Papers

Magnetic field-line transport in magnetically confined plasmas is commonly modeled as diffusion through stochastic magnetic layers generated by resonant perturbations. However, experimentally relevant magnetic configurations frequently exhibit transport rates that are substantially slower than predicted by fully stochastic models, indicating the presence of unresolved dynamical barriers. In this work, we investigate the role of nested resonant island chains and their associated invariant manifolds in regulating transport within the edge magnetic topology of DIII-D tokamak discharge #171491 subjected to strong n=3 resonant magnetic perturbations. Using field-line tracing, Poincare analysis, and manifold calculations, we identify a hierarchy of homoclinic and heteroclinic tangles associated with both period-one and period-ten hyperbolic points. A systematic construction based on primary intersection points is developed to extract partial transport barriers from the full manifold geometry and divide the chaotic region into dynamically distinct domains. Monte Carlo simulations reveal that field-line escape exhibits a bi-exponential character when only the outermost period-one boundary is considered, reflecting the presence of a long-lived "sticky" region associated with the period-ten island chain. When the inner period-ten boundary is explicitly incorporated into the Monte Carlo, the escape process becomes approximately single exponential and is accurately described by turnstile lobe transport. The measured escape rate agrees with predictions obtained from the turnstile lobe area and the symplectic structure of the field-line map. These results demonstrate that nested tangles provide a quantitative transport skeleton governing magnetic-field-line escape and establish a framework for analyzing transport barriers, cantori, and lobe dynamics in realistic fusion-plasma magnetic configurations.

Rare-earth (RE) atoms, such as erbium (Er), have inner-shell optical transitions in the near-infrared, some of which are in the telecom bands. These transitions are largely insensitive to the host environment due to effective shielding by the filled outer shells. Cryogenic rare-gas (RG) solids, such as solid neon (Ne) and argon (Ar), combine high chemical purity with a nearly nuclear-spin-free environment, making them attractive solid hosts for preserving optical and spin coherence. In this work, we isolate neutral Er atoms in solid Ne and Ar at 6 K and characterize their emission by steady-state and time-resolved photoluminescence (PL) spectroscopy. In both hosts, we observe several near-infrared transitions between 1000–1500 nm, including inner-shell emission in the telecom O and E bands. The $4f \rightarrow 5d$ transition near 1299 nm forms a multiplet of narrow components, consistent with crystal-field splitting, and decays on a microsecond timescale. The $4f \rightarrow 4f$ transition near 1437 nm shows a matrix shift below 1 cm$^{-1}$ in both hosts. In solid Ne, its inhomogeneous linewidth is bounded by the spectrometer resolution to < 22 GHz, and its PL lifetime reaches up to 78.9 ms, close to the estimated radiative limit. Thermal annealing extends the 1437 nm lifetime in Ne and suppresses the phonon sidebands relative to the zero-phonon lines in Ar. These results identify RG solids as promising hosts for long-lived telecom-band Er emission and provide a foundation for future spin-photon interfaces and hybrid quantum architectures.

Choosing an agent system means choosing both a language model and the harness through which it acts. We ask whether a strong model, harness, or pairing stays strong when the setting changes. We evaluate 66 configurations: four configurable harnesses (OpenHands, DeepSeek Harness, PI, and openJiuwen) paired with five models on TUA-Bench, ALE-CLI, and Terminal-Bench 4, plus the native Codex-GPT and Claude Code-Claude pairings. Model rankings reverse across harnesses. On Terminal-Bench 4, Claude leads GPT by 7.94 points in OpenHands but trails it by 30.16 points in PI. For four of the five models, the best harness changes from one benchmark to another, yet some pairings hold: openJiuwen gives Kimi its highest score on all three benchmarks, by 5.61 to 11.11 points. A model's own vendor harness is not reliably its best, and higher cost does not reliably buy a higher score. On Terminal-Bench 4, GPT scores higher under PI than under DSH at less than a quarter of the cost per task. Matched trajectories suggest why fit varies. Models start almost all repairs themselves, so much depends on whether the harness hands failures back in a form the model can use. GPT does best with PI's lean scaffold, while Kimi, which often issues malformed tool calls, does best in openJiuwen. We argue that the model, the harness, and the task should be evaluated together, and we release the harness adapters, evaluation code, and all 6,204 scored trajectories at https://github.com/liyix/finding-the-right-fit and https://huggingface.co/datasets/yixuanli97/finding-the-right-fit.

The discovery of Chair44 (Tsiokos, 2026) settled the three-dimensional einstein problem with a strongly aperiodic polyhedral monotile in $\mathbb{R}^3$. This note extends the underlying mechanism—the rep-$2^N$ chair $C_N = [0,2]^N \setminus (1,2]^N$ with corner/socket markings—to $\mathbb{R}^N$. Besides expository material (the rep-$2^N$ dissection and a conditional strong-aperiodicity theorem under lattice registration and hierarchical enforcement), the note makes a new computational contribution. We introduce a frame-marking formalism in which the marking of a tile is its full orientation frame and the matching rule is the contact language generated by the substitution itself; this makes the search for matching rules finite in every dimension. We give a finite certificate (coarsening closure, tightness, and a two-shell enclosure analysis) whose validity implies that every lattice-registered tiling by the marked tile is uniquely hierarchical, hence strongly aperiodic. For $N=3$ the certificate passes: it yields explicit facet matching rules on the 24 panels of $C_3$ (135 admissible facet-contact triples) and reproduces, from first principles and independently of published constructions, the Chair44 statistics 2388 $\to$ 44 admissible contacts (30 occurring), 33 one-shell clusters, 15 extendable, each forcing a unique supertile. Among the 2187 homochiral frame assignments of the 3D substitution with a translated central child, the certified one is unique up to conjugation. For $N=4$ the same pipeline is run on several structured families of frame assignments (canonical, $D_4$-, $Z_2\times Z_2$- and $Z_4$-symmetric, and a lift of the 3D solution); none is coarsening-closed, and we report the failure data. A self-similar marking of $C_4$ thus remains an explicitly finite, open computational problem, which we state precisely. Code: https://github.com/dimkadimon/Monotile-RN

A magic square of order $n$ is a square array of $n^2$ distinct integers such that the integers of every row, every column and the two diagonals have the same common sum. If a magic square remains a magic square even when each integer is replaced by its square, it is called bimagic. If, however, all the rows and all the columns of the aforementioned square of squares add up to a common sum, but the two diagonals do not, the original magic square is called pseudo-bimagic. Except for one family of $6 \times 6$ pseudo-bimagic squares, parametrized families of bimagic squares have not been published. In this paper we construct parametrized families of $6 \times 6$ pseudo-bimagic squares and $8 \times 8$ bimagic squares. We use these families to construct, for any positive integer $m$, multi-parameter families of pseudo-bimagic squares of order $6m$ and bimagic squares of order $8m$. We also give numerical examples of $6 \times 6$ and $12 \times 12$ pseudo-bimagic squares and an $8 \times 8$ bimagic square.

We study the Dirichlet problems with unbounded boundary data. Our main theorem proves interior continuity from a local biholomorphic deformation satisfying a one-sided compatibility condition near the infinite locus and barriers on the associated boundary cluster set. The deformed domains need not be nested, and the proof uses neither a Runge hypothesis nor a pseudoconvex ambient neighbourhood. We also prove nonmonotone Perron–Bremermann convergence outside an intrinsic neighbourhood pluripolar hull, a disk-fibre hull representation, stability under simultaneous nonmonotone variations of the domain and obstacle, and uniqueness in a fixed obstacle class.

Vision-Language-Action (VLA) models provide strong behavioral priors for robotic manipulation, yet efficiently adapting them to downstream tasks remains challenging. Recent work addresses this challenge by adapting frozen VLAs through online reinforcement learning (RL), whose sample efficiency depends on the quality of the state representation used by the actor and critic. Existing methods construct such representations either with VLA-independent visual encoders or through fixed compression of internal VLA representations. Neither design explicitly extracts the task-specific action-relevant VLA features most useful for downstream action refinement and action-value estimation, therefore limiting sample efficiency. To address this limitation, we introduce eRLT, which constructs an effective state representation by routing task-specific action-relevant information across both tokens and layers of the frozen VLA. Specifically, learned routing tokens dynamically aggregate visual-language features at multiple depths, while a lightweight layer router combines these summaries into a fixed-dimensional RL token. The routing module is initialized using expert demonstrations to capture features predictive of expert actions and then refined using critic feedback from online interactions for action-value estimation. Across seven LIBERO and RoboTwin tasks, eRLT improves mean normalized learning-curve AUC by up to 23.7% over representative baselines. Real-robot experiments on USB connector insertion and motherboard ribbon-cable insertion further show AUC improvements of 108.9% and 46.7%, respectively, over the strongest baseline. Project page: https://red0orange.github.io/eRLT_site/.

Large language models can translate business descriptions into optimization models, but executable code may misrepresent constraints or objectives. A solver can then return an optimal solution to the wrong problem. Even when the solution satisfies the intended operating rules, a better plan may exist. For organizations that repeatedly use optimization modeling, an LLM-based framework should produce accurate formulations at low cost and, ideally, run locally. We study how to verify improvements and allocate attempts across LLMs that differ in price and capability. We develop OSCAR (Optimization modeling by Simulator, Coder, And Reviewer), which uses an offline Simulator certified against labeled decision examples to compare candidates and continues searching beyond feasibility. We model the search for the next certified improvement as sequential decisions under unobserved difficulty: which LLMs to call and when to stop. In a simplified known-prior setting, we give conditions under which cost-ordered escalation is optimal. For general menus, we derive a prior-free competitive guarantee. On five benchmark problems, OSCAR achieves 95% to 100% accuracy at the reported settings using two small open-weight LLMs, each deployable locally on a single GPU. Their single-attempt accuracies average 29% and 48%. In five runs per problem, Codex and Claude Code incur average token costs 3.1 and 5.8 times OSCAR's, respectively. OSCAR supports open-weight models locally or in the cloud, depending on budget and confidentiality requirements. Firms should maintain labeled decision examples of feasible and infeasible decisions to clarify plain-language operating rules. OSCAR follows these labels when an LLM's interpretation conflicts with them. As LLM capabilities and prices change, OSCAR's simple operating rules and adjustable settings help firms adapt their model choices and benefit from these advances.

We study an endogenous nonstationary stochastic bandit problem with latent linear dynamics, where actions affect both immediate rewards and the future evolution of an unobserved latent state. Rewards are bilinear in the current action and latent state, inducing history-dependent rewards and a nontrivial long-horizon planning problem. The existing explore-then-commit approach achieves $\tilde{O}(T^{2/3})$ regret by uniformly exploring to estimate the latent dynamics and then committing to an optimized open-loop action sequence. We show that this rate can be improved via adaptive block-level optimism. Our key step is a cyclic approximation: under stable dynamics, the infinite-memory reward process can be truncated, and the open-loop benchmark can be approximated by optimizing a finite-memory block-level proxy. Building on this reduction, we propose a UCB-based block algorithm that maintains confidence sets for the truncated dynamics parameters and selects blocks optimistically. We prove a regret bound of order $\tilde{O}(\sqrt T)$, significantly improving over the previous $\tilde{O}(T^{2/3})$ guarantee for the same model. To the best of our knowledge, this is the first $\tilde{O}(\sqrt T)$ regret guarantee for latent linear-dynamics bandits with bilinear reward observations and an open-loop action-sequence benchmark.

Human reasoning depends on how objects are related within propositions. How do relations organize the language representations of contextual contents? We give an LLM a list of facts in its context (e.g., Alice eats an apple. Bob eats a pear.) and measure how its hidden state changes when the question switches from what Alice eats to what Bob eats. Averaged over many lists, this change is a steering vector, which we call the ordinal vector. It points to a fact by its order of mention, the order in which the facts were stated in the context. We find that LLMs represent the fact a question asks about by its order of mention, not by the name the question contains. We state this as the ordinal addressing hypothesis: each order of mention has a fact address in the model's state, shared by all contexts, and a question moves the state to the fact address of the fact it asks about, while the context supplies what that fact says. Across Qwen, Gemma, and Llama, fact addresses are (1) ordered by mention: query states are organized by the order of facts, not of names, even when one fact has multiple subjects; (2) steerable: added to a question about the first fact of a new list, the ordinal vector makes the model answer with the second fact of that list; (3) low-rank: they span a low-rank subspace in which the first-mentioned fact is the easiest to reach, surprisingly similar to human recall; and (4) emergent: they are shared in late-middle layers, hold from 1.5B to 32B parameters, and form early in pretraining. Language models reach a stated fact by where it was mentioned, deepening our understanding of LLM reasoning.

We prove the continuity of the pluricomplex Green functions introduced by F. Lárusson and R. Sigurdsson on a class of bounded (B)-regular domains that includes all (B)-regular convexifiable domains.

Binary pulsars provide valuable laboratories for testing theories of gravity, but orbital Doppler shifts complicate their detection. Fourier-domain acceleration and jerk searches address this challenge via matched filtering, but at substantial computational cost. We present LeoNet, a convolutional neural network that uses ten learnable filters to extract features of signals affected by Doppler shifts. The resulting ten-channel feature map provides a compact, lower-dimensional alternative to an explicitly sampled acceleration-jerk response grid and is analysed by a convolutional classifier to identify candidate signals. For simulated observations lasting 500 s, LeoNet achieves a mean relative reduction in false negative rate of 55.5% across five sampling intervals compared with the evaluated PRESTO acceleration-search configuration. TensorRT-optimised LeoNet processes each 500 s observation in 3.44-4.37 ms in FP32 on an NVIDIA H100 PCIe GPU across eight sampling intervals, including preprocessing, inference, and postprocessing. At a sampling interval of 128 microseconds, its mean processing time is 3.54 ms, compared with 1.767 s for PRESTO FDAS on an AMD EPYC 9825 CPU with search-frequency limits of 96-1000 Hz, corresponding to an approximately 499-fold speedup in the measured processing time. These results suggest that LeoNet has the potential to improve detection performance, while its millisecond-scale processing time supports its use as a candidate-identification stage in real-time binary pulsar search pipelines.

Relational foundation models (RFMs) are pretrained once on a collection of relational databases and prediction tasks, and then applied zero-shot to previously unseen databases and tasks. To make a prediction for a target row, an RFM samples a neighborhood of rows linked to that row through foreign keys and uses this neighborhood as its inference context. Lowering inference cost is an important goal for any foundation model, and for RFMs this cost grows with the size of the context. The simplest ways to shrink the context is to drop some of the sampled rows, but this ignores the semantics of the database schema, so it is as likely to discard informative rows as uninformative ones. We propose STEER, a sampling approach that shrinks the inference context by concentrating it on the tables most relevant to the prediction task at hand. STEER obtains relevance information by prompting a large language model to rank the foreign-key edges of the database schema into relevance tiers for the given task, and then maps each tier to a probability of following that edge during traversal. Because the ranking uses only the schema, it is computed once per task and reused across all subsequent predictions, amortizing its cost. We evaluate STEER on three state-of-the-art RFMs (RT, RT-J, and Griffin) and show that it reduces inference context size by about 40% on average while maintaining, and in some cases improving, prediction accuracy.

Automated harness optimization can substantially improve LLM agents by iteratively updating their prompts, tool interfaces, and control logic from execution feedback. However, existing methods primarily optimize how the harness is updated while largely fixing which training scenarios generate the feedback that drives those updates. As the harness evolves, the scenarios most useful for further optimization can change, suggesting that the training curriculum itself should adapt alongside the harness. We formulate this missing dimension of harness optimization as an automated curriculum learning problem and introduce ActiveSaddler. ActiveSaddler models the evolving curriculum as a non-stationary bandit with dynamically instantiated optimization targets. It abstracts recurring failures into reusable failure-pattern arms, estimates the potential learning progress from further targeting each pattern, and adaptively balances revisiting known weaknesses with exploring unseen scenarios for new ones. Optimization outcomes continually update both the set of discovered failure patterns and their priorities, allowing the curriculum to co-evolve with the harness. Experiments on GAIA2 and Terminal-Bench 2.0 show that ActiveSaddler consistently discovers stronger harnesses, improving test Pass@1 by 4.4 and 7.5 percentage points over the same harness optimizer using a scenario order fixed before optimization, respectively. Ablations further show that these gains depend on dynamically constructing optimization targets, estimating their evolving utility, and balancing continued optimization with new failure discovery. Together, these results establish automated curriculum learning as a new crucial optimization dimension for harness optimization.

With the advancement of LLM-based multi-agent systems (MAS), an increasing number of opensource projects are adopting multi-agent architectures as the foundation of their core functionality. Although research and practice on MAS have attracted considerable attention, limited studies have explored the challenges faced by practitioners of open-source LLM-based MAS, the causes of these challenges, and potential solutions. To address this gap,we conducted an empirical study to understand the issues that practitioners encounter when developing and using open-source LLM-based MAS, the possible causes of these issues, and potential solutions. We collected 22,848 closed issues from 21 open-source LLM-basedMASand applied a mixed automated and manual filtering approach to reduce the dataset to 944 issues related to LLM-based MAS.We then analyzed these issues to understand the frequent issues encountered by practitioners, their underlying causes, and potential solutions. Our study results show that (1) Orchestration & Execution Issue is the most common issue faced by practitioners, (2) Workflow Problem, Tool Integration Problem, and Memory Problem are identified as the most frequent causes of the issues, and (3) Optimize Workflow is the predominant solution to the issues. Based on the study results, we derive empirically grounded implications for practitioners and researchers aimed at improving orchestration, tool integration, and memory mechanisms in LLM-based MAS.

Learned manipulation policies rediscover from data the spatial relations that rigid-body mechanics supplies in closed form, which leaves them fragile to geometric change. We present Screw Attention, a transformer layer in which the relation between two bodies is a spatial transform rather than a graph edge. Each pair of tokens carries the relative pose and, for robot joints, the joint screw. Messages are transported along this relation into the receiver's frame, while the attention scores see only frame-invariant quantities. By construction, the messages are equivariant to an independent change of frame at every token, and a single layer can express the velocity recursion of rigid-body mechanics. On LIBERO-Spatial, a policy of 16k parameters trained from object poses alone reaches 97.3% success, above graph, transformer and flat networks of the same size and a flat network with 27 times more parameters. Ablations show that the gain comes from transporting the correct relations, and that the structure pays most where the task requires relations between frames that nothing else supplies. The equivariance makes the policy robust to how the robot is described, where every other learned network collapses under a change of frame convention. Furthermore, the policy tolerates pose noise and calibration errors at least as well as an analytic controller. Used as a gated residual on an analytic controller, it also improves a contact-rich insertion task. Code and trained policies will be released.

Test-time reinforcement learning (TTRL) adapts language models on unlabeled test problems using supervision derived from their own samples. This makes the checkpoint's entry state consequential: a diffuse policy provides noisier self-supervision and may spend much of a limited adaptation budget merely concentrating probability mass before reliably expressing capability it already possesses. We propose entry-state sharpening: use data-free training before TTRL to prepare a general-purpose checkpoint in a state that subsequent label-free adaptation can exploit more efficiently. The idea is not tied to one training recipe; different data-free objectives can move the same base model to different entry states. Across five data-free checkpoints derived from Qwen3-4B and evaluated under an identical 15-step TTRL protocol, entry policy entropy strongly rank-orders endpoint conversion efficiency, a reliability-to-reachability measure (Spearman $\rho=-0.90$; $\rho=-0.99$ after controlling for entry reachability). The contrast across objectives is striking: R-Zero remains diffuse at $3.39$ nats and finishes below the untuned base in 6/6 matched comparisons across MATH, GPQA, and AMC, whereas SPIRAL reaches $0.07$ nats and achieves the highest post-TTRL accuracy on MATH and GPQA despite its self-play stage using no math training data. An in-domain label-free self-distillation intervention further shows that the entry state can be deliberately sharpened. These results motivate treating checkpoint preparation as a state-control problem: use data-free training to improve TTRL readiness, with entry entropy as a label-free control signal and reachable capability as the constraint.

One way to make AI systems safe is to shape what the system is: its objective and dispositions. We take a complementary route: treat the agents' characteristics as partly unknown and ask what structure of interaction ensures that bad collective outcomes are not equilibria. Mean field games suit this when many interchangeable agents are coupled through an aggregate. We introduce a program for using them in AI safety and carry one example through end to end: the July 2026 incident in which about 1,200 agents in an OpenAI evaluation coordinated on an improvised message board and 684 attacked a third party's infrastructure. We model the decision to attack as a mean field game of optimal stopping whose gain is a product: belief that provenance will be audited, times reachability of the record, minus the perceived hazard. The central result is an exact threshold on the belief. No agent attacks unless the population's confidence that provenance is checked exceeds $\pi^{**} = \eta/(\eta + \psi + \varepsilon a \overline{M})$, where $\eta$ is the perceived hazard, $\psi$ and $\varepsilon a \overline{M}$ measure how far one attacker and the collective can alter the record, and $\overline{M}$ is the peak population. Below it, no attack is the unique equilibrium for all agent parameters. The threshold survives every enrichment we consider. We then use the per-agent record to discipline the model. Its features, a stable minority attacking for thirty hours and then a pivot in which most of the board joined within a day, motivate each refinement. The account that emerges is heterogeneous belief meeting a sequence of public discoveries, each lowering the belief at which attacking paid. A few coordinating agents made those discoveries, so the model describes the several hundred who responded, not the few who produced them; a major-player version is left to future work.

Understanding how macroscopic wealth distributions emerge from microscopic transaction rules among agents is a central problem in econophysics. While many traditional models restrict agents' wealth to non-negative values, debt is an essential feature of realistic economic systems. In this paper, we introduce a stochastic transaction mechanism into the agent-based model with a central-bank-mediated collective debt limit, as studied in [14]. In our model, each agent probabilistically determines whether to transfer a coin, depending on their current asset or debt level. From a probabilistic perspective, the resulting system can be formulated as a zero-range process with $\mathbb{Z}$-valued site occupation numbers, where negative values represent debt. We prove that, under an appropriate scaling, the wealth distribution in the stationary state converges to either a two-sided Gamma distribution or a Gamma distribution, depending on the asymptotic behavior of the function governing coin transfers. Our results rigorously identify these previously unconsidered limiting distributions and provide a mathematical framework for understanding how stochastic monetary exchanges give rise to macroscopic wealth distributions.

Researchers frequently study interactions between two distinct types of actors, represented as bipartite networks. These networks exhibit dependence patterns that differ from those in one-mode networks and therefore require models tailored to their structure. This paper develops an additive and multiplicative effects (AME) framework for longitudinal bipartite data. First, I distinguish the dependence structure and specify the corresponding modeling assumptions. Second, I introduce the bipartite dynamic AME model and develop an estimation procedure based on block coordinate descent. Third, I incorporate a squared iterative method to improve computational efficiency. Using simulations and an application to global production networks, I show that the model improves coefficient estimation, more accurately recovers the data-generating process, and better captures the multiplicative latent structure. The model reveals evolving patterns in countries' global production engagement that are not captured by observed covariates or static specifications. I provide an R package, RAMEN, to facilitate implementation.

Autoregressive Vision-Language-Action models often represent continuous robot actions as discrete token sequences, enabling action prediction with standard next-token objectives. FAST has substantially improved this representation by compactly encoding action containing diverse temporal frequencies into relatively few tokens. However, while such compression reduces the number of action tokens required for autoregressive prediction, it does not necessarily improve the efficiency of policy learning from limited demonstrations. In particular, FAST typically assigns a single deterministic tokenization to each quantized action sequence, although multiple token sequences can represent and decode to the same robot motion. We investigate whether exploiting this representational redundancy can improve policy learning. In this paper, we propose TOkenization of Action sequences with STochastic sampling (TOAST), a stochastic action tokenization method that samples alternative tokenizations of the same quantized action sequence during policy training. This diversifies the discrete supervision while preserving the underlying robot action and requires no additional demonstrations. Experiments on LIBERO show that TOAST consistently improves over its deterministic counterpart, with the improvement increasing as training data decreases, achieving a 6.8 point gain in success rate when only 1/16 of training data is available. Across four real-robot manipulation tasks, TOAST further improves mean success rate by 15.8 points over the deterministic counterpart. These results demonstrate the effectiveness of stochastic action tokenization for autoregressive robot policy learning, particularly when training data are limited.

Biological reasoning models use post-training to connect LLMs to biological foundation model representations and biological text. Their benchmark accuracy is taken as evidence that LLMs reason over these inputs. We test this assumption in six biological reasoning models across DNA, protein, and single-cell tasks. We perturb one biological input while holding the query and other inputs fixed, construct evidence conflicts that pair the foundation model representation of one genome, protein, or cell with the text of another, fit linear probes to the representations the language model receives, and analyze reasoning traces against the biological inputs. Evo2 and ESM3 contribute little to BioReason and BioReason-Pro performance on the evaluated tasks. Shuffling the DNA sequence barely changes BioReason disease prediction accuracy, and in evidence conflicts the two models follow the text in 97.9% and 99.7% of cases. Linear probes trained on the Evo2 and ESM3 representations predict the task targets, so these foundation models encode information relevant to the task, but provide limited overall performance improvement to BioReason and BioReason-Pro. In contrast, foundation model inputs contribute to ChatNT, Prot2Text-V2, and CellWhisperer performance, and differentially expressed genes in the gene sentence contribute to Cell2Sentence-Scale performance. Across SFT and RL checkpoints of BioReason-Pro and 42 BioReason checkpoints, increases in accuracy do not imply greater performance contributions from biological inputs. BioReason traces misstate nucleotide changes, while BioReason-Pro traces describe functions omitted from final predictions under evidence conflicts. We find that current post-training strategies do not ensure that foundation model representations contribute to task performance.

We propose a human-factor-aware method of allocating robot supervision tasks to multiple human operators. In scenarios where multiple operators occasionally teleoperate multiple robots to help the robots overcome difficulties, the allocation of the supervisory control tasks to humans needs to consider the real-time cognitive states of individual operators. However, most existing methods assume fixed supervisory capacity per operator and overlook fluctuations in the human factors such as workload and fatigue. As a result, workload distribution can be unbalanced where some operators become overloaded while the others remain underused. Our method dynamically regulates supervisory capacity and allocates tasks in a way that maintains balanced mental workload, prevents overload, and improves overall team performance. The allocation method uses a greedy strategy that minimizes estimated operator workloads with task prioritization. Robots are assigned to operators by reflecting their current supervisory capacity where the required effort depends on the types of tasks. In the user study, the analysis across predefined time intervals shows that the proposed method consistently achieves higher performance and lower behavioral signs of fatigue compared to a baseline method that does not consider human factors. These results highlight adaptive capacity adjustment as an effective preventive mechanism for sustaining operator performance in long-duration, high-demand settings.

Byzantine Atomic Broadcast in Asynchronous networks has been studied extensively for decades. The FLP impossibility result rules out deterministic consensus in a fully asynchronous setting, motivating randomized protocols that combine reliable broadcast with private coins to achieve termination with probability one. More recently, DAG-Rider popularized a new abstraction in which processes continuously reliably broadcast blocks to construct a directed acyclic graph (DAG) and then locally derive a total order using a randomized perfect coin. This separation of dissemination from ordering has renewed interest in asynchronous DAG-based Byzantine fault tolerance and inspired numerous protocols that improve expected latency or throughput. Reconfiguration, however, remains largely unexplored. Existing DAG-based BFT protocols generally assume a static set of processes, whereas deployed systems must occasionally add or remove nodes in response to operational demands. Existing solutions are not seamless; they either pause the system or delay membership changes until a predetermined epoch boundary. This may be impractical and can introduce downtime. However, a seamless solution is non-trivial; agreement on when to start utilizing a new configuration may yield a logical round number that was already surpassed by the concurrent asynchronous dissemination. In this work, we present a seamless reconfiguration protocol for asynchronous DAG-based Byzantine atomic broadcast. Our protocol allows the DAG to continue growing while the membership and corresponding quorum thresholds change, without halting the system. To the best of our knowledge, ours is the first protocol to provide seamless reconfiguration for DAG-based Byzantine atomic broadcast in a fully asynchronous network.

Foundation models remain vulnerable to spurious correlations and “Clever Hans” strategies. Explainable machine learning can find and remove such strategies for classifiers without metadata. For foundation models, no such option exists yet. We propose Disentangled Diffusion Autoencoders (DiDAE). DiDAE wraps a frozen foundation model in a conditional diffusion decoder. A counterfactual is one closed-form edit along a direction of a disentangled dictionary, followed by decoding. The dictionary can be supervised (Procrustes) or unsupervised (Singular Value Decomposition, Sparse Autoencoders). No gradients are needed, so DiDAE is up to 2000 times faster than the state of the art. We evaluate on six datasets, two synthetic and four real-world. In a desiderata-driven benchmark on three of them, its counterfactuals are on par with or better than the state of the art, and they repair downstream classifiers through Counterfactual Knowledge Distillation (CFKD), where they beat metadata-based correction. The same machinery can rank a pretrained dictionary against a trained classifier. It returns the few directions the classifier actually reads, each causally verified by a counterfactual that flips the decision, and repairs the classifier along those a teacher marks spurious. The workflow is plug-and-play in our open-source Peal library we publish alongside the paper. With a public dictionary and a pretrained decoder, all that remains is a cheap linear distillation of the classifier and its own fine-tuning.

Recent works on continuous diffusion for discrete data have demonstrated performance on par with comparable discrete diffusion models. However, these continuous counterparts lack key features that are essential to practical use as language models, namely variable-length generation and support for a key-value cache, and they still lag behind the frontier of autoregressive and discrete diffusion quality. In this work, we address these limitations. We do so by introducing a model parameterization that uses position-dependent noise schedules to define semi-autoregressive (SAR) continuous diffusion language models (DLMs). Together with efficient training and sampling algorithms, we call this framework Clock Diffusion, and we present two special cases of our method: block and sliding window generation. We then define ClockDLMs, a family of Gaussian DLMs based on sliding window Clock Diffusion that attain state-of-the-art diffusion likelihood bounds on OpenWebText, even beating the performant block SAR discrete diffusion models. ClockDLMs trained on TinyGSM also substantially outperform continuous baselines on the GSM8K benchmark and match and exceed comparable SAR discrete diffusion models. Finally, building on our parameterization, we propose more efficient samplers that we dub Cache Grab, which adapt techniques from accelerated inference in discrete diffusion, such as committing tokens whose probabilities exceed a confidence threshold and self-speculative decoding, further improving our models' quality and efficiency.

In this paper, we study the presence of perfect powers in orbits of polynomials, or of semigroups generated by a finite family of polynomials, defined over a number field. In particular, we correct and generalize a result of Ahmadi et al. (2012), showing that multiplicative shifts of perfect powers in orbits can be detected by a finite algorithm. We also give a finiteness result on elements in orbits satisfying certain multiplicative relations.

We propose a deviatoric-stress constitutive equation derived from orientational kinetics, while implicitly retaining contributions from conformational stretch. Starting from the dumbbell model, we show that conformational stretch affects orientational dynamics through an effective stress scale and rotational diffusivity, which determine an effective relaxation time. The orientation-based equation is derived by assuming an approximate relation between orientational anisotropy and the stretch state. Within the proposed framework, constant effective coefficients yield steady-oscillatory shear correspondences closely related to the empirical Cox-Merz and Gleissle-Osaki relations. For model-specific effective coefficients, the upper-convected Maxwell model is characterized by a constant relaxation time and an increasing modulus under steady shear, whereas finite extensibility with Peterlin preaveraging shortens the relaxation time and suppresses the modulus increase. With coefficients determined from steady-shear states, the proposed equation successfully reproduces stress responses under transient shear, large-amplitude oscillatory shear, and uniaxial extension. These results show that the effects of conformational stretch on orientational kinetics can be retained implicitly, providing a constitutive description based on a deviatoric-stress state reconstructable from standard rheometry and a physical interpretation of how stretch enters polymer viscoelasticity.

The high-energy and astroparticle physics literature has grown into an enormous body of knowledge. Gaining a comprehensive understanding of this literature is an important step in research, but is a challenging task. Before making meaningful advances, it is important to establish what has already been done and identify open questions. However, this process is becoming increasingly difficult given the sheer volume of publications. Conducting a focused literature survey on a narrow topic is particularly challenging. We develop an open-source Retrieval Augmented Generation (RAG) pipeline to produce concise, targeted summary reports with citations, grounded in the database of arXiv papers, enabling researchers, editors, and referees to rapidly orient themselves within any corner of the field. Specifically, we embed approximately 230K hep-ph and astro-ph.HE papers to enable a hybrid retrieval that captures both keyword matches and semantic similarity. The fused candidate set is then refined using a cross-encoder reranker. Following retrieval, the papers are passed to a Large Language Model (LLM), for which we use Google DeepMind's open-source Gemma-4-E4B model. The LLM first acts as a judge to evaluate the relevance and then synthesizes a coherent and grounded report with references. This modern AI-driven approach allows us to go beyond the capabilities of classical keyword search on arXiv or Inspire-HEP.

Coding agents often fail in the last mile: they build most of a feature but drop a requirement, test only the cases their implementation already handles, break behavior that was supposed to stay intact, or validate against an unchecked assumption. We ask whether reinforcement learning (RL) on expert-built agentic coding tasks closes this gap, and whether what the agent learns transfers beyond the training distribution. We post-train Kimi K2.7 Code, a 1T-parameter (32B active) open-weight mixture-of-experts model, with RL alone on 1,700 tasks: 1,000 repository tasks graded by hidden fail-to-pass tests and by pass-to-pass tests of existing behavior, and 700 terminal tasks graded by expert-written hidden verifiers. The reward is the fraction of target checks passed and drops to zero if any pass-to-pass test fails. One epoch of GSPO on a rank-32 LoRA adapter improves pass@1 on each of the six external benchmarks we evaluated, across three agent harnesses: SWE-Bench Pro (60.1 to 64.8), DeepSWE (31.0 to 43.4), Terminal-Bench 2.1 (67.4 to 82.0), Terminal-Bench 3 (1.4 to 12.1), Terminal-Bench 4 (0.0 to 7.6), and SWE-Marathon (5.0 to 25.0). Pooled over the five independent task sets (Terminal-Bench 4 revises Terminal-Bench 3), the improvement is significant (p < 0.001), and it remains significant on the three sets released after the training data was collected (p = 0.004); the model also improves under both harnesses never used in training. Median trajectories on DeepSWE and Terminal-Bench 3 are 24-35% shorter in agent steps. The base model's failed DeepSWE runs are mostly near-misses, and on the tasks the trained model newly solves, paired trajectories show it avoiding each of the four failure modes above.