagation of deeply sub-diffractional optical modes, offer the opportunity for extreme confinement via the stimulation polaritonic modes, with substantially higher confinement obtained through the efficient excitation of of the higher-order (shorter wavelength) hyperbolic polaritonic modes, which they can support. However, while these higher-order hyperbolic polaritons (HO-HPhPs) form high-momentum ray-like propagation within the bulk, efficient excitation of these modes, especially in contrast to the long-wavelength lower-momentum surface polariton propagating modes, has remained a challenge. Critically, the large momentum mismatch between these modes and free-space light, alongside the spatial mismatch between the sub-diffractional scatterer and the distinct modal distribution of HO-HPhPs, lead to a suppressed launching efficiency of these higher-order modes, limiting their use in nanophotonic applications. Here, we report the experimental observation of a 10-fold enhancement in the excitation efficiency of HO-HPhPs through the use of subsurface scatterers over traditional surface scattering (e.g. a flake edge or gold launcher) within single-crystalline {\alpha}-MoO3 slabs. We employ full-wave numerical simulations to investigate the role of the spatial overlap between HO-HPhP modal distributions and the scatterer placement upon excitation efficiency. Furthermore, we develop a generalized process using transfer matrix method to selectively design modal HO-HPhP excitation, which advances the capabilities of HPhP multiplexing for on-chip applications.
Read the full abstract
Number Theory
Trilinear Kloosterman fractions II: subdyadic intervals and nearly balanced convolutions
This paper broadens the range on which Fouvry and Radziwiłł's results on nearly balanced convolutions apply. In particular, let $\alpha_m$ and $\beta_n$ be sequences supported on $m\sim M$ and $n\sim N$ where $\beta_n$ is equidistributed for small moduli, and let $Q=X^{\frac 12+\varepsilon}$. We find that \begin{gather*}\sum_{q\sim Q}\left|\mathop{\sum\sum}_{\substack{n\sim N,m\sim M \\ mn\equiv a\pmod q}}\alpha_m\beta_n-\frac{1}{\phi(q)}\mathop{\sum\sum}_{\substack{n\sim N,m\sim M \\ (mn,q)=1}}\alpha_m\beta_n\right|\ll \frac{X}{\log^A X} \end{gather*} if $N=X^{\frac 12+\delta}$ and $M=X^{\frac 12-\delta}$ with $0<\delta<\frac 1{68}$, which improves Fouvry and Radziwiłł's $0<\delta<\frac 1{112}$. To prove this, we sharpen Bettin and Chandee's famous result on trilinear forms with Kloosterman fractions in the case where some of the sums are over subdyadic intervals.
Read the full abstract
High Energy Physics (Theory)
Self-Dual Yang-Mills in AdS$_4$
Quantum Yang-Mills theory in Euclidean AdS$_4$ with Neumann boundary conditions is studied as a function of the complex couplings $\frac{1}{g_\pm^2} \equiv \frac{1}{g^2}\mp \frac{i \theta }{8 \pi^2}$. We define a "self-dual limit" by $g_- \to 0$ with $g_+ ~{\rm fixed}$. We argue that the limiting theory is well-defined perturbatively and equivalent to the BF formulation of self-dual Yang-Mills theory with a particular boundary condition relating $B$ and $F$ at the boundary $\partial$AdS$_4$. $g_+$ is the loop-counting parameter of the self-dual theory. This is in stark contrast to flat space, where $\theta$ has no effect on perturbative dynamics. In the self-dual limit, it is shown that all tree-level zero-plus and single-plus boundary correlators vanish. A closed form expression is given for any number of gluons for the double-plus tree correlators. Finally, we show that in the flat space limit of AdS$_4$, the total energy poles in the tree boundary correlators correctly reduce to the single-minus gluon amplitudes.
Read the full abstract
Astrophysics of Galaxies
Overmassive supermassive black holes in SDSS close galaxy pairs
Supermassive black holes (SMBHs) and their host-galaxies coevolve through channels such as hierarchical merging and AGN feedback, establishing tight scaling relations between SMBH masses and properties of the host-galaxy. These scaling relations, particularly those with host-galaxy bulge mass (M${_\text{b}}$) and stellar velocity dispersion ($\sigma$), are widely used to estimate SMBH masses. Galaxy mergers drive gas into the centers of galaxies, enhancing SMBH growth and stellar bulge growth but not necessarily in tandem, so it is necessary to determine if galaxies undergoing a merger still follow SMBH-host-galaxy scaling relations. In this study, we explore scaling relations in galaxy mergers by taking advantage of the well-established value-added catalogs in the Sloan Digital Sky Survey. These catalogs provide close galaxy pairs, AGN broad lines for SMBH mass measurements, and bulge-disk decomposed stellar masses. We compare SMBH mass estimates from M$_\bullet-$M${_\text{b}}$ to SMBH mass estimates from AGN broad lines and find that SMBH mass growth outpaces the bulge for AGN residing in the secondary (less massive) bulge and for smaller physical separations between galaxies in a pair. Using our full sample of close galaxy pairs, we find that M$_\bullet-$M${_\text{b}}$ and M$_\bullet- \sigma$ predict significantly different fractions of major and minor black hole mergers with M$_\bullet-$M${_\text{b}}$ predicting black hole mass ratios closer to 1:1 and $6-20\%$ more major mergers overall. These results have major implications, including for predictions of astrophysical gravitational-waves and high-redshift overmassive SMBHs.
Read the full abstract
Computation and Language
Below the Noise Floor: Bimodal Seed Collapse and Distinct Failure Modes in Small-Model Knowledge Distillation
Function routing – selecting the correct API call from a fixed catalog given a natural-language request – is a deployment problem where small students are attractive but knowledge distillation gains are typically reported single-seed, at scales where seed variance is unknown. On a 740-instance healthcare API routing task with a 1.5B Qwen student and a 20B teacher, we compare eight KD variants against supervised cross-entropy, using three to six seeds for key configurations. We find: (i) per-seed standard deviation ranges from 2.8 to 48.7 percentage points, swallowing every claimed KD gain below five points; (ii) three of seven KD variants exhibit bimodal collapse, with at least one in three to five seeds falling below 55% accuracy while the others train normally, and a fourth showing elevated variance; (iii) collapse has distinct modes – wrong-function selection for ce_kd and ce_paraphrase, and a previously undocumented output-truncation mode for reasoning_kd, where the model emits reasoning but terminates before producing a function name (0.9% accuracy); (iv) only progressive_kd and rank_kd avoid collapse across observed seeds, with sigma <= 3.9 pp; (v) a naive cross-split +3.78 pp gain from input enrichment reverses to -2.70 pp under controlled within-split multi-seed re-testing. Single-seed evaluation is therefore unable to detect central failure modes in small-model KD.
Read the full abstract
Machine Learning
Diffusion Distillation for Efficient Weather Ensembles
Diffusion models generate skillful weather ensembles but require costly iterative sampling. We introduce a supervised energy-distance distillation method that compresses a multi-step diffusion teacher into a single-step student by aligning student forecasts with teacher samples and ground-truth observations. Experiments on global forecasting and typhoon-track prediction show that our student outperforms existing distillation methods and preserves skill for extreme events. It matches or surpasses the teacher across key metrics using only one neural function evaluation per autoregressive step.
Read the full abstract
Artificial Intelligence
Probing Perceptual Priors of MLLMs via Gibbs Sampling with Interpretable Generative Controls
A model's behavior on a task is jointly determined by the input it receives and the prior it brings in, i.e. the distribution over stimuli it implicitly expects. Interpretability research has traditionally studied models by holding inputs fixed and examining model responses either mechanistically, probing how internal structure represents inputs, or behaviorally, measuring how variation in inputs leads to variation in outputs. Neither reconstructs the prior distribution itself, since internal structure shows what a model can represent, not what it expects, and any fixed stimulus set leaves most of the possible input space unseen. In particular, such an input space in real-world settings, such as images seen by VLMs, is extremely high-dimensional and diverse. These priors thus remain a poorly understood component of models that nonetheless influence real-world behavior. We propose a method to sample from models' perceptual prior distributions directly, by steering a generative model to produce stimuli along controllable axes and running Gibbs sampling over that space with the model under study as the judge. We apply this to a variety of categories and target variables (such as trustworthiness in faces and cheapness in art images) and recover both canonical biases and surprising novel priors invisible to direct prompting, warranting further investigation of their downstream effects.
Read the full abstract
Robotics
Coordinated Motion Planning for Multi-Arm Systems via Iterative LQ Games
Multi-agent motion planning for high-degree-of-freedom robotics manipulators in shared workspaces remains a fundamental yet challenging problem. Centralized planners often suffer from poor scalability, while decentralized approaches face robustness and safety concerns. Game-theoretic formulations offer a promising approach for modeling agent interactions, potentially overcoming these limitations. However, their application to articulated multi-arm systems remains limited. This paper presents an iterative Linear Quadratic (LQ) game framework for multi-manipulator motion planning, where each manipulator is modeled as an independent agent optimizing its own objective while interacting with other agents based on shared global states and collision constraints. The method solves a series of local LQ games by linearizing the dynamics and approximating the cost around a nominal trajectory, with Riccati backward recursions yielding feedback Nash strategies. To address the challenges of articulated systems, we incorporate differentiable penalties for self-collision and inter-arm collision into the optimization pipeline, enabling coordinated, collision-aware trajectory generation. Experiments demonstrate that our framework produces smooth, safe, and efficient trajectories in high-dimensional settings, outperforming traditional methods. This highlights the effectiveness of differential game formulations for multi-robot manipulation.