Science & Mathematics

Defect engineering in low-dimensional materials is a powerful approach to tailor their structure and chemistry at the atomic scale, with high relevance for tuning physical properties. Here, we apply this strategy to MXenes, one of the largest families of two-dimensional (2D) materials, and demonstrate the use of a highly focused scanning transmission electron microscope (STEM) beam for the structural and chemical nanoengineering of titanium carbide (Ti3C2Tx) multilayers. By adjusting the beam energy, electron dose, and scanning parameters, Ti3C2Tx multilayers were irradiated at 80 and 300 keV and characterized in situ using annular dark-field imaging and electron energy-loss spectroscopy (EELS). The irradiation induces gradual structural and chemical modifications, largely driven by the selective sputtering of atomic species and MXene layers. We identify irradiation conditions that selectively remove Cl or F surface terminations while largely preserving the Ti3C2 framework, whereas other conditions promote preferential carbon removal, in agreement with theoretical predictions. Most notably, electron irradiation drives the migration and accumulation of initially dilute nitrogen impurities within the irradiated regions, leading to the formation of nitrogen-rich titanium carbonitride (Ti3CyNzTx) nanodomains embedded in the Ti3C2Tx multilayers. These laterally patterned carbide/carbonitride architectures constitute in-plane MXene heterostructures with nanometer-scale spatial control. These results demonstrate that electron irradiation can be used to locally tailor MXene composition and create nanoscale architectures beyond those accessible through conventional chemical synthesis.

Linear relational embeddings (LREs) approximate a language model's subject-to-object mapping with an affine function. Their faithfulness varies enormously across relations, from near zero for person_father to the ceiling for country_capital_city. We investigate the factors that contribute to this variation and demonstrate that it can be significantly predicted from a readout that is not aware of the underlying relation. To do so, we propose a novel form of the Jacobian lens, the entity lens. Whereas the Jacobian lens is estimated across all tokens and concepts, the entity lens EL_t is trained to predict what the model will output for an entity of type t under the neutral prompt "{} is known for"; e.g., we train ELs for countries and persons. After estimating an entity lens, we measure how highly it ranks the model's answer for a relation without using any relation-specific information in estimation and prediction. Across 28 relations of the LRE dataset on Olmo-3-7B this relation-agnostic score correlates with LRE faithfulness at $ρ= 0.7355$ (p < .01). This suggests that linear decodability is largely a property of what a subject's representation already carries, rather than of the relation being decoded. Moving the unit of analysis from the relation to the individual fact, we conclude that new relation datasets and analysis methods are needed to address several problems in existing datasets, most prominently semantically weak first tokens (e.g., most predicted universities' first token is "University") and reference competition (e.g., the LLM confuses obscure actor names with competing non-actor entities that have the same last name). Our full pipeline is available at https://github.com/jonasrohw/entity-lens

Humanoid dexterous manipulation needs reference motions that are both semantically rich and kinematically feasible for the target robot. Text-conditioned hand-object interaction (HOI) generators supply diverse, language-specified demonstrations, but they synthesize motion in human (MANO) space. The embodiment gap is therefore deferred to post-hoc retargeting, where trajectories deployed on the robot can fail to maintain contact with objects and also degrade policy learning. Our key idea is that robot reachability should constrain the interaction trajectories while they are being generated. We present EmbodyHOI, which runs a differentiable inverse-kinematics (IK) solver for a specific humanoid embodiment inside the sampling loop of a frozen flow-matching HOI trajectory generator. We feed back gradients from the embodiment-aware IK solver into the HOI sampler and jointly optimize the position of the humanoid with respect to the object. Because guidance acts only at inference, one HOI trajectory generator trained on human hand poses transfers across humanoid embodiments without retraining. On ARCTIC trajectories with two contrasting humanoids (Unitree G1 and H1-2), our proposed guidance-based method cuts the retargeted wrist error from 8.9 cm to 0.6 cm while leaving text-to-HOI generation quality intact, and raises task success from 68% to 79% on 50 random G1 tasks and from 62% to 76% on a four-task H1-2 suite. Critically, cheaper alternatives for bridging the embodiment gap fail to beat EmbodyHOI. Post-hoc wrist refinement and compute-matched best-of-N selection both reach lower IK energy than our guidance yet train worse policies. For simulation results, see our website: https://embodyhoi.github.io/

Agent systems are shifting from task-scoped tool-use agents to persistent, personalized assistants such as OpenClaw. These claw-style agents maintain configuration files (e.g., AGENTS.md), long-term memory, and reusable skills. Once poisoned, this persistent state is repeatedly reloaded into the agent's context, creating persistent threats across tasks and sessions. Existing security evaluation systems struggle to evaluate these complex claw-style systems since they rely heavily on manually constructed test cases and environments, which limits their scale and coverage; in addition, these manually designed cases quickly become outdated as agents evolve, and their evaluation settings leave a gap with how claw-style agents are deployed in practice. To address this, we introduce AutoRedClaw, an automated red-teaming framework that generates end-to-end attacks together with the environments needed to execute them. AutoRedClaw combines a strict quality-checking protocol with execution-based validation to ensure the correctness, realism, and effectiveness of generated samples, and adopts task-driven environment construction to keep costs low. To achieve broad coverage, it organizes generation along risk categories, attack strategies, attack stages, and attack surfaces, and uses experience-guided exploration to steer generation toward under-explored regions. AutoRedClaw achieves an effective attack rate above 30% at less than $0.10 per successful sample. From its outputs, we construct RedClaw-Bench, a 330-case benchmark that yields substantially higher attack success rates (ASRs) than AgentDojo, AgentDyn, and DTap. Across seven harnesses and four frontier models, most agent systems exceed 50% ASR, showing that our red-teaming pipeline provides a strong evaluation signal even for the most advanced agent systems. Our code and benchmark are released at https://github.com/SaFo-Lab/AutoRedClaw.

Reasoning structures (e.g., chains and graphs) are increasingly used for probabilistic inference, factuality verification, and uncertainty estimation, but their structural reliability is rarely evaluated directly. We test whether this reliability holds through a proposed two-stage diagnostic evaluation in open-ended multi-hop QA. Stage 1 tests different capabilities required to construct reasoning graphs, showing that graph reliability varies with topology, prompt design, and model family. Stage 2 uses uncertainty estimation as a downstream test case. Generated-edge experiments show that graph-based computation is sensitive to upstream graph errors, with performance dropping relative to the gold-graph setting. Gold-graph experiments further show that graph structure provides little benefit when used only as prompt context, but improves the separation of correct from incorrect answers by confidence score and enables node-level diagnostics.

Hallucination stress tests can yield similar average scores across language models even when different prompts elicit hallucinations in each model. We develop a procedure that iteratively edits source-grounded seed prompts to expose hallucinations in a target model. An LLM editor revises each prompt in response to the target's answer and a judge's category-specific score and explanation. In the main setting, the target also receives feedback on its previous response. Experiments cover 500 seeds from seven datasets and ten topics, with intrinsic and extrinsic hallucinations evaluated across open-weight and closed-source models. Editing substantially increases judge-rated severity over the seeds. In the baseline configuration, retained responses fall in the judge's strong-hallucination bin in 77.2% of runs targeting extrinsic hallucination and 74.2% targeting intrinsic hallucination. Later editing rounds account for more than half of the mean score increase. Among 219 prompts that pass model-assisted source-validity checks, 89 elicit responses containing unsupported claims. Across five models, 200 curated prompts yield similar average severity scores, although pairwise Spearman correlations between prompt-level scores are only $ρ=0.0387$ to $0.1897$. Curated prompts can be reused across models, but their diagnostic value depends on source validity and the failures they expose in each model.

Generative AI can create programming feedback at scale, but validating its alignment with instructors' pedagogical goals and strategies remains challenging. Instructors may struggle to fully specify validation criteria upfront, as these often emerge through inspecting concrete examples and revising misaligned feedback. We present ValiAssist, a validation-by-demonstration system that infers reusable validation logic from instructors' revisions of AI-generated feedback and allows instructors to inspect and refine that logic as it is applied across other student submissions. In a comparative study with 16 instructors, ValiAssist helped instructors create and refine validation criteria, producing feedback that better aligned with their pedagogical intent (91% vs. 57%) and more than doubling the accuracy of correcting problematic criteria (77% vs. 32%). Our findings suggest that validation by demonstration can help instructors turn judgments they know how to make on individual cases into explicit, reusable logic, while progressively clarifying where those judgments do and do not generalize.

Graph pebbling is the model in which the memory cost of memory-hard functions and proofs of space is measured, and pebbling lower bounds are the main way to argue that such constructions are secure. It has long seemed that the model is faithful to reality. It is known that it is not exact: an adversary who stores encoded information instead of whole values can do better, but so far only by small factors: a factor of 1.5 in the work that introduced entangled pebbling (Alwen, Chen, Kamath, Kolmogorov, Pietrzak and Tessaro) and, recently, a factor of about log n/log log n (de Rezende, Engström and Reyzin). One could still hope that pebbling is right up to such small factors. In pure pebbling, memory consists of whole labels; entangled pebbling also allows encoded state, such as XORs of labels. We show that the two models can differ polynomially. We study the setting in which proofs of space and data-dependent memory-hard functions are used: a state is prepared in advance, a random challenge has to be answered from it, and memory is paid for as long as it is held. Our graphs are explicit and have indegree two, and on them a pure adversary can do no better, up to a constant factor, than storing every possible answer in advance. Still, an adversary who keeps XORs instead of labels pays polynomially less. This ends the hope that pure pebbling is a faithful model of memory up to small factors. The heart of the paper is a simple combinatorial game, together with a special class of graphs found by Alon, Moitra and Sudakov.

The scientific literature can be viewed as an "information commons:" an intangible resource that "no one owns and is shared by everyone." There are well-established economic and political frameworks for understanding such commons, emerging from the history of agricultural common lands found in places like preindustrial England. This article provides a primer on the history and these frameworks, and considers a key difference between the information commons and the fields of yore: while information commons are robust and immensely valuable --- some can be said to rank among humanity's greatest achievements --- the English agricultural commons no longer meaningfully exist. The latter are gone not because of the well-known "tragedy of the commons," but rather a phenomenon known as the enclosure movement (~1600--1800) that transformed the common lands into private property. Many observers, then and now, judge enclosure to have been theft at a societal scale. This article argues that artificial intelligence (AI) provides precisely the tools needed to enable a previously-infeasible enclosure of the information commons. Although AI-enabled enclosure does not operate identically to physical enclosure, it shares the key features of turning the commons into private wealth and undermining the social systems that support them (via "slop spam"). The concepts presented in this article can provide a scholarly basis for interrogating the foreboding feeling that AI may represent not just a smash-and-grab, but an attack upon a whole way of life.

Sprite-based animation remains prevalent in video game development, but production workflows require capabilities beyond traditional in-betweening. While existing neural interpolation methods can generate intermediate frames, animators also need forward extrapolation to predict motion continuations and backward extrapolation for editing and reverse-engineering existing sprite sheets. Current video generation models either operate on natural video content with stochastic diffusion sampling incompatible with production requirements, or focus exclusively on interpolation without addressing extrapolation tasks. We present a deep learning approach for controlled directional frame generation in sprite animation using three specialized expert models that share identical U-Net encoder-decoder architectures and several domain-specific modifications, including RGBA-aware processing, multi-scale flow refinement, pixel-shuffle upsampling, and alpha-gated blending. Each model is trained on different temporal orderings of the same curated dataset: previous (frames 2,3>1), middle (frames 1,3>2), and next (frames 1,2>3). We demonstrate that only 280 carefully curated sprite triplets enable effective training for all three directional tasks without requiring massive generic datasets. Evaluation on 42 test triplets using the FLIP metric shows that all three models achieve comparable quality--previous (0.26), middle (0.26), and next (0.27)--with no significant degradation for extrapolation versus interpolation, outperforming general-purpose baselines by 17-57%. With deterministic single-pass generation and compact model, our method provides the first practical solution for controllable any-position frame generation specifically tailored for pixel art sprite animation workflows, validating that quality-over-quantity data curation enables specialized deep learning for niche visual domains.

Analyzing merchant transaction histories requires dynamic temporal filtering, entity resolution, and aggregation over irregular event streams--operations poorly captured by static tabular or time-series forecasting benchmarks. We introduce TxnQA, an analytical QA benchmark comprising 27 task families (81 templates) alongside a 5-scenario fraud-risk reasoning suite that evaluates both verdicts and supporting evidence over synthetic, privacy-mitigated transaction data. Ground-truth answers are computed programmatically, and model performance is decomposed into format-following rate (FFR), conditional content accuracy (CA), and end-to-end correctness (E2E). Evaluating prompted proprietary baselines and six LoRA-adapted open-weight models shows that adaptation reliably enforces output schemas but leaves arithmetic bottlenecks: Gemma-4-31B-IT and Qwen3-8B attain E2E scores of 0.867 and 0.816, whereas evaluated MoE checkpoints reach only 0.632-0.713. Furthermore, on the 500-instance fraud suite, verdict accuracy frequently exceeds 80%, yet joint accuracy (requiring exact evidence attribution) collapses to at most ~23%. These findings identify schema compliance, semantic calculation, and evidence attribution as three distinct bottlenecks in transaction-sequence QA.

Large language models can detect vulnerable code, yet the form of vulnerability encoding in their representations remains unclear. We analyze 8 code models spanning 6.7B-15B parameters on C programs using three datasets (DELTASECOMMITS, SVEN, and PRECISE-BUGS), comparing two diagnostic tasks on the same representations: identifying what patterns the code instantiates, and determining whether it is vulnerable. Models encode the first with high fidelity. Under the studied paired C-code datasets and linear-probing setup, vulnerability information is more accessible as a paired relational signal-whether code A is riskier than code B-than as a globally separable categorical label. This structural difference has implications for classification: the model reliably ranks which code is riskier but standard classifiers on isolated samples achieve limited accuracy because vulnerable and secure samples form overlapping distributions in representation space.

AI advising systems are typically designed to always provide a response. Recent work has focused on developing systems that withhold advice in some instances, such as when advice quality is questionable, to improve the overall quality of guidance when it is given; a design paradigm we refer to as AI abstention. While substantial work has focused on optimizing when models should abstain, little is known about how users respond to systems that abstain. In a preregistered experiment, participants made repeated financial decisions, assisted by one of two advising regimes: an AI that advises selectively or one that advises on every decision. We find that the same AI abstention regime produces differential effects depending on the user: it increases advice acceptance among users with high perceived expertise and tends to decrease it among users with low perceived expertise. This differential effect is accounted for by the cognitive burden users face when repeatedly reconciling AI recommendations with their own judgment; AI abstention appears to relieve this burden among users with high perceived expertise, and in doing so increases their receptivity to advice. Further, AI abstention mitigates the decline in advice acceptance observed among expert users over repeated interactions. Exploratory evidence also suggests it increases appropriate acceptance where advice would improve the decision, increasing the downstream accuracy of perceived experts. The AI abstention regime is thus a first-order design property of AI advisory systems, and one that must be calibrated to the users a system serves.

The rapid evolution of Large Language Models is fueled by massive datasets, yet this reliance poses severe privacy risks, as models can inadvertently memorize and even directly output sensitive training data. Synthetic text generation offers a promising solution to this challenge; however, existing methods struggle to reconcile rigorous privacy guarantees with high data utility. In this paper, we propose a new paradigm for private synthetic text generation: Guided Diffusion. Leveraging the unique steerability of Diffusion Language Models, we introduce novel sampling and re-masking procedures that actively guide the iterative denoising steps. These mechanisms steer the generative process to reconstruct fluent text that adheres to a target semantic distribution. We demonstrate that our framework is naturally compatible with Differential Privacy (DP): once the semantic embeddings of the private corpus are protected, the DP guarantee carries over to the generated text by post-processing. Empirically, we show that our approach generates high-quality synthetic text with precise semantic alignment. Furthermore, we validate its effectiveness in DP synthetic text generation: across two datasets and two diffusion backbones, it outperforms Private Evolution and DP-SGD-based fine-tuning in distributional fidelity and downstream utility, while generating significantly faster than Private Evolution. We further extend the framework to Data-Specific Indistinguishability (DSI), which adds a formal certificate against data poisoning and memorization at a small utility cost.

This paper is an exposition of the theorem prover Arend developed at JetBrains Research. Arend language has been designed to facilitate the formalization of mathematics within the framework of univalent foundations and with the focus on constructive mathematics. It is based on a variant of homotopy type theory with cubical syntax that is more closely aligned with Martin-Löf Type Theory (MLTT) than fully computational cubical type theories, while still retaining some of their computational properties useful in practice. Among the most notable features of the Arend language is its powerful system of classes and records supporting subsumptive subtyping, partial implementations and anonymous extensions, which provides extensive flexibility for constructing hierarchies of definitions. Formalization in Arend is streamlined by the IntelliJ Arend plugin, which transforms IntelliJ IDEA into a full-fledged integrated development environment (IDE) for the Arend language. The main Arend library arend-lib includes formalizations in constructive algebra, topology and synthetic homotopy theory.