Combinatorics
Papers
Fresh from arXiv, typeset properly. Highlighted words trade on the market.
All of arXiv
Tuesday 10 November
The directed Hamilton–Waterloo problem asks for a factorization of the complete symmetric digraph $K_v^*$ into directed $2$-factors, each of which is isomorphic to one of two given directed $2$-factors. In its uniform version, one seeks a decomposition of $K_v^*$ into $r$ factors consisting of directed $m$-cycles and $s$ factors consisting of directed $n$-cycles, where $r+s=v-1$. Let $m=3$, and let $n \geq 5$ be odd and not divisible by $3$. We show that, for every such $n$, the problem reduces to four decompositions of order at most $12n$: factorizations of $K^*_{3n}$ and $K^*_{6n}$ with every admissible number of factors of each kind, and two factorizations of the complete symmetric equipartite digraph with four parts of size $3n$. We construct these decompositions for $n=7$, and thereby give a complete solution for $(m,n)=(3,7)$: the obvious necessary conditions are also sufficient, with the single exception of $(v,r,s)=(6,5,0)$.
High Energy Physics (Phenomenology)
The DIS structure functions $F_4$ and $F_5$ at NLO and their impact on $\tau$-neutrino scattering
We present next-to-leading-order calculations of heavy-quark contributions to the deep-inelastic scattering structure functions $F_4$ and $F_5$ within the Aivazis-Collins-Olness-Tung scheme, implemented in the open-source library APFEL++ and evaluated using CT18NLO parton distribution functions. These structure functions are suppressed by lepton mass effects in light-lepton processes, but become important in $\tau$-lepton and neutrino scattering at facilities such as SHiP, IceCube, and DUNE. We furthermore quantify the direct impact of $F_4$ and $F_5$ on the charged-current $\tau$-neutrino cross section, for the first time presented here, showing they can suppress it by up to $-40\%$ near the kinematic boundary at low $Q^2$ and large $x$.
Read the full abstract
High Energy Physics (Phenomenology)
Self-interacting dark matter and neutrino masses from MeV-scale lepton-number breaking
We show that a scalar whose vacuum expectation value (vev) breaks lepton number at the MeV scale, generating the light-neutrino masses through the inverse seesaw, can also serve as the light mediator of self-interacting dark matter. Dark matter is a pseudo-Dirac fermion whose mass splitting arises from the same vev and is predicted to be proportional to the self-interaction coupling times the vev. A strong central interaction between the dark states requires this splitting to be sizable. The CP-even mediator decays into neutrinos at a rate fixed by the neutrino masses, independent of its mixing with the Higgs boson. Fixing the coupling by the relic abundance and evaluating exact partial-wave cross sections, the model reproduces the velocity dependence of the self-interaction cross section inferred from dwarf galaxies to clusters along a narrow band, $m_s\simeq1.9~{\rm MeV}\times(m_\chi/{\rm GeV})$, extending up to $m_\chi\simeq12$~GeV with a best fit at $m_\chi\simeq7.6$~GeV and $m_s\simeq15$~MeV, and on a sequence of resonant islands of comparable fit quality at higher masses. The cluster data bound the dark matter mass from above, so that the allowed masses extend from about $3$ to $70$~GeV ($95\%$ confidence level), with correlated mediator masses of $6$-$28$~MeV. Combining constraints from the effective number of neutrino species, SN~1987A and direct detection, we find the lepton-number-breaking scale confined to a window of a few hundred MeV and a strong preference for normal neutrino mass ordering. The allowed parameter space is testable by future measurements of $N_{\rm eff}$, low-threshold direct detection and the neutrino mass ordering.
Read the full abstract
Trading and Market Microstructure
Optimal Trading Speed on Blockchain
On blockchain-based exchanges, faster blocks weaken the priority-fee competition among arbitrageurs and shift rents from the block producer to them. Equilibrium liquidity peaks at an interior block speed; beyond a threshold, an active liquidity provider cream-skims and passive capital shrinks. Evidence from the Base Flashblocks upgrade, a tenfold block-speed increase, confirms the predictions. Priority-fee bids decrease sharply and arbitrageurs' net profit rises. The active liquidity provider rebalances faster, avoids toxic flow, and earns a higher profit. Pool depth falls and effective spreads widen despite closer tracking of external prices.
Read the full abstract
Computation and Language
Synapse-Control: Making Agent Experience Actionable with Lightweight Runtime Control
Large language model agents reuse execution experience to improve reliability on complex tool-based tasks. Systems such as Sentry call large models to generate failure diagnoses before retrieving advice, adding inference cost to experience reuse. We present Synapse-Control, an agent runtime whose lightweight control plane selects advice and determines when to send it to the task-executing model. Rules shortlist relevant advice, and a lightweight 1.7B classifier fine-tuned on teacher-labeled state-suggestion pairs scores those candidates instead of generating a diagnosis. Micro-contracts encode reusable interface checks, such as required arguments or valid options, without prescribing complete task programs. A Prompt Value Table records earlier guidance and feedback to limit repeated suggestions through cooldown and token-budget checks. Across 1,308 attempts per method in AppWorld and WebShop, Synapse-Control achieves 270 successes (29.8% more than Base), replaces 4,987 generative diagnosis calls, and reduces auxiliary control time relative to Sentry by 50.2% and 51.0%, respectively.
Read the full abstract
Audio and Speech Processing
Emilia2: A Large-Scale Dataset for Controllable Long-Form Speech Generation
Generating long-form speech requires models to keep content, speaker identity, turn-taking, and expressive style consistent over time. Yet most open training corpora contain examples lasting only tens of seconds and open long-form benchmarks cover at most two speakers per recording. We introduce Emilia2, a 229K-hour multilingual dataset for controllable long-form speech generation. From in-the-wild data, Emilia2 constructs short utterances, single-speaker long-form speech, and multi-speaker dialogues, preserving transcripts, speaker identities, pauses, and acoustic transitions. Examples longer than five minutes account for 50.0% of single-speaker long-form hours and 66.8% of dialogue hours. To support expressive control, Emilia2 provides natural language instructions for 4.77M examples and nonverbal-vocalization tags for 933K examples. We also introduce Emilia2-testset, 23 hours of bilingual recordings of up to 30 minutes and four speakers, each speaker paired with reference audio. We train autoregressive (AR) and non-autoregressive (NAR) models on Emilia2. Our 0.6B-backbone NAR model achieves the lowest Chinese CER and English WER on TTSD-Eval (5.27%/9.93%) among compared systems, while our AR model is on par with Qwen3-TTS-12Hz-VD on InstructTTSEval (zh/en: 76.2/78.3 vs. 77.1/77.9). These results validate Emilia2 as a comprehensive data foundation for long-form speech generation and minute-scale expressive control. Demo: https://charlesnii.github.io/Emilia2-demo/.
Read the full abstract
Artificial Intelligence
Megrez: A Knowledge-Guided Agent for Formal Mathematics over MK Set Theory
General-purpose coding agents equipped with proof-assistant tools have shown strong performance on standalone competition problems, yet cumulative textbook formalization inside a low-resource mathematical library still faces three obstacles: aligning textbook concepts with library declarations, keeping long-horizon search synchronized with live proof states, and integrating locally verified proofs reliably into a growing codebase. We present Megrez, a knowledge-guided formalization agent for Rocq libraries built on Morse–Kelley (MK) set theory. Megrez coordinates a general coding agent through three mechanisms: (1) a source-rebuildable knowledge graph merging formal declarations, compiler references, and curated metadata to support retrieval via Chinese aliases, concept clusters, proof dependencies, and multi-lane full-text ranking; (2) a feedback-driven 1–5 tactic short-burst proving loop paired with a goal-local branch ledger and exponential-interval structured reflection; and (3) a three-check acceptance pipeline (statement fidelity, axiom auditing, and full builds) coupled with a multi-corpus Provider stack and a candidate-bound recoverable batch state machine. Deployment and evaluation on a point-set topology corpus demonstrate effective cross-chapter formalization and scalable premise retrieval.
Read the full abstract
Artificial Intelligence
Earth Embeddings in Environmental Remote Sensing
Earth observations from satellite platforms have profoundly transformed research and applications across numerous domains. However, uneven observation coverage and frequent data gaps continue to limit their ability to capture Earth's surface dynamics. Moreover, processing multi-modal Earth observations requires substantial computational resources and sophisticated data harmonization frameworks, which may not be readily accessible to many remote sensing practitioners. Earth embeddings have emerged as promising solutions to these challenges. As general-purpose representations converted from multi-source Earth observations, Earth embeddings can support a wide range of downstream tasks with no requirement of task-specific feature design and preprocessing. In this study, we first introduce the theoretical foundations of Earth embeddings. Then, we review several representative Earth embedding products. We next examine the applications of Earth embeddings across six major domains of environmental remote sensing. Finally, we systematically assess the performance of Earth embeddings in four downstream tasks, including hyperspectral image classification, individual tree crown segmentation, crop yield prediction, and aboveground biomass mapping. Our review and case studies suggest that, despite several advantages over conventional Earth observation data, current Earth embeddings do not always outperform remote sensing imagery or hand-crafted features. Earth embeddings should not be viewed as a direct replacement for conventional EO data, but rather as representations that provide stable information across space and time. Remote sensing practitioners may consider evaluating Earth embeddings' suitability for specific tasks and using them strategically, rather than assuming superior performance by default.
Read the full abstract
Materials Science
How anion composition governs structural frameworks and Li-ion transport in oxychloride solid-state electrolytes
Mixed-anion lithium oxychlorides exhibit high Li conductivity, but the relationship between oxygen content, structural framework, and Li-ion transport remains unclear. Here, crystal structure prediction and molecular dynamics simulations with machine-learning interatomic potentials are used to examine $\mathrm{Li}_yM\mathrm{O}_x\mathrm{Cl}_{6-2x}$ ($M$ = Nb and Ta [Group 5] and $M$ = Ti, Zr, and Hf [Group 4]). Increasing oxygen content changes isolated chloride-rich polyhedra into one-dimensional chains and subsequently more highly connected two- and three-dimensional frameworks. In the Group 5 series, fast Li diffusion occurs over $x \approx 0.67$–1.5, where the Li sites form a three-dimensional percolating network. Potential of mean force maps and nudged elastic band calculations reveal that fast Li-ion transport requires a low-dimensional framework supporting long-range migration through undercoordinated Li sites with low barriers, together with similar energies across Li sites. Finally, we perform fine-grid compositional search, which identifies $\mathrm{LiNbO}_{0.67}\mathrm{Cl}_{4.67}$, $\mathrm{LiTaO}_{0.75}\mathrm{Cl}_{4.5}$, $\mathrm{LiNbO}_{1.5}\mathrm{Cl}_3$, and $\mathrm{LiTaO}_{1.5}\mathrm{Cl}_3$ as stable candidates with predicted room-temperature self-diffusivities of up to $\sim 10^{-6}\,\mathrm{cm}^2/\mathrm{s}$. These findings establish how oxygen content controls framework and Li-ion transport, guiding the design of lithium oxyhalide electrolytes.
Read the full abstract
Robotics
From Egocentric Human Video to Robot Manipulation in the Era of Embodied Intelligence
Egocentric human video (EHV) records close-range hand-object interaction (HOI) and the temporal structure of everyday activities, offering a potentially scalable source of experience for robot manipulation. Using this experience for robot control requires recovering useful supervision without recorded robot actions and accounting for differences in morphology, viewpoint, and action space. We define Ego2Robot learning as the process of learning or adapting robot manipulation policies from EHV, organized into four learning paradigms: explicit embodiment transfer, explicit action and task supervision, egocentric video representation pretraining, and latent action and dynamics modeling. The four paradigms differ in what they derive from human video: robot-format demonstrations, interpretable supervision, visual representations, or predictive models of actions and dynamics. For each paradigm, we trace what EHV information is extracted, how robot policies use it, and when the additional data is required. Besides, we review the benchmark datasets that cover EHV, robot demonstrations, and synthetic and simulated resources. This paper links EHV data's contribution to the task information preserved and the robot-specific grounding used during training or deployment. Finally, we identify challenges in physical feasibility, agent coordination, in-context adaptation, controlled evaluation, and predictive control.
Read the full abstract
Computation and Language
Post-OCR Correction of Mixed-Script Uzbek Text: Byte-Level ByT5 with an Empirical Homoglyph Noise Matrix
In Uzbek official and technical documents, a large share of OCR errors is caused not by poor image quality but by characters that are visually identical across the two alphabets in use. Because the script transition is still under way, Cyrillic and Latin appear on the same page and often within a single sentence, and in our measurements the density of homoglyph errors on such pages is close to five times higher than in single-script text. We study this problem in a setting where the corpus of engineering-geological and ICT infrastructure reports may not leave the organisation, so every stage was executed on local hardware. The proposed system combines a byte-level ByT5 model with a noise generator whose sampling weights are estimated from 359 manually classified error events, together with a Levenshtein-bounded conservative decoding rule. On a Gold set of 5,204 line pairs it lowers the character error rate from 0.0578 to 0.0369, restores 91.3% of the homoglyphs and introduces only 84 new errors.
Read the full abstract
Combinatorics
Triangles without a common vertex
Let $G$ be a graph with $m$ edges, $t(G)$ triangles, and adjacency spectral radius $\lambda(G)>\sqrt m$. Li, Feng and Peng (2025) conjectured that $t(G)\ge\sqrt m-O(1)$ if no vertex belongs to every triangle of $G$. We prove the stronger bound $t(G)\ge\lambda(G)-C$ for an absolute constant $C$, and give an explicit choice of $C$. The proof bounds the sum of the squared adjacency eigenvalues other than the largest and smallest. Interlacing and matrix perturbation estimates then restrict the neighbourhood of a vertex corresponding to a largest entry of a Perron vector to two disjoint edges and isolated vertices. A weighted count of the triangles completes the proof.
Read the full abstract
Computation and Language
One Byte, One Token: Byteifying a 27B Hybrid-Attention LLM with Self-Generated Data
Byte-level BPE fragments Chinese text (for Qwen3.8 on Chinese Wikipedia, 3.84 bytes per token on average; 64% of tokens are single characters) and hides spelling from the model. We ask whether an already-trained 27B subword model can be converted directly into a pure byte-level model (one byte, one token), without a patching structure or a local encoder, while keeping most of its ability. Starting from Qwen3.8-27B (48 Gated DeltaNet layers + 16 full-attention layers), we shrink the vocabulary from 248,320 to 384 entries (the 256 byte symbols already in the BPE vocabulary plus special ids), leave every other weight untouched, and continue training in a two-stage curriculum (encoding first, then reading) on 4.77 GB of supervised bytes: all dialogue/behavioral supervision is generated on-policy by the original model itself (character dictionary, word dictionary, general dialogue, coding sessions), and knowledge reading uses encyclopedia text (54% of supervised bytes). Training takes 178.6 h on 8xH200. On the same GPU and serving stack, the byte model's gap to the original falls within the paired bootstrap interval on MATH-500, HumanEval+, MMLU-Pro and non-thinking GSM8K, and it loses 4-6 points on C-Eval, thinking-mode GSM8K and non-thinking IFEval. Without thinking it leads on every character-level task (CUTE 75.2 -> 84.0, EXECUTE-zho 81.6 -> 93.9, our Chinese character suite 54.4 -> 64.0), but lookup knowledge about Chinese characters such as pronunciation and stroke count drops, and thinking does not recover it. The cost is throughput: per-step speed is identical, but each step yields 3-5x less output; Chinese characters per second is about 1/4 of the original, and prefill at 64K bytes is 6.5x slower. We also document four byte-model failure modes (greedy runaway, sampling breaking UTF-8, ineffective penalty terms on a small vocabulary, repetition) and their fixes.
Read the full abstract
Number Theory
A Threshold-Crossing Condition for the Least Colossally Abundant Counterexample to Robin's Inequality
We study the constraints that a hypothetical least Colossally Abundant (CA) counterexample to Robin's Inequality would place on the non-trivial zeros of $\zeta(s)$. In the single-prime case we express the least CA counterexample $N$ as $M\cdot p$, where $M$ is its CA predecessor and $p$ is prime. Using the von Mangoldt explicit formula for $\psi(x)$, we express $\ln p$ in terms of a sum over the non-trivial zeros of $\zeta(s)$ and split that expression into its baseline contribution $\mathcal{R}_0(p)$ and its off-line distortion. When both converge, we show that the off-line distortion is negative whenever $1+1/(e^{\mathcal{R}_0(p)}+e^{2\mathcal{R}_0(p)}+\cdots+e^{(v+1)\mathcal{R}_0(p)})\le\ln(\ln M+\mathcal{R}_0(p))/\ln\ln M$, where $v$ is the $p$-adic valuation of $M$. In the case where the least CA counterexample is the product of its CA predecessor and two primes (the semiprime case), the logarithm of the product of the two primes is the sum of two von Mangoldt expansions. An analogous inequality implies that at least one of the off-line distortions of the two expansions is negative.
Read the full abstract
Analysis of PDEs
Existence and uniqueness of asymptotic traveling waves in a time-fractional bistable reaction-diffusion equation
We study asymptotic traveling waves for a reaction-diffusion equation with a Caputo time derivative. For general $C^1$ bistable imbalanced nonlinearities, we prove the existence of an asymptotic traveling wave connecting the two stable states. The asymptotic wave speed is unique, and the asymptotic profile is strictly increasing and unique up to translation. The proof uses continuation from a stationary front and the sliding method. For a cubic nonlinearity, we also derive an asymptotic formula for the speed near the balanced case, with an explicit leading coefficient. Finally, we perform numerical simulations to compare the speed predicted by the asymptotic formula with the front propagation speed in the original equation.
Read the full abstract
Number Theory
A counterexample to the Erdős–Mahler conjecture
We disprove a 1939 conjecture of Erdős and Mahler by constructing, for every real $\tau\ge4$, a Cantor set of numbers of irrationality exponent $\tau$, each having infinitely many regular continued-fraction convergents of the form $2^a/3^b$.
Commutative Algebra
Relative Derived Torsion in Categories of Reduced Modules: Comparison and Non-Noetherian Examples
Let $I$ be an ideal of a commutative ring $R$. We use relative derived torsion in exact full abelian subcategories of $I$-reduced modules to study ambient derived ideal-power torsion. Exact inclusion forces $IM=I^2M$ whenever $I^2$ is finitely generated. Over a Noetherian ring, every finite object consequently decomposes into its torsion part and $IM$, and its positive-degree ambient cohomology vanishes. For $R=S[t]$ and $I=JS[t]+(t)$ with $J^2=J$, we compute every stage of the canonical Ext system and its transition maps, obtaining a comparison with $\Ext_S^q(S/J,-)$. Valuation-domain and product-of-fields examples yield nonzero cyclic first cohomology. A further construction has cyclic coefficients and cyclic ambient cohomology in every positive even degree, while its positive-degree relative cohomology vanishes. Applications determine kernels of canonical Ext maps for these coefficient modules and identify the obstruction to ring-valued specialization in the questions of Eisenbud, Mustaţă, and Stillman.
Read the full abstract
Nuclear Theory
Nuclear medium and final state interaction effects in $\eta$ production induced by $\nu_\mu(\bar\nu_\mu)$ at SBN and DUNE energies
We investigate nuclear-medium modification and final-state interaction in charged- and neutral-current $\nu_\mu(\bar{\nu}_\mu)$-induced $\eta$-meson production on $^{12}\mathrm{C}$ and $^{40}\mathrm{Ar}$. The nuclear medium is described within a local density approximation, incorporating in-medium modifications of the $N^*(1535)$ resonance, which is known to play the dominant role in $\eta$ production through its $N \eta$ decay channel. The subsequent propagation and absorption of the $\eta$ meson in the nucleus are simulated using a Monte Carlo approach. We find that nuclear-medium modifications and final state interaction lead to substantial suppression of the production cross section, highlighting the importance of nuclear effects in interpreting $\eta$ production data. Our results provide valuable theoretical input for the analysis of $\eta$-production at the ongoing SBND and MicroBooNE experiments and the forthcoming DUNE experiment.
Read the full abstract
Representation Theory
Regular loci of affine Springer fibers at very special parahoric level
For split groups, we prove that centralizers act transitively on regular loci of affine Springer fibers at very special parahoric level. As an application, we prove Chi's conjecture on the number of top-dimensional irreducible components of affine Lusztig varieties.
Computational Geometry
Parameterized Complexity and Geometric Approximation of Riesz Energy Subset Selection
We choose $k$ of $n$ sites to minimize Riesz $s$-energy, in which a pair at distance $t$ contributes $t^{-s}$. For $s=2$, exact selection is polynomial on a line but NP-hard in the plane. Adapting the proxy lemma of Ceccarello et al. gives an efficient parameterized approximation scheme (EPAS) for fixed Euclidean dimension and fixed $s>0$. To sharpen the bound, we start with a geometric packing. Farthest-first traversal gives $2k$ centers separated by at least the last insertion distance $r$. If one site of an optimum is removed, each of the $k-1$ remaining sites excludes at most one center closer than $r/2$, leaving at least $k+1$ centers free. The new step is a discrete exchange bound: the removed site's interaction energy is at most the average it would contribute at those free centers. In the plane at $s=2$, this reduces the representative bound from $O(k^2\log(2k)/\varepsilon^2)$ for a packing estimate of total energy to $O(k\log(2k)/\varepsilon^2)$, independently of $n$. Rational planar examples show that the resulting separation scale $r/\sqrt{\log(2k)}$ is sharp. A $(1+\varepsilon)$ approximation on rational planar sites takes $O(nk+n\log n)+(C\log(2k)/\varepsilon^2)^k$ operations of polynomial bit cost. The observed number $M$ of representatives determines the search size $\binom{M}{k}$. With one million sites, $k=5$ and $\varepsilon=1/2$, the exchange grid left 252 subsets on clustered inputs and about $9.3\times10^{26}$ on uniform inputs. The method extends to metrics of bounded doubling dimension and to compact domains with effective geometric routines, including filled simple polygons. At $s=2$, exact decision on metrics with nonzero distances in $\{1,2\}$ is W[1]-complete in $k$. A gap that shrinks with $k$ rules out a deterministic EPAS on unrestricted metrics unless FPT = W[1].
Read the full abstract
Machine Learning
Unified Optimization of Traffic Signal and Lane Channelization for an Isolated Intersection Using Reinforcement Learning
Optimizing traffic signal control has been a long-lasting critical issue in traffic engineering, and a number of robust methodologies have emerged to solve the problem so far. On the other hand, the road design for lane channelization is a static problem that is difficult to solve jointly with the real-time traffic signal control. If traffic volume does not vary, it is possible that both problems can be solved in an integrated manner as a single static optimization problem. Recently, researchers have devised a mixed integer linear problem (MILP) to jointly determine the traffic signal and road width designs. However, this joint optimization problem must be solved repeatedly whenever the traffic flow pattern varies. Solving the MILP may also take impractical computing time as the problem size gets larger. We present a robust methodology for solving the joint optimization problem using reinforcement learning (RL) on a real-time basis. Abrupt changes in lane markings with an incompatible signal phase can jeopardize drivers. To sidestep such a dangerous action change, we train a model with human feedback to filter infeasible action transitions in a rule-based manner. We devised a deep Q-network that can accommodate joint actions for traffic signal and lane marking plans. To ensure the implementation of the proposed operation, a variable sign for lane markings is necessary at each stop line of an intersection. The experimental results confirm that the proposed methodology enhances the intersection performance by increasing throughput.
Read the full abstract
Classical Analysis and ODEs
Weak $L^{2/3}$ bounds for the bilinear Hilbert transform
For Schwartz functions $f_1,f_2$ on $\mathbb{R}$, define the bilinear Hilbert transform by $$ \mathbf{BHT}(f_1,f_2)(x) =\operatorname{p.v.}\int_{\mathbb{R}} f_1(x-t)f_2(x+t)\,\frac{dt}{t}. $$ We prove the weak-type estimate $$ \left\|\mathbf{BHT}(f_1,f_2)\right\|_{L^{2/3,\infty}} \lesssim_{p_1,p_2} \|f_1\|_{p_1}\|f_2\|_{p_2} $$ for $1<p_1,p_2<2$ with $1/p_1+1/p_2=3/2$. The key ingredient is a new local supremum estimate for packet sums, which we use to establish the forest estimate.
Read the full abstract
Computation and Language
Compute the Interactions, Not the Attributes: What a Fact Layer Must Supply in the Chinese Sexagenary System
Language-model systems in rule-governed domains often pair the model with a deterministic fact layer, but what that layer should compute is rarely measured. We measure it in the Chinese sexagenary system underlying BaZi chart analysis, where every fact has one table-checkable value: 682 items over twelve fact categories and 160 transit–natal collision items, each posed on a ladder of rungs that differ only in how much of the computation is done for the model. Answering directly, none of eleven open-weight models gets more than 4% of sexagenary days or 1% of collisions right, and supplying the attributes of every pillar does not change the collisions. A complete list of computed relations is not enough either: it leaves Qwen3-32B at 1%, with relations listed for other pairs misattributed to the queried pair, whereas a query-filtered layer from follow-up runs, stating only the queried relations pillar by pillar, gives 92%. On chess positions, two predictions locked in advance hold: attributes do not help, and unqueried relations are again misattributed. Three proprietary reasoning models compute 68–95% of collisions unaided, while reasoning in smaller open models overrides supplied facts. A fact layer should compute the interactions and return the ones the question asks about.
Read the full abstract
Artificial Intelligence
Towards Reliable VLM Judges: State-Conditional Invariance and Presentation-Aware Diagnostics
Vision-language models (VLMs) are increasingly used as automated judges for multimodal and embodied tasks, yet aggregate accuracy or consistency alone does not reveal whether a judge changes its verdict for the right reason. We propose State-Conditional Invariance (SCI), a reliability principle requiring a judge to remain stable under semantics-preserving changes in presentation while responding to causal interventions that change the task outcome. We instantiate SCI in GridWM-Judge, a simulator-grounded diagnostic benchmark built from deterministic MiniGrid trajectories with aligned Full, evidence-removal, and counterfactual variants, plus presentation and representation probes. Across five diagnostic tasks, GridWM-Judge tests local perception, action-conditioned transitions, presentation stability, representation exchangeability, and Full-CF causal discrimination. Our diagnostic evaluations reveal heterogeneous failure modes, including sensitivity to evidence presentation and encoding, low activation on successful trajectories, and weak paired counterfactual discrimination. Perception-controlled analyses further indicate that visual grounding and objective recovery are substantial, model-dependent bottlenecks. These results motivate jointly reporting correctness, presentation stability, and causal sensitivity instead of relying on a single score.
Read the full abstract
Computation and Language
IFRS Compliance Verification: An Open Multi-LLM Benchmark and Content-Conditioned Ensemble Classifier
Verifying corporate financial disclosures against the International Financial Reporting Standards (IFRS) is a core auditing task in which experts cross-reference disclosures with hundreds of fine-grained regulatory paragraphs: a laborious, costly, and error-prone process that motivates growing interest in large language models (LLMs) as compliance assistants. Yet single-model accuracy remains insufficient for high-stakes use, and naive ensembling is vulnerable to systematic per-model biases that can drag aggregate accuracy below that of the strongest individual annotator. This gap is addressed with a multi-stage framework, supported by a new benchmark of 19,998 IFRS regulation-disclosure pairs comprising 11,161 unanimous multi-LLM labels and 8,837 human annotations on items where the LLMs disagreed. The framework couples a three-stage domain-adapted sentence embedding pipeline (BERT masked language modeling on SEC EDGAR filings, unsupervised TSDAE and SimCSE adaptation, and supervised contrastive alignment) with a learned ensemble classifier that fuses frozen embeddings and LLM probability features from GPT-4.1, Claude Sonnet 4, and Gemini 2.5 Pro via Tree-structured Parzen Estimator (TPE) optimized gradient-boosted trees. On the disagreement set, the ensemble achieves 81.7% weighted F1, exceeding the strongest single LLM by 10.4 points and majority voting by 14.6 points, and outperforms eleven unsupervised aggregation baselines; the retrieval pipeline attains Recall@15 of 45.3% on unseen regulations. The dataset, model checkpoints, annotation platform, and classifier implementation will be publicly released.
Read the full abstract
Artificial Intelligence
Transformers Provably Learn Graph Search: Training Dynamics and the Exponential Power of Depth
Large Language Models (LLMs) have achieved remarkable progress on complex reasoning tasks, yet the mechanisms by which they acquire reasoning capabilities and the efficiency of their reasoning remain poorly understood. In this paper, we investigate the path-finding problem over directed graphs, a symbolic abstraction of multi-step reasoning. We first characterize the architectural requirements for this task, proving that one-layer transformers require $\Omega(N^2)$ size on dense graphs to implement the key DFS child-selection primitive. While a two-layer transformer implements full DFS with $O(N\log(N))$ size. Thus, a second layer is necessary and sufficient for near-linear DFS reasoning. We then extend this expressivity analysis to deeper models: for every $k\geq3$, a $k$-layer transformer implements a bounded-reachability DFS policy with look-ahead horizon $1+2^{k-3}$, establishing an exponential increase in reachability horizon with depth. Finally, going beyond expressivity, given appropriately curated training data, we show that two-layer transformers can provably learn to execute DFS via gradient flow and generalize to unseen graphs. Together, our results provide a mechanistic account of how transformers implement graph search, how depth expands bounded look-ahead, and how such reasoning algorithms can emerge through training.
Read the full abstract
Combinatorics
Nonnegative weak tilings and the Coven-Meyerowitz condition (T2)
We develop a general method, in finite cyclic groups, for constructing nonnegative rational functions $f$, normalized by $f(0)=1$, that satisfy the Coven-Meyerowitz condition $(T1)$ but fail $(T2)$, such that $f$ is a weak complement for a set tile $A$. Motivated by the unsupported-divisor questions in arXiv:2507.11672, we show that if a set tile $A$ admits an unsupported divisor, then a suitable dilation of $A$ admits such a rational weak complement $f$. Using this method, we construct an explicit example in the group $\mathbb{Z}_{180}$. The weak complements in our explicit perturbative constructions are not positive definite because of restrictions on the Fourier support of the perturbations. Thus these constructions do not resolve the proposed extension of the Coven-Meyerowitz conditions to positive-definite weak tilings (arXiv:2411.03854, Problem 1.2).
Read the full abstract
Robotics
Few-Shot Residual Adaptation of Pretrained Latent World Models
Pretrained latent world models can support visual planning and control, but their predictions may fail when physical dynamics change at deployment. We study whether such models can adapt from only a few interactions without modifying their pretrained weights or identifying the changed physical parameters. We introduce a few-shot residual adaptation approach that learns self-supervised corrections from latent prediction errors. The corrections are applied both inside the latent predictor and directly to its predicted latent state, while the pretrained world model remains frozen. We evaluate the approach in simulation and on a real robot under center-of-mass and friction shifts. These changes reduce the success of the frozen model to near zero. Using three interaction episodes for adaptation, our dual-site approach recovers 90% success under the center-of-mass shift and 100% under the modified friction on the real robot. Ablations show that correcting both locations outperforms single-site adaptation. Our diagnostics further suggest that much of the learned correction reduces a general prediction mismatch, while successful control also depends on shift-specific information captured during adaptation.
Read the full abstract
Mesoscale and Nanoscale Physics
Topological flat bands from nonlinear periodic motion
We use the Lax formulation of classical integrable systems to characterize the topology of nonlinear periodic orbits. For a Hamiltonian system with one degree of freedom, we express the chiral-symmetric Lax operator along a periodic orbit in a frequency basis. Energy conservation makes both bands of the resulting tight-binding model exactly flat. The harmonic oscillator corresponds to a fully dimerized topological Su-Schrieffer-Heeger model, while nonlinearity generates long-range hopping through higher harmonics. In the symmetric double well, the winding number changes as intrawell periodic orbits give way to an orbit spanning both wells, and the zero-energy edge mode of the semi-infinite chain disappears. We derive the edge-mode profile and show that its localization length diverges logarithmically as the separatrix is approached. At the separatrix, the period and the hopping decay length diverge while the band gap remains finite; this singular limit of the periodic motion allows the winding number of the infinite frequency chain to change. Canonical quantization of the Lax operator also connects the chiral winding numbers to supersymmetric quantum mechanics: for a bound system on the real line with simple turning points, their sum over all periodic orbits at a common positive energy equals the Witten index. We also apply the construction to periodic nonlinear waves through traveling-wave reductions. The finite-period winding number becomes undefined in their solitary-wave limits.
Read the full abstract
Dynamical Systems
Fractional-Order Hybrid SIR Forecasting of COVID-19 in Bangladesh and the United States: Memory Effects and Out-of-Sample Evaluation
Fractional-order epidemic models are often justified by memory, yet whether memory improves forecasts is rarely tested out of sample. We extend a hybrid SIR framework, in which a daily transmission rate is reconstructed from reported cases and then forecast, to Caputo fractional order $\alpha\in(0,1]$. We prove well-posedness, conservation, positivity and a linearized threshold at disease-free states, and derive a Grünwald-Letnikov reconstruction that reduces exactly to the integer-order one at $\alpha=1$. Our main theoretical result shows that for $\alpha<1$ the influence of the unknown initial infectious count vanishes, but no faster than $t^{-\alpha}$; after 332 daily steps it is about $4.4\times10^{8}$ times larger at $\alpha=0.9$ than at $\alpha=1$. In out-of-sample forecasting over 27 weekly origins, the effect of memory depends on the setting. In Bangladesh the integer order gives the lowest error for linear and constant-rate forecasts in 47 of 48 settings. In the United States, where reported cases reached about 18% of the population, a fractional order lowers the linear forecaster's error in all 12 comparisons in 2021 but raises the constant-rate forecaster's, with no difference significant after Holm adjustment. Against standard baselines, the SIR-based linear forecaster was more accurate than the classical SIR model and a neural network without SIR structure, and competitive with ARIMA: in the United States the fractional linear forecaster had the lowest error of all methods (seven-day MAPE 7.54% against 9.14% for ARIMA), whereas in Bangladesh ARIMA had the lowest MAPE, with no significant difference from ARIMA in either country. The choice of a fractional order should rest on out-of-sample evidence rather than in-sample fit.