Papers

Reinforcement learning with verifiable rewards (RLVR) provides reliable supervision for software engineering agents through executable tests, but binary rewards fail to distinguish trajectories with substantially different debugging quality. We investigate whether the current policy can judge its own trajectories and use these judgments to improve task performance without a separately trained reward model. To this end, we introduce Behavioral Self-Rewarding (BSR), a reward-refinement method that reuses the policy to compare observable debugging behavior among trajectories from the same task with the same verifier outcome. BSR aggregates pairwise preferences into task-local scores and converts them into bounded auxiliary rewards, introducing reward contrast while preserving the verifier-defined pass/fail ordering. This provides additional supervision from already-collected trajectories without requiring new environment interactions. Experiments on SWE-bench Verified and SWE-bench-Live show consistent improvements across the evaluated Qwen3-14B and Qwen3-32B configurations, supporting using the policy's own comparative judgments as a complementary learning signal for training software engineering agents.

Quantum error-correcting codes on graph geometries beyond regular euclidean lattices have emerged as a promising route toward robust quantum information processing. A major challenge in quantum error correction is the trade-off between favorable properties of codes constructed in high dimensionality and the need for locality in practical implementations on near-term hardware. Sparse intermediate geometries may combine the advantages of both, but how their structure governs quantum information dynamics and code performance remains unclear. Here we show that the spectral dimension $d_s$, defined from the low-energy spectrum of the graph Laplacian, is the universal parameter connecting these properties. Using Floquet random Clifford circuits on two distinct sparse graph architectures with tunable spectral dimension, we find that $d_s$ governs both quantum information spreading and the scaling of the resulting quantum error-correcting code distance. We identify a geometry-induced entanglement phase transition from localized area-law to ergodic volume-law, whose universality is determined by $d_s$ and exhibits long-range percolation criticality. In the ergodic phase, the circuit acts as a CSS code encoder whose code distance is set by the competition between two logical sectors — the cost of severing the graph and the length of its shortest winding cycle — which connectivity pushes in opposite directions. The spectral dimension organizes this trade-off, highlighting a maximum at intermediate $d_s$ retaining sparse connectivity and physical realizability.

Agentic large language models now tackle complex tasks, yet strategic reasoning over long horizons is under-studied: most agentic benchmarks are short and episodic, and performance on long tasks is hard to dissect into stored knowledge, reasoning and luck. We introduce SpireBench, in which models play the complete commercial deck-builder Slay the Spire (StS), where early decisions decide long-term success. Across 20 large language models (LLMs) at the game's highest difficulty, frontier models beat heuristic baselines and match the average human player but remain far behind experts. This gap widens as the run lengthens: the strongest models win the first act as often as experts, then fall behind. What separates strong models from weak is how well they play each fight; all of them misjudge risk over a run, preserving health rather than spending it to grow stronger. A comprehensive, self-consistent renaming of every game term shows that part of their performance rests on prior knowledge of the game, and models do not improve from the memories they write across runs, which are often wrong and overly cautious. Our harness is the first to give a model the complete original game, and we compare it with existing harnesses. Overall, SpireBench offers an unsaturated long-horizon benchmark with tools to locate where and why agents fail.

We present the first inference of the Hubble constant and matter density parameter from a strongly lensed supernova by exploiting the mass model of a galaxy cluster enhanced by an extended surface brightness reconstruction. Building upon the recent mass model of the galaxy cluster MACS J1149.5+2223, relying on the identified image positions only, we include the surface brightness distribution of the SN Refsdal host, to infer some cosmological parameter values within a flat LCDM framework. We consider two cosmological parameterizations, with either H0 and {\Omega}m or H0 alone treated as free parameters, acknowledging the milder dependency on {\Omega}m when incorporating additional constraints from the SN host. We explore four different mass-model realizations to assess some of the systematic uncertainties associated with the constraint weighting and source reconstruction. Although known as one of the main possible systematic effects, including influences from the external convergence, we deliberately keep the cluster total mass parametrization fixed to that of the previously published point-like mass model to enable a fair comparison. This reveals that incorporating 77,000 HST pixels reduces the uncertainty on the inferred value of the Hubble constant by a factor of 2.0, while all model realizations remain mutually consistent. Our final measurements are $H_0 = 65.0^{+2.0}_{-1.8}$ km/s/Mpc and $\Omega_m = 0.30^{+0.08}_{-0.09}, in agreement with previous results. This corresponds to a substantially improved precision, reaching a remarkable 3.0% uncertainty on the value of H0 given our cluster total mass parameterization. The contributions of the uncertainties on the lens total mass and on the measured time delays to the final H0 error are now comparable, indicating that future improvements in cluster-scale time-delay cosmography will increasingly depend on (abridged version).

In incomplete markets, contingent claims cannot in general be replicated exactly, and hedging must therefore be formulated as an optimization problem over admissible trading strategies. This paper studies the relation between quadratic hedging and entropic hedging in a discrete-time incomplete market when strategies are restricted to bounded finite-dimensional parametric classes. These restrictions induce admissible gain sets that are compact and convex in \(L^2\), providing a common well-posed setting for both the restricted mean-square hedging problem and the restricted exponential-utility hedging problem. The main contribution is to establish a uniform small-risk-aversion expansion of the entropic risk functional over the restricted gain set and to use this uniformity to obtain convergence of the corresponding minimizers. Under a uniform exponential integrability assumption, the restricted entropic hedging gains converge in \(L^2\) to the unique restricted quadratic hedge as the risk-aversion parameter tends to zero. The theoretical result is illustrated numerically in a regularized discrete-time stochastic volatility model of Heston type.

The continued improvement of large language models increasingly depends on synthetic experience, yet effective data synthesis for interactive agents requires not only executable and verifiable trajectories, but also interactions that provide learning value for the current task agent. Existing methods increasingly incorporate environment feedback, data validation, or learner failures, but still face a feedback-to-decision gap: observed weaknesses do not directly specify how future synthesis should change, or whether they can be represented by the current synthesis space. To address this gap, we propose EvoSyn, a learning-feedback-driven multi-agent framework that organizes synthesis around a revisable state specifying what interaction factors can be deliberately generated and where synthesis effort should be allocated. A role-specialized multi-agent system translates learner feedback into synthesis revisions, while a Designer allocates synthesis directions and a Planner grounds them into executable experience. This process enables within-space adaptation toward weak or under-covered regions and synthesis-space expansion that promotes newly discovered factors into explicit generation controls. Experiments on $\tau^2$-Bench and CAR-Bench show that EvoSyn improves average task success by 13.3 points over the strongest synthesis baseline, while producing higher-quality and structurally broader synthetic experience and discovering controllable factors missed without learner feedback.

This paper introduces an enhanced channel estimation method for reconfigurable intelligent surface (RIS)-assisted multiple-input multiple-output (MIMO) communications. Considering the presence of both the direct link from the base station (BS) to the user equipment (UE) and the indirect link via the RIS, we formulate a tensor-based receiver for joint direct and indirect channel estimation called the TenDICE algorithm. The proposed receiver relies on an extended parallel factor analysis (PARAFAC) tensor model for the received pilot signals, from which the involved (direct and indirect link) channels are estimated using a simple alternating least squares scheme. The proposed TenDICE method is compared with the state-of-the-art least squares (LS) method, Khatri-Rao factorization (KRF) method, enhanced trilinear alternating least squares (E-TALS), and the theoretical Cramér-Rao lower bound (CRLB) based on normalized mean square error (NMSE) and spectral efficiency (SE). The proposed TenDICE outperforms the LS and E-TALS methods and performs similarly to the KRF methods while estimating the direct channel.

The post-reionization epoch ($z\sim4$-6) marks a critical phase in galaxy evolution, during which galaxies rapidly assemble their stellar mass, enrich their interstellar medium, and build up substantial dust reservoirs. The physical processes responsible for such rapid dust enrichment remain uncertain. We investigate the dust enrichment history of the ALPINE-CRISTAL-JWST (ACJ) sample using the semi-analytic galaxy formation model GAlaxy Evolution and Assembly (GAEA), which follows galaxy assembly, chemical enrichment, and dust evolution. We select model galaxies matching the observed stellar masses, star formation rates, and metallicities of the ACJ sample, compare their dust properties with observed scaling relations, and trace their enrichment histories and environments. The model reproduces the observed dust mass-stellar mass relation and the dust-to-stellar mass, dust-to-gas, and dust-to-metal ratios as functions of stellar mass and metallicity. Reproducing these constraints requires efficient dust growth in the interstellar medium; stellar dust production alone cannot account for the observed dust reservoirs, even under the most optimistic assumptions. The combination of stellar mass, star formation rate, and metallicity provides a good predictor of galaxy dust content. Our model suggests rapid dust build-up consistent with the previously noted rapid metal enrichment, with evidence for a transition around $z\sim8$-9 below which dust growth becomes the dominant dust formation channel. Finally, ACJ-like galaxies preferentially reside in locally overdense environments, suggesting that environment may have contributed to their accelerated evolution. Taken together, these results support rapid and coupled stellar, metal, and dust build-up in ACJ galaxies during the post-reionization era.

We prove that a tubular group is linear if and only if it is residually finite, equivalently virtually primitive. Every such group embeds in $\SL_n(\Z[u])$. For a primitive tubular decomposition with at most $m$ edge slopes at each vertex, our representation has dimension $2\max\{2,m\}$. In particular, Gersten's group embeds in $\SL_6(\Z[u])$, and this dimension is minimal.

Reversible motion can produce a directed circulation of surviving ancestry in a \FV\ particle system. We prove that two particles are the exact threshold for universal ancestral balance in finite systems with uniform copying. For two particles, the stationary current of surviving labels vanishes for every finite irreducible reversible Markov chain and every positive killing vector. On a fixed complete three-state graph, nonzero circulation occurs for each population size at least three on an open dense set of killing vectors of full Lebesgue measure. A single killing vector produces circulation simultaneously for every population size at least three, and such vectors form a dense $G_\delta$ set of full measure. The current is divergence free and vanishes on bridges, so three sites and three particles are the smallest possible setting for circulation. We also identify the probabilistic effect of this circulation: for sufficiently rare transplants, migration followed by fixation is more likely in one direction than in the reverse direction.

Near-field communication with large antenna arrays requires beamforming techniques that account for spherical wavefront effects. While predefined codebooks are widely used in multiple-antenna wireless systems, the performance of randomly generated codebooks in the near-field region has not been investigated. In this work, we study four randomly generated codebooks for near-field communications: random vector quantization (RVQ), random phase (RP), random far-field (RFF), and random near-field (RNF) codebooks. We use the known RVQ gain distribution as a benchmark. For RP, we derive a large-array bulk approximation and characterize the high-gain tail behavior of the gain distribution. For RNF, we characterize the high-gain tail behavior of the gain distribution. For RFF, we derive an analytical approximation for the maximum achievable beamforming gain when far-field beams are used to serve near-field users, along with the corresponding maximizing far-field direction. We also analyze the maximum-gain scaling with codebook size. Monte Carlo simulations verify the analytical results and show that RVQ and RP degrade significantly as the array size increases, while RNF achieves the most favorable high-gain scaling. RFF can outperform RVQ and RP for finite codebook sizes, but its maximum gain saturates below one due to the phase mismatch between the far-field beams and the near-field channel. These results show how codebook structure, array size, and codebook size affect random beamforming performance in the near-field.

Mental accounting is usually explained by self-control, and two-stage budgeting by separable preferences. We propose a third explanation based on sellers' market power. A consumer who commits to spend a fixed number of dollars in a category has unit-elastic demand, so a seller in that category gains nothing by restricting quantity. A consumer who instead targets a fixed physical quantity has perfectly inelastic demand, a case that a monopolist can exploit. We characterize the commitments that remove a seller's incentive to mark up prices, and show that the fixed dollar budget is the least restrictive. We then show that what matters is budget rigidity rather than separability of preferences. A consumer who re-optimizes her category budget when prices change is insulated from market power only if categories are substitutes; a rigid budget insulates her in every case. Mental accounting can therefore be read as a commitment device against markups, a rationale that is separate from the self-control motive in the existing literature.

Event sequences are chains of random events, each described by heterogeneous (numerical and categorical) features. The problem of event sequence forecasting occurs in finance, healthcare, business process intelligence, and other fields. In this work, we study multi-horizon forecasting of event sequences as a conditional generative modeling problem. We evaluate forecasts not only for prediction accuracy, but also for distributional realism. To this end, we present a benchmark comprising metrics for both quality aspects, strong statistical baselines, recent models adapted from adjacent fields, and curated real-world datasets. Careful evaluation reveals that no method excels at both accuracy and realism, highlighting a fundamental trade-off between the two. Finally, we introduce Wasserstein BaryBooster, an algorithm that aggregates samples from a realism-strong diffusion model into a state-of-the-art predictor, demonstrating that distributional realism can be converted into prediction accuracy. Code is available at https://github.com/xaosina/event-sequence-forecasting.

Semi-Lagrangian (SL) schemes advect a distribution function along characteristics and remap it to a fixed grid. They are unconditionally stable without a global solve, and with discontinuous Galerkin (DG) integration the transfer along a magnetic field line conserves mass exactly. This paper presents the parallel and GPU design of a 3X+2V SLDG drift-kinetic tokamak solver on unstructured, flux-surface-aligned poloidal planes of shaped cross section, extruded toroidally. Transport follows magnetic field lines across the planes, and each velocity channel advances a fixed distance per step, proportional to its parallel speed, so the halo depth a channel needs grows with its speed. Velocity-dependent ragged halos exploit this, sending each channel only its own upwind depth. The 4D decomposition includes the toroidal angle, the 2D poloidal plane and the magnetic moment, and leaves the parallel velocity whole: no stage transports across the magnetic moment, while a parallel-velocity partition would put a halo inside the acceleration stage. On NVIDIA A100s the design strong-scales over the magnetic-moment axis from one to 12 four-GPU nodes, and a weak series of three points refining all five phase-space dimensions together, 16x in total per series point, holds a weak-scaling efficiency of 75% on 16 GPUs and 57% on 256, from a one-GPU base case. The pointwise interpolations within the poloidal plane do not conserve mass, and the full step's mass drift stays below 2.1x10^-5 at every series point. The design was implemented with an AI code assistant under a machine-checked regression discipline, in which every change is held to one of four exactness classes, from bit-identical to a physics rate within a stated tolerance.

Many neck disorders, from chronic neck pain and headache to thyroid and nodal malignancy and radiation-induced fibrosis, both arise from and alter the mechanical properties of neck tissues, yet clinical assessment still relies on morphology, palpation and symptoms. Ultrasound and magnetic resonance elastography (MRE) now allow these properties to be measured in vivo, but the neck literature differs widely in methodology and remains dispersed across modalities and tissues. This review brings together the biomechanical basis of neck tissue behavior, the clinical conditions that alter it, and the principles and applications of ultrasound elastography and MRE in the neck. We show that ultrasound elastography has demonstrated diagnostic value for thyroid nodules and cervical lymph nodes, while neck MRE has been shown to be feasible but remains at an early stage. Across both modalities, reported values depend strongly on fiber orientation, posture, muscle activation and acquisition conditions. We therefore propose reporting recommendations and discuss how measurements beyond stiffness, including anisotropy, viscosity and nonlinearity, could improve disease management and inform computational models of the neck.

We introduce the global obstruction property (GOP) for Artin algebras: for every finitely generated module of infinite projective dimension, two distinct syzygies have a common nonprojective indecomposable direct summand. We prove that GOP implies the generalized Auslander–Reiten condition, and hence the Auslander–Reiten conjecture. We give criteria for GOP to hold for pullbacks over semisimple algebras and construct an example which is not syzygy-finite. For each Artin algebra $A$, we construct an algebra $\mathcal{G}(A)$ satisfying GOP by using triangular matrix extensions. We show that $\text{fin.dim}\mathcal{G}(A)=\text{fin.dim} A+2$ and that $A$ and $\mathcal{G}(A)$ are singularly equivalent. Finally, we give a pair of derived equivalent finite-dimensional algebras for which GOP holds on only one side.

A latent recurrent language model applies a shared block of layers repeatedly, so one checkpoint can trade inference compute for accuracy by changing the loop count. At one loop, however, such models typically fall well short of a non-recurrent model of the same size, a one-loop tax that makes a separate non-recurrent model the better choice whenever latency matters. We present Elastic Latent Recurrence (ELR), a recipe for training looped language models without this tax. At one loop, an ELR model costs the same to serve as a non-recurrent model of the same size and, trained on the same data, reaches similar accuracy. Each additional loop then buys more accuracy for more inference compute. Three choices make this work: a gated state update with a learned step size, loop counts mixed within each optimizer update, and conditioning each pass on its loop index and budget. The first two preserve one-loop accuracy, and the third lowers loss at every additional loop. Mixed depths can be introduced late in pretraining to limit the added training cost, and with zero-initialized conditioning added at the switch, this late schedule reaches lower loss than mixed-depth training from scratch. In MoE models, reusing first-loop expert routes on later loops cuts mixed-depth update time by 11.7% with no measurable change in accuracy. At inference, keeping only the last loop's attention KV cache holds attention KV memory at its one-loop level while retaining most of the gain from additional loops. On 4B dense and 30B-A3B MoE hybrid models, ELR at one loop is 0.52 points below and 1.04 points above its matched baseline on a ten-task average, and four loops add 7.25 and 4.75 points. Carried through post-training, the recipe yields Koi-4B, a single checkpoint that scores an estimated 18 on the Artificial Analysis Intelligence Index at one loop, matching Qwen3.5-4B, and 29 at four loops.

We show that if a contractible Stein manifold $X$ with the density property admits a proper holomorphic embedding into $\mathbb{C}^k$, then $X \times \mathbb{C}^k$ is biholomorphic to $\mathbb{C}^{n+k}$ with $n=\dim X$. We also prove that McLean's exotic affine variety $K_n$, $n\ge4$, a smooth contractible affine variety which is not biholomorphic to $\mathbb{C}^n$, has the density property. Hence for $n \ge 4$ there exist an integer $m \in \{ n, \dots, 2n\}$ and an affine variety $X$ such that $X \times \mathbb{C}$ is biholomorphic to $\mathbb{C}^{m+1}$ and $X$ is not biholomorphic to $\mathbb{C}^m$. This settles the Holomorphic Cancellation Problem in the negative. We also construct holomorphic actions of $\mathbb{C}^*$ and finite cyclic groups on complex Euclidean spaces that are not topologically conjugate to linear actions.

Event ontology is the foundational framework for defining events in a dataset. Event type induction generally involves domain experts who define guidelines for creating event types and their names based on the dataset's domain and scope. However, this process is resource-intensive in terms of annotation costs and is prone to human biases. To alleviate the need for manual event ontology creation, we propose the CLUE framework for event type induction, naming, and annotation. Our main contributions in this work are: 1) designing a novel joint embedding representation for expressing both the verb-centric view and the semantic view of events; 2) event type induction using contrastive loss guided clustering; 3) event naming and refinement. Experiments demonstrate the effectiveness of our method. Our code is available at https://github.com/PratikMaitra/CLUE.

Spin-dependent forces enable independent control over the effective frequencies of bosonic motional modes and their coupling strengths to spins in trapped-ion spin-boson quantum simulators. We show that the relative signs of these effective frequencies provide an important programmable resource for engineering distinct multimode dynamics. Focusing on a two-mode quantum Rabi model, changing the frequencies of the two modes from the same sign to opposite signs reorganizes the resonance structure of the spin-mediated bosonic interaction, converting the dominant resonant process from excitation exchange to correlated pair creation. In the dispersive regime, the resulting effective interaction realizes a two-mode parametric amplifier with stable oscillatory, critical, and exponentially amplifying regimes separated by a spin-dependent instability threshold. Across these regimes, the dynamics generate collective-quadrature squeezing and bipartite motional entanglement. Preparing the spin in a coherent superposition further allows different spin components to experience distinct stability regimes, generating tripartite spin-boson entanglement, while spin postselection can coherently combine the conditional bosonic branches to generate non-Gaussian entangled bosonic states. Finally, we outline an implementation using parameters within reach of current trapped-ion platforms.

Frustrated Ising models of social networks mix ferromagnetic (agreement-seeking) and antiferromagnetic (antagonistic) couplings, and the antagonistic bonds are almost always placed independently at a fixed rate, yielding a +/-J spin glass. Real signed networks (Epinions, Slashdot, Wikipedia) instead show an excess of balanced triads, and perfectly balanced placement is gauge-equivalent to an unfrustrated ferromagnet. We compare three sign-placement conditions on a frustrated random-field Ising model of a polarised population, sharing an identical graph, field assignment, and edge-weight magnitudes: an independent baseline; a balance-structured condition whose camp membership coincides with each voter's field sign; and the same condition with camp labels randomly permuted, so triadic balance is identical by construction while the camp-field correlation is removed. On a confound-controlled Kolmogorov-Smirnov test of the sampled energy-gap distribution (simulated annealing), the aligned condition differs from the baseline at every density tested (p = 7.3e-8), while the permuted condition shows no significant difference (paired mean difference 0.003, 95% CI [-0.041, 0.046]). Field-camp alignment, not triadic balance alone, drives the effect. The qualitative direction follows from a known mapping; the contribution is a quantitative, confound-controlled measurement of the effect's size away from that limit, robust to parallel-tempering sampling, system sizes up to N = 300, and structural noise calibrated to real networks, while triadic balance alone produces no detectable effect at N = 90 and at most a small one at N = 180-300 (paired change in KS distance about 0.04, 95% CI upper bound below 0.09).

The newly measured correlation between the neutrino and X-ray luminosities of Seyfert galaxies has yielded the first neutrino-based measurement of the Hubble constant. Building on this result, we forecast the reach of a global neutrino-telescope network by 2030, utilizing full-sky IceCube exposures combined with KM3NeT-ARCA and Baikal-GVD. We find that this 2030 configuration sharpens the neutrino-measured Hubble constant from today's $H_0 = 66^{+44}_{-37}$ to $H_0 =66^{+15}_{-13}$ km s$^{-1}$ Mpc$^{-1}$. Additionally, we project that the disk-corona slope of the neutrino X-ray luminosity relation ($L_\nu\propto L_X^\beta$) will reach $\beta=0.72\pm0.04$, enabling a $\geq \!5\sigma$ exclusion of the calorimetric value $\beta=1$. Finally, we show that, if a high-redshift population of neutrino sources, such as JWST's "little red dots," is identified and obeys the same luminosity relation, we can obtain a new independent measurement of the matter density $\Omega_m$ and the dark energy equation-of-state parameter $w$.

We prove that every connected, locally finite, quasi-transitive graph which is quasi-isometric to a graph excluding a countable clique as a minor is quasi-isometric to a planar graph. Consequently, every finitely generated group with such a quasi-isometry is virtually planar, meaning that some finite-index subgroup admits a planar Cayley graph. We also show that a connected, locally finite, quasi-transitive graph admitting a regular map into a bounded-degree graph excluding some finite minor admits a regular map into a bounded-degree planar graph. If the source is accessible, the target need only exclude a countable clique. Combined with work of Mackay, Spriano, and the present author, this shows that a finitely presented group admits a regular map into a bounded-degree graph excluding a countable clique if and only if it is virtually planar.

Modern video-language models are increasingly treated as general-purpose video reasoners, yet it remains unclear whether they robustly capture temporal dynamics in safety-critical settings such as driving. We study this question through controlled ego-motion tasks that require identifying actions from driving videos. Across video-language models, we find that standard prompting, visual prompting, few-shot prompting, and chain-of-thought prompting remain unreliable and highly task- and model-dependent, with persistent class biases, although newer models show substantial task-specific few-shot gains. In contrast, lightweight LoRA finetuning and simple probes trained on intermediate visual features recover strong motion-discriminative performance, indicating that motion-relevant cues are often present inside the models but are not reliably used by the final video-to-language pathway. These results suggest that, in driving, current video-language models remain unreliable temporal reasoners and often behave more like frame-level, bias-driven recognizers than dependable temporal world models.

Multi-teacher distillation allows transferring knowledge from multiple "teacher" networks to a single "student" network. This method is promising in fields such as digital pathology, where many powerful foundation models were proposed recently. However, as we show in this paper, the standard multi-teacher distillation approach reacts poorly to an increase in the number of teachers used to train a student encoder. We demonstrate that learning teacher-specific representations is a key to scaling in MTD. Importantly, different design choices, such as learnable teacher tokens, a tailored attention scheme, additional mixture-of-experts layers and a contrastive loss, are proposed to better learn such teacher-specific representations. We show that our method better scales with respect to the number of considered teachers, allowing us to train compact student encoders, up to 10x more efficient than larger teacher foundation models, while matching their performance and even outperforming them on a large set of tile-level and slide-level tasks. We conduct a thorough empirical validation, evaluating more than 10 foundation models on 39 tasks spanning tile and slide levels. As a result, we release a new collection of strong and efficient foundation models, named OMNI, trained from the knowledge of 10 state-of-the-art foundation models. We finally conduct an analysis of the learned teacher-specific representations, highlighting their complementarity and explaining why they can be easily aggregated at downstream time.

We formulate a theory of spin-polarized tunneling between a nanowire tip of a topological Kondo insulator SmB$_6$ and the bi-collinear antiferromagnet Fe$_{1+x}$Te in which the tunneling conductance which measures magnetic contrast changes sign when the bias is reversed. First we show that no time-reversal-symmetric tip, however strong its spin-momentum locking is, can produce experimentally observed magnetic contrast, which implies that time reversal must be broken at the tip surface. Within a mean-field treatment of the surface theory of SmB$_6$ we find that the interaction between conduction $d$- and $f$-electrons drives a surface phase transition near $T_c\approx10$ K in which the time-reversal-odd component of the Kondo hybridization amplitude condenses in the same symmetry channel as an out-of-plane surface magnetization, gapping out the Dirac surface states. Below the transition the spin-resolved spectral function of each gapped surface Dirac cone is odd in energy about its node. For a charge-neutral apex the Fermi level lies between the surface $\overline{\Gamma}$ and $\overline{X}$ Dirac nodes, so the spin-polarized conductance, which is determined by this function at the bias energy, reverses its sign close to zero bias. This sign change is thus a spectroscopic signature of time-reversal symmetry breaking on the surface of SmB$_6$.

Online reinforcement learning (RL) fine-tunes diffusion models with black-box rewards, but reverse-process methods such as Flow-GRPO are limited in efficiency by poor credit assignment, while forward-process methods such as DiffusionNFT are efficient yet unstable without carefully tuned exponential-moving-average (EMA) off-policy training. We show that NFT-style objectives contain a hidden forward–reverse consistency force that joint positive–negative training cancels only partially, and that EMA stabilizes training through restoration rather than off-policy rollouts. We introduce DiffusionART (Diffusion Reinforcement with Aggregation-First Reward Tilting), which aggregates each prompt's rewards into a shared reward-improving direction for reverse-process states, removing the consistency force by construction, and adds a trust-region-masked second update for drift-adaptive restoration without EMA tuning. It requires no differentiable reward model, remains effective with fewer samples per prompt, and supports pairwise reward models (DiffusionART-Pairwise). Across image and video generation, DiffusionART trains efficiently and stably without per-task tuning. In particular, on HunyuanVideo-13B with equally weighted visual and motion quality rewards, DiffusionART reaches the late-stage mean reward levels of DiffusionNFT and Flow-GRPO-Fast with 12.02 and 11.72 times fewer GPU hours, respectively, while achieving 28.8% and 53.0% higher mean training rewards.

Most systems trained by equilibrium propagation or coupled learning are in the same class as multi-layer perceptrons: while they may be universal function approximators, they struggle to learn associative tasks which are easy for transformers. Equilibrium propagation relies on a bidirectional flow of information through the physical system between input and output. The reverse flow is broken by more complex nonlinear systems, such as those performing multiplication which is needed to express dot-product attention. Here we explicitly restore this bidirectional flow by introducing extra circuit elements for this purpose into a double or dyadic circuit. We then use these ideas to construct an attention mechanism exploiting the exponential current law of sub-threshold MOSFET devices. We show that in simulations this circuit can be trained to perform a simple in-context learning task, despite using realistic reverse components which transmit only an approximate gradient. For this training to succeed we need cross-entropy loss (not mean squared error). We generate nudges appropriate for this loss using a novel floating circuit element, based on the same MOSFET sub-circuit used for softmax. We train the resistors in this circuit using the Lion optimiser, which can be implemented using one capacitor per parameter to store local memory of past gradient estimates.

Speculative decoding accelerates large language model inference by using a lightweight drafter to propose multiple tokens that are subsequently verified in parallel by the target model. Recent diffusion-based drafters make this paradigm particularly attractive: deep attention-based neural backbones provide strong predictions, while drafting multiple future positions in parallel avoids the sequential cost of autoregressive generation. This parallelism, however, comes at a cost. Tokens within a draft block are predicted without access to the realized identities of preceding draft tokens, removing intra-block dependencies that would naturally be available under autoregressive generation. We introduce DCOCO (DFlash Contextual Corrector), a lightweight mechanism that restores these dependencies while preserving parallel drafting. Rather than committing to each position's top prediction, DCOCO retains multiple candidate tokens and evaluates their contextual compatibility across adjacent draft positions in parallel, enabling contextually informed token selection at low additional cost. Across four verifier backbones and ten benchmarks, DCOCO achieves the highest average throughput in every evaluated configuration, spanning greedy and stochastic decoding with and without extended reasoning, with speedups of up to 6.6x over vanilla decoding. Its advantage persists under concurrent serving: DCOCO outperforms the strongest competing method by 6-23% across 4-32 concurrent requests, while other sequential correctors fall below uncorrected DFlash at these loads. These results demonstrate that recovering intra-block dependencies through parallel contextual scoring delivers consistent acceleration across decoding regimes and serving loads.

Static coding benchmarks are fixed at release. Over time their tasks leak into training data, and stronger agents eventually saturate them. We argue that a benchmark should instead aim to be self-scaling: its tasks and oracles are created at evaluation time, so no content exists in advance to leak, and its difficulty is set by the participating agents rather than fixed at release. We introduce SWE-Duel, an arena in which two coding agents evaluate each other over mirrored Red-Blue turns. In each turn, the Red agent works on a real repository across three context-isolated sessions. It proposes and implements a new feature with feature tests, then embeds a subtle bug in that feature along with sealed bug-triggering tests. The resulting pull request is admitted only if (1) it passes the feature tests and the repository's existing tests while failing exactly the bug tests, and (2) a fresh session of the same agent can repair the bug blind, which rules out unsolvable tasks. The Blue agent then reviews the pull request without being told a bug exists. It wins only if its revision passes all three test suites, and then the two agents swap roles. We run a tournament among 11 model-harness participants on 15 repositories in five languages for 1,500 turns. The leaders win in different ways. GPT-5.5 with Codex CLI ranks first as the strongest solver, repairing 98.3% of the defects it reviews. Claude Opus 4.8 with mini-swe-agent ranks second as the strongest task author, with a 27.3% Red win rate. Finally, within one model family, tasks written by the newest version defeat a fixed-reference solver far more often than those of older versions (42.9% vs. 4.0%), giving preliminary evidence that task difficulty scales with task authoring capability.