Computational Complexity
Papers
Fresh from arXiv, typeset properly. Highlighted words trade on the market.
All of arXiv
Friday 6 November
We show that over the field of rational numbers, there is a deterministic algorithm for factorization of constant depth circuits that makes calls to a (whitebox or blackbox) PIT algorithm for such circuits. The time complexity of the algorithm (and hence the number of calls to the PIT oracle) is polynomially bounded in the size and degree of the input circuit. Our reduction to PIT is fairly simple and general and applies as it is to other natural classes of algebraic circuits like formulas and algebraic branching programs. Prior to this work, the known deterministic factorization algorithms for constant depth circuits crucially relied on blackbox PIT [Bhattacharjee et al.]. Our main technical ingredient is an efficient, deterministic algorithm that, when given access to sufficiently high degree truncations of all the power series roots of a constant depth circuit, uses oracle calls to any (whitebox or blackbox) PIT algorithm for these circuits, and partitions the roots into the sets of conjugates; thereby identifying the irreducible factors. For our proof, we rely on a root recombination criterion using logarithmic derivatives [Sasaki et al., Bostan et al., Lecerf] and show that this criterion can be efficiently and deterministically implemented for power series roots of constant depth circuits using access to any PIT oracle. On the way, we discuss an algorithm that may be of independent interest: it takes as input a matrix whose entries are multivariate constant depth circuits over a field $\mathbb{F}$, requires oracle access to any PIT algorithm for these circuits and outputs a basis for the $\mathbb{F}$-kernel of this matrix in polynomial time.
Mathematical Physics
The dynamics through the first two Lagrange necks in the small-mass restricted three-body problem
We study the planar circular restricted three-body problem (PCR3BP) for sufficiently small mass ratios, from low energies to slightly above the second Lagrange value. We show that retrograde orbits around the primaries and Lyapunov orbits at the necks form the binding of a finite energy foliation. We construct these foliations using the theory of pseudoholomorphic curves in symplectizations developed by Hofer, Wysocki and Zehnder. Slightly above the second Lagrange value, the dynamics is transverse to families of disks bounded by the retrograde orbits, and cylinders connecting the retrograde orbits to the Lyapunov orbits, forming a $3-2-3-2$ foliation. The proof uses the rotating Kepler problem and Hill's lunar problem to control the singular limit as the mass ratio $\mu \to 0$. As a consequence of the foliation, both Lyapunov orbits have infinitely many homoclinic orbits.
Read the full abstract
Nuclear Theory
Baryon Number Susceptibilities of Hot and Dense Nuclear Matter in Relativistic Mean-Field Models
We develop a fully analytical framework for calculating baryon number susceptibilities in relativistic mean-field (RMF) models of hot and dense nuclear matter at finite temperature and baryon chemical potential. The formalism is applied to representative nonlinear RMF parameterization sets to investigate the model dependence of baryon number fluctuations over a wide range of temperatures and baryon chemical potentials. All the considered RMF parameterizations reproduce the overall qualitative beam-energy dependence of the cumulant ratios. The observed skewness and kurtosis within the RMF models satisfy the experimental values. Specifically, the FSU-family parameterizations exhibit relatively small deviations from the measured heavy-ion observables i.e. cumulant ratios.
Read the full abstract
High Energy Physics (Theory)
Induction instead of Feynman diagrams, L-functionals instead of Hilbert space
The main result of this paper is a construction of an inductive process that can replace Feynman diagrams. Combining this result with the notion of the L-functional, one can give an exposition of QED that does not suffer from infrared divergences
Number Theory
The existence of small defining polynomials for number fields, II
We prove that every number field of odd degree $n\geq 3$ and absolute discriminant $D$ has an integral defining polynomial whose coefficients are $O_n(D^{(n-1)/(2n-1)})$ in size. This strengthens, in odd-degree cases, the recently established bound of $O_n(D^{1/2})$ for all degrees $n$, which was conjectured in 1998 by Ruppert and is now known to be optimal in even degrees. More generally, we prove that this strengthened bound holds for all degree-$n$ fields having a real embedding and no subfield of index two. The proof makes use of the lattice shapes of the ring of integers and its trace dual, the inverse different. Assuming the quasi-Riemann hypothesis for Dedekind zeta functions, as recently announced by Cureton and Thorner, we obtain an exponent smaller than $1/2$ for every field with no index-two subfield, of any signature. In the case of primitive fields, we obtain the exponent $(n-1)/(2n)$. In particular, we prove that the optimal exponent for cubic fields is $1/3$.
Read the full abstract
Robotics
A Controlled Study of Models, Skills, Feedback, and Curriculum in Agentic Robotics
Agentic robotics asks coding agents to implement and iteratively improve robot policies from simulated and real-world physical interactions. We study the self-learning loop of agentic robotics through a controlled simulation study spanning nine manipulation tasks, 631 different experiment configurations, and 28K coding agent rollouts, systematically varying model capability, skill priors, feedback structure, and curriculum. The experiment data suggests that the bottleneck is often not the access to sufficient information, but interpreting it into the right policy update. Across these factors, a successful self-learning loop depends on how effectively agents attribute failures, reason task-relevant spatial and geometric structure, and selectively restructure code to improve policies.
Read the full abstract
Analysis of PDEs
Delaunay classification of positive singular solutions to the critical Lane–Emden system
We classify positive classical solutions of the Lane–Emden system \begin{align*} \begin{cases} -\Delta u=v^p & ~~\text{in }\mathbb{R}^n\setminus\{0\},\\ -\Delta v=u^q &~~ \text{in }\mathbb{R}^n\setminus\{0\}, \end{cases} \end{align*} with a nonremovable singularity at $0$ (i.e., $\lim_{x\to0}u(x)=\lim_{x\to0}v(x)=+\infty$), where $n\ge3$ and $p,q>0$ satisfy \[ \frac1{p+1}+\frac1{q+1}=\frac{n-2}{n}. \] Without additional growth or integrability assumptions, we establish radial symmetry about the origin, strict radial decrease, and global two-sided power bounds for both components. Furthermore, every such solution can be written as \[ u(x)=|x|^{-\alpha}U(\log|x|),\qquad v(x)=|x|^{-\beta}V(\log|x|), \] where $\alpha=\frac{n}{q+1}$, $\beta=\frac{n}{p+1}$, and $(U,V)$ is either the positive equilibrium $(A,B)=((\frac{\alpha}{\beta})^{\frac{\alpha}{2}}, (\frac{\alpha}{\beta})^{\frac{\beta}{2}})$ or a positive nonconstant periodic solution of the associated Emden–Fowler system. These periodic solutions form a Delaunay family: for every $\varepsilon\in(0,A)$, there is a unique profile pair, up to translation in $t=\log|x|$, with $\min_{\mathbb{R}}U=\varepsilon$. The Hamiltonian energy provides an alternative parameter, decreasing continuously and strictly from zero to the equilibrium energy as $\varepsilon$ increases from $0$ to $A$. The common least period of the two components tends to infinity as $\varepsilon\downarrow0$ and to the period of the linearized oscillation at $(A,B)$ as $\varepsilon\uparrow A$. The classification includes the case of a sublinear exponent.
Read the full abstract
Artificial Intelligence
Scout Before You Route: Predicting Task Success and Token Count from a Small LLM Prefill
Routers typically choose a model per query by trading predicted correctness against cost. Most price each model at a fixed cost, as if every query used the same number of output tokens, but the output length of reasoning models varies by an order of magnitude across queries. When is per-query cost worth predicting? What factors drive cost? How cheaply can predictions be made? We predict both the success probability and the output length of candidate models from the prefill activations of a single small frozen LLM, the scout. With reasoning candidates, perfect per-query cost knowledge would save 29-61% of spending on eight of nine datasets, and much less with non-reasoning candidates on the same problems. On those eight datasets, our predicted costs save 9-33% at matched accuracy relative to median-length pricing, up to 69% of what perfect knowledge would save. They are significantly cheaper than each estimator we reimplemented or adapted from published work, including ZeroRouter and a MixLLM-style predictor, on a majority of datasets. Output length depends mostly on the problem. On most math and code datasets, harder problems are longer, so predicted success alone prices cost nearly as well. On knowledge and mixed-format datasets they are not, and predicted success falls well behind a probe trained on length. The probe keeps most of its saving with 20 labelled problems and transfers without labels to four of seven new benchmarks. On a mixture of all nine benchmarks, it saves 22% beyond pricing each benchmark at its own median length. Savings stop growing at about a 2B scout, so above 4B a larger scout costs more than it adds.
Read the full abstract
Computation and Language
Language, Not Code Abstraction: Natural-language Training Shapes Brain-aligned Geometry
Code-trained language models produce positive brain scores, which raises the question of whether abstract structure learned from programming languages can substitute for natural-language training. We trained 25 GPT-2-small models on TinyStories, Python, C functions, x86-64 assembly, and the corresponding machine code, and scored them using a shared tokenizer and encoding procedure against matched random-weight models. On Pereira-243 and Tuckute2024, TinyStories exceeds the mean of the four code models by $\Delta r = 0.053$ and $0.059$ (Holm-corrected $p<0.001$), while code training leaves alignment near the random-weight level. The language advantage is benchmark-specific: Blank2014 shows no corrected language-code difference, and Fedorenko2016 shows no language benefit. Geometry tracks the encoding result on Pereira-243: TinyStories seeds converge on the most similar representations, and its fit-free correspondence with neural dissimilarity peaks at $\rho = 0.14$, twice the code and random-weight levels. Binary-trained models have about three effective dimensions, so seed agreement alone does not imply neural correspondence. Updating only the embedding matrix on English improves mean-over-depth alignment for every code model more than an equal update on its own corpus. Within this architecture and training setup, natural-language training produces a distinct representational geometry with closer neural correspondence on Pereira-243
Read the full abstract
Cryptography and Security
Beyond a Single Commit: Anchor-Guided Reconstruction of Multi-Commit Patches
Vulnerability patch identification plays a critical role in software security by linking reported vulnerabilities and the corresponding fixes, enabling reliable vulnerability analysis, mitigation, and automated security tool development. As software systems evolve in complexity, vulnerability remediation has increasingly shifted from single-commit fixes to multi-commit fixing processes. Although many studies have been proposed to identify fixing commits associated with target vulnerabilities automatically, they often suffer from inaccurate semantic analysis or excessive workload for language models. These limitations reduce both the effectiveness and comprehensiveness of fixing commit identification. To improve patch identification, we propose an effective and comprehensive patch identification framework, HARPOON, which not only enhances the accuracy of identifying individual fixing commits, but also investigates dependencies among commits to achieve comprehensive commit recognition for target vulnerabilities. Experimental results show that HARPOON significantly outperforms the state-of-the-art approaches. HARPOON successfully identified 895 vulnerability patches that were not previously maintained in public vulnerability databases. Among them, 663 are previously unrecorded fixes associated with 323 CRITICAL- or HIGH-severity vulnerabilities.
Read the full abstract
Materials Science
Symmetry-Gated Sliding: Decoupling Interlayer Translation from Polarization Reversal in Boron Nitride Ferroelectrics
Electric-field-driven polarization switching in sliding ferroelectrics depends on how local interlayer motion is accommodated by the surrounding structure. Yet the relationship between sliding, domain-wall (DW) propagation, and collective domain evolution under changing vector fields remains unclear. Here, we show that local symmetry governs the coupling between field-induced forces and interlayer sliding, while DW connectivity determines how local motion contributes to net domain conversion. These effects are examined across three structural scales, from monodomains and isolated DWs to reconstructed moire networks. We investigate bilayer hexagonal boron nitride (h-BN) using molecular dynamics with a machine-learning interatomic potential that incorporates field-dependent electric enthalpy. An initially monodomain bilayer switches through directional wrinkling and interlayer sliding, while the barrier along a constrained flat path remains finite. Pre-existing DWs propagate with velocities governed by their shear and tensile character, and an additional in-plane field tunes wall velocity and width according to its orientation. In reconstructed triangular moire networks, shared junctions and boundary curvature mediate domain evolution and recovery after field removal or reversal. Most strikingly, a rotating in-plane field translates the layers and displaces the moiré nodes by approximately 1 AA yet leaves the AB/BA domain-area ratio without detectable change, because threefold connectivity cancels the contributions of individual walls. We term this response polarization-silent sliding. These results distinguish local sliding, DW translation, and net domain conversion as different responses to electric fields and establish how local structural symmetry and DW connectivity shape their selective control.
Read the full abstract
Machine Learning
Knowledge Distillation from Neural Networks to Differentiable Logic Networks
Differentiable logic networks (DLNs) are interpretable, edge-friendly classifiers that learn both their logic operators and their connections through gradient descent, achieving accuracy competitive with multi-layer perceptrons at orders-of-magnitude lower inference cost. Their accuracy, however, is bounded by the problems of vanishing gradients, scalability, and lack of an ecosystem, leaving a gap relative to conventional neural networks on several datasets. We introduce knowledge distillation for DLNs to narrow this gap. We distill from a trained multi-layer perceptron teacher into a DLN student, integrating the distillation objective into a two-phase training loop without altering the architecture or the inference procedure. The teacher is discarded after training. Hence, the distilled model retains the interpretability, sparsity, and inference cost of a conventionally trained DLN. We investigate this approach through four variants: a joint hyperparameter search that optimizes the distillation weight and temperature alongside the architecture, a grid search over these two hyperparameters with the architecture fixed, and two single-phase variants that employ distillation during only the neuron-function or only the connection phase. On 20 public tabular classification tasks, distillation improves balanced-class accuracy at negligible cost to gate-level operation counts and model size, though the benefit varies by dataset. Performance improvements of up to 15% are observed, in some cases surpassing the teacher itself.
Read the full abstract
Optimization and Control
Constrained Nested-Logit Pricing: An Exact Reformulation and Global Optimization Method
Sellers must price substitutable products while meeting operational requirements such as price limits and sales targets. Nested logit captures stronger substitution among similar products, but these requirements can undermine the structure used to solve unconstrained pricing problems. We develop an exact reformulation and a global optimization method for a fixed assortment with price bounds and weighted price caps. The key novelty is to concentrate the difficult interactions at the nest level while retaining convex optimization over individual products, without discretizing prices or approximating demand. This structure makes the number of nests, rather than products, govern the exponent in a conditional worst-case complexity bound. The method returns feasible prices together with a bound on how much profit could still be gained, allowing decision makers to assess solution quality even when computation stops early. Extensions cover purchase-probability bounds and price-order rules, with explicit accounts of their computational cost. In synthetic experiments, the method certifies solutions for assortments of up to 500 products with few nests. Projecting unconstrained prices onto the feasible set loses up to 32.60% of profit relative to direct constrained optimization, highlighting the value of jointly optimizing substitution and feasibility.
Read the full abstract
Logic in Computer Science
Optimal Probabilistic Answer Set Programming: Complexity and Implementations
Probabilistic Answer Set Programming (PASP) extends Answer Set Programming with the possibility of representing uncertain data. Among the possible available semantics, we focus here on the credal semantics. In this setting, prior work has focused exclusively on plain ASP programs, i.e., programs without optimization constructs. However, many real world tasks define optimality criteria over the possible solutions, which cannot be conveniently modeled with existing probabilistic ASP frameworks. In this paper, we close this gap and propose optimal Probabilistic Answer Set Programming, which extends PASP by also allowing weak constraints, which are the standard optimization construct of ASP. We define the semantics of such programs, study the computational complexity of the inference and most probable explanation tasks, and propose practical implementations based on Algebraic Model Counting and Answer Set Programming with Quantifiers. An empirical evaluation of the proposed approach on three real world tasks confirms the applicability of our framework.
Read the full abstract
Combinatorics
Infinitely many convex polytopes with dihedral angles bounded above by $\pi/2+\delta$
For every $\delta>0$, we construct infinitely many pairwise combinatorially distinct convex three-dimensional polytopes whose dihedral angles are all less than $\pi/2+\delta$.
Computational Complexity
Reconfiguration of Time-Bounded Descriptions
Time-bounded Kolmogorov complexity measures a string by the length of its shortest program that prints it within a deadline. An adjunction-universal machine (Bauwens, Kozachinskiy, and Shen, arXiv, 2025) is a Turing machine that simulates every Turing machine $M$ on every input $z$ with polynomial overhead when given a code of $M$, computable from $M$ in polynomial time, followed by $z$ unchanged. Any two adjunction-universal machines give the same measure up to an additive constant and a polynomial change of the deadline. However, this measure records only the shortest program for a string. We study all programs of one length that print a string within one deadline, join two of them when they differ in one bit, and ask when this graph is connected. Fix any adjunction-universal machine. For almost all $m$-bit strings and every deadline beyond some polynomial, the graph is disconnected at program length $m+O(1)$ and has at least $2^{m-O(1)}$ components at length $2m+3$. In contrast, we build an adjunction-universal machine in which every such graph is connected or empty, at every deadline, once the program length reaches a computable threshold that is at most exponential in the length of the string. Exponential thresholds are unavoidable: for every adjunction-universal machine with such thresholds and every large $m$, more than three quarters of the $m$-bit strings need a threshold exponential in $m$. For our machine, above the threshold, every program for the string is within three bit flips of a connected set of fastest programs for the string, along a path inside the graph, at every deadline by which that program halts. For this machine, deciding whether two given programs are connected in this graph is $\mathsf{PSPACE}$-complete when the input is the string, the deadline in unary and the two programs; for each fixed string it takes polynomial time.
Read the full abstract
Signal Processing
Capacity of Multi-Layer Manhattan-Grid UAV Corridors Served by Terrestrial 5G NR Networks
Advanced air mobility (AAM) is expected to organize unmanned aerial vehicle (UAV) traffic into corridors at multiple altitudes. In this paper, we study the downlink (DL) and uplink (UL) capacity of a multi-layer Manhattan-grid corridor mesh served by a terrestrial fifth-generation (5G) new radio (NR) macro network. The corridor spacing and the numbers of layers, UAVs, and base stations are configurable. We combine the 3GPP aerial channel and antenna models, correlated UAV mobility, and frequency-selective fading with an NR physical-layer abstraction that includes channel aging and link adaptation. Results show that UAVs above 90 m experience a DL signal-to-interference-plus-noise ratio (SINR) about 17 dB below that of ground users, and that a dense macro network with $10^\circ$ downtilt leaves 17-29% of corridor positions without DL service under full-buffer load, regardless of corridor spacing. On a common layout, the carrier frequency (700 MHz, 3.5 GHz, or 7.2 GHz) changes the UAV SINR by less than 0.5 dB, since the corridor links are strongly interference-limited. The UL is the bottleneck: at 85% satisfaction, a 20 MHz carrier supports only about 3.4-4.1 UAVs/km$^2$ with 4 Mb/s of UL video per UAV.
Read the full abstract
Quantum Physics
Perfect Wave Transfer in Nonconformal Tomonaga-Luttinger Liquids
Perfect wave transfer (PWT) reproduces spatial reflection at a common finite time under a fixed Hamiltonian. We classify this process for open compact bosons and inhomogeneous Tomonaga–Luttinger liquids, distinguishing transfer of oscillator and neutral observables from transfer on the charged field algebra. In the single-charge model, the first two notions coincide. Global transfer additionally requires rephasing of the compact zero mode. For regular reflection-symmetric channels, exact oscillator transfer forces an odd multiple of the optical crossing time and a strictly ordered, parity-compatible spectrum differing from the conformal ladder at only finitely many levels. This spectral condition is also sufficient when the Robin boundary data induced by the physical profile are included. Requiring Neumann conditions after the Liouville transformation instead forces conformality. We construct an explicit, regular one-defect family for every $r>1/2$. Its transfer parameters are precisely the reduced fractions $r=p/q$ with odd numerator and denominator, with shortest transfer time $pL/v_0$. At the delay $3L/v_0$, the $r=3/5$ profile has numerical contrast $K_{\max}/K_{\min}\simeq 12.913$, compared with $81$ for $r=3$. Two independent conservative discretizations show second-order convergence on fixed mode bands. Occupation- and charge-constrained error estimates quantify robustness, while an exact homogeneous-chain calculation explains the $J^{2/3}$ resolved-band scale also observed in the engineered channels. These results provide a spectral design method and finite-device benchmarks for quadratic quantum channels. An interacting microscopic realization remains open.
Read the full abstract
Computer Vision
Frontier3DBench: Benchmarking Coding Agents for 3D Creation across Objects, Scenes, and Interactive Open Worlds
Coding agents combine visual understanding with program execution, offering a new route to automated 3D creation. By translating images and instructions into editable programs, they bring reconstruction, generation, and editing into a shared workflow. Yet existing benchmarks fragment evaluation across task scopes and protocols, leaving their capabilities across construction levels unclear. We propose Frontier3DBench, a unified benchmark of 145 tasks across 9 categories spanning object, scene, and interactive open-world creation, with joint evaluation of appearance, structure, and behavior. Experiments with 11 model-harness configurations reveal strengths that vary by level: GPT-6.1-sol leads the object and scene levels and achieves the highest overall score of 66.02/100, whereas GPT-6-astra leads interactive world construction. However, all configurations score lower on reference reconstruction than on text-driven generation in object and scene tasks. Producing plausible content thus remains easier than reproducing specified geometry, layouts, and visual details. Moreover, generated worlds average 83.08/100 on local mechanisms but only 20.82 on motion clearance and 30.26 on stateful workflows: individual mechanisms often function, whereas complete activities that require coordinated actions and persistent state remain difficult. These results identify reference fidelity and coordinated activity completion as central bottlenecks for coding agents in 3D creation.
Read the full abstract
Quantum Algebra
Twisted Verlinde Formula for Finite Cyclic Orbifolds of Vertex Operator Algebras
Let $V$ be a simple, rational, $C_2$-cofinite and self-dual vertex operator algebra of CFT type, and let $G$ be a finite cyclic automorphism group of $V$. We establish a twisted Verlinde formula for cyclic groups of arbitrary finite order, expressing fusion coefficients among twisted $V$-modules in terms of the $S$-matrix in the orbifold theory. As an application, we compute fusion rules among twisted modules for the permutation orbifold $\left(V^{\otimes k}\right)^{\langle g\rangle}$ associated with a $k$-cycle $g=(1\,2\,\cdots\,k)$.
Read the full abstract
Artificial Intelligence
SIMBAD: Spatio-Temporal Traffic Forecasting Robust to Aperiodicity
Accurate traffic prediction is essential for urban planning but remains challenging due to the irregular patterns caused by unexpected events. While Spatio-Temporal Graph Neural Networks (STGNNs) effectively model periodic data, they often struggle with these aperiodic fluctuations. To address this, we propose SIMBAD, which improves generalization irregular patterns via two key mechanisms: 1) adaptive regulation of periodic signals to prioritize recent signals, and 2) dynamic spatial modeling that adjusts influence between connected nodes based on contextual relations. Built upon simple similarity metrics to adapt to aperiodic patterns, SIMBAD outperforms existing methods on real-world benchmark datasets, demonstrating superior performance in predicting unseen and irregular traffic events.
Read the full abstract
Computer Vision
Did I See It or Think It? Probing Reality Monitoring in Vision-Language Models with Concept Injection
Recent work tests whether language models can introspect by injecting a concept direction into their activations and asking whether they notice it and can name it. These studies are text-only. Vision-language models (VLMs) face an additional problem, known in cognitive psychology as reality monitoring: they must keep internally generated content separate from what they actually see. We therefore ask: can VLMs distinguish injected concepts from image content? We introduce a paradigm that varies the image and the injected concept independently. On each trial, the model reports what it sees, whether a thought was injected, and what that thought is. This design scores both whether the model identifies the concept and which source it assigns to it. Across six VLMs, only Gemma-3-27B reliably identifies injected concepts (35% vs. 3% for steering alone), with almost no false alarms. Even then, the source it reports is fragile. Strong or visual-pathway injections are reported as seen, and a visible object nearly eliminates correct thought reports. The same failure appears in ordinary captioning, where injected objects appear in up to 87% of captions of real photographs. Evaluating multimodal self-reports therefore requires checking both their content and their reported source.
Read the full abstract
Machine Learning
KnowEvo: Knowledge Evolution for Protein Optimization
Large language models (LLMs) have shown promise in scientific discovery, yet systematically accumulating and refining reusable scientific knowledge remains challenging. Existing approaches primarily improve LLM-based problem solving through parameter optimization or feedback-driven refinement of prompts, programs, and solution trajectories, leaving acquired expertise either embedded in model weights or scattered across iterative problem-solving processes. As a result,knowledge is difficult to systematically inspect, evaluate, and refine. We propose KnowEvo, a framework that treats externalized knowledge itself as the optimization target. Our key insight is that scientific knowledge is inherently modular: multiple interacting mechanisms jointly determine the design outcomes. This makes knowledge evolution fundamentally challenging, as improvements cannot be attributed to individual components in isolation. To address this, KnowEvo represents knowledge as executable expertise blocks and organizes their refinements into strategy mutation trees, enabling structured exploration over a combinatorial knowledge space. The framework combines a Thompson-style forward allocation strategy that adaptively distributes evaluation budgets across strategy variants with a backward revision process, in which a code agent operates in an isolated sandbox to translate rollout evidence into localized program updates. This design enables persistent accumulation, evaluation, and refinement of mechanistic expertise across iterative experimentation. We evaluate KnowEvo on protein optimization tasks including stability and solubility. Results show that knowledge evolution consistently improves over initial strategy blocks and outperforms monolithic program evolution and prior LLM-guided evolution baselines across metrics such as pass@k, MRR, MAP,and NDCG@5. Code is available at https://github.com/recynie/KnowEvo.
Read the full abstract
Machine Learning
Towards Robust Reinforcement Learning Market Making: Lessons from Loss Function Experimentation
Market makers play a vital role in ensuring efficient trade within financial markets. Recent literature has delved into training market-making systems using reinforcement learning. This paper delves into the impact of various loss function formulations on risk management behavior within reinforcement-based market making. Through empirical evaluation, we contrast profit-maximizing strategies with risk-averse approaches, finding that penalties for inventory and cash management foster more resilient and adaptable market-making strategies. Our results underscore the significance of striking a balance between profit maximization and risk management in algorithmic trading, providing actionable insights for refining strategies in practical settings.
Read the full abstract
Numerical Analysis
On optimal convergence rates for parametrized variational problems with applications to nonlocal models
We develop an abstract variational framework for convergence rates of parametrized problems and their Galerkin approximations, with applications to nonlocal diffusion. The main objective is to quantify modeling and discretization errors in terms of the nonlocal horizon, the mesh size, the Sobolev regularity of the local solution, and the compatibility of boundary data. We establish error estimates in energy and $L^2$ norms and develop an adaptation of the Aubin–Nitsche argument that uses regularity of the limiting local dual problem. This approach yields improved $L^2$ estimates without requiring additional regularity of the nonlocal dual solutions, including for integrable kernels. We apply the framework to two constant-horizon nonlocal Laplacians with exterior boundary data and to heterogeneous localization models with classical Dirichlet, Neumann, and Robin conditions. The resulting estimates provide quantitative guarantees of asymptotic compatibility and improve existing error bounds under weaker regularity assumptions. Numerical experiments illustrate the convergence rates and examine the sharpness of the estimates.
Read the full abstract
Classical Analysis and ODEs
Walsh Analysis in the Twenty-First Century: Convergence, Summability, Nonlinear Approximation, Universality, and Digital Applications
Walsh analysis connects ordered Fourier convergence, summability, nonlinear approximation, and universal representation through a common character algebra. This survey examines developments from 2000-2026 against their classical foundations. Binary kernel geometry explains Hardy-space losses and summability thresholds; interval and packet methods control spectral selection; localized blocks produce correction and universal representations with distinct coefficient and exceptional-set requirements. Selected proofs and counterexamples compare these mechanisms and identify where an implication needs an additional hypothesis. The same distinctions organize digital integration, sparse measurements, Boolean spectra, and generalized Walsh bases. Quantitative thresholds, source conventions, and published theorem pointers are collected in comparative tables. The concluding problems separate source-listed questions, proposed refinements, and elementary obstructions. Supplementary derivations preserve normalization checks and finite covariance calculations while keeping the main exposition focused on the analytic mechanisms.
Read the full abstract
Analysis of PDEs
Critical Resonance and the Radial Branch in Maz'ya's Weighted Hessian–Gradient Inequality
We study the sharp constant $C(p,\gamma)$ in Maz'ya's planar weighted inequality \[ \int_{\mathbb{R}^2}|\nabla\psi|^p|x|^{p(\gamma-1)}\, \mathrm{d} x \le C(p,\gamma)\int_{\mathbb{R}^2}|D^2\psi|_{\mathrm{F}}^p|x|^{p\gamma}\, \mathrm{d} x, \] where $|\cdot|_{\mathrm{F}}$ denotes the Frobenius norm. For every fixed $1\le p<\infty$, with $a=\gamma-1+2/p$, we prove \[ C(p,\gamma)^{-1}=2^{p/2}|a|^p+O_p(|a|^{p+1})\qquad(a\to0). \] The leading term comes from the first angular harmonics after the Emden–Fowler transform. We also determine explicit exterior parameter rays on which the unrestricted constant equals the radial value $(1+a^2)^{-p/2}$. The proof uses an entropy identity and a homogeneous calibration. In the remaining transition region, we obtain symmetry-breaking intervals and a necessary and sufficient condition for nonnegativity of the periodic second variation at the radial state. An explicit example shows that global symmetry breaking can occur even when this second variation is nonnegative.
Read the full abstract
Distributed Computing
ISER-MoE: Dynamic In-Switch Replication of Hot Experts for Load Balancing in Expert-Parallel
Expert-parallel mixture-of-experts (MoE) layers run at the pace of the most loaded GPU, and a few hot experts carry most of the excess load. Replicating them at every MoE layer balances the load only if their weights arrive within the layer, and broadcasting from the owner is the bottleneck: point-to-point copies overload it, and baseline broadcasts pay latency or bandwidth overheads. We present ISER-MoE (In-Switch Expert Replication for MoE), built in TensorRT-LLM, which multicasts hot experts through the NVLink switch to fixed GPU groups, so each byte leaves its owner once. A GPU scheduler assigns hot experts to groups per layer, and a multi-port broadcast spreads the injection over the group's ports, both overlapped with the shared expert. In trace replay, the scheduler removed 98-99% of the excess load; in prefill serving of DeepSeek-V4-Pro, throughput rose 17.4% on a math benchmark and 7.2% on agentic prompts.
Read the full abstract
Number Theory
An Elementary Proof of the Binary Goldbach Conjecture
We establish the binary Goldbach conjecture which asserts that every even positive integer greater than four can be expressed as a sum of two odd primes. This goal is achieved by proving that the number of ordered Goldbach pairs that add up to $2n$ is an $\mathcal{O}\left(\frac{n}{\ln^{2} n}\right)$ function. As an application, we provide a simpler verification of the ternary Goldbach conjecture that was first established in [6,7].
Read the full abstract
Information Retrieval
Rethinking Latent Reasoning for Generative Recommendation: Beyond Teacher-Generated CoT and Recurrent Hidden-State Feedback
Latent reasoning in generative recommendation faces two practical barriers: constructing teacher-generated chain-of-thought (CoT) supervision is costly, and recurrent hidden-state feedback requires a specialized execution loop. We revisit both the supervision source and the computation interface. A step-level diagnosis measures how recommendation CoT supports next-item prediction and trace coherence. In the analyzed traces, history analysis is the only step with positive mean contributions on both dimensions, while later preference and intent steps have negative mean answer contributions. Guided by this finding, we propose PHiT (Prefilled, History-Informed Thought Tokens), which learns intermediate states from semantic targets extracted directly from observed item histories. PHiT aggregates cached sparse item encodings to construct these targets and supervises the contextual states of distinct vocabulary thought tokens. All thought tokens are supplied together as ordinary input IDs during a single causal prefill, removing the need for recurrent final-state feedback. With supervised fine-tuning alone, PHiT outperforms the compared baselines across four recommendation benchmarks. Controlled ablations support the value of input-matched semantic supervision, while cost measurements show fast target construction with training time comparable to the recommendation-only thought-token baseline.