Papers

Generating corner-case scenarios with appropriate adversity in a simulation environment is critical for testing an autonomous vehicle (AV) software stack's safety performance before deployment. Existing autonomous-driving scenario generators can enforce specific behavior, adversity, or feasibility conditions, but they provide limited control over how extreme a generated scenario is relative to plausible futures in the same traffic context. This study represents the adversity of a generated scenario as its percentile in the conditional distribution of future risk given the observed history. This view supports calibrated answers to two questions: how "corner" a generated corner-case scenario is and how its "cornerness" can be fine-tuned. To this end, we formulate history-conditioned risk-percentile requests and learn a reference risk distribution that maps each requested percentile to a physical risk target. We then use a percentile-conditioned joint diffusion model with sampling-time risk guidance to generate multi-agent futures, together with a reference-based criterion for evaluating percentile realization. Experiments use the minimum post-encroachment time (PET) between the ego and its surrounding vehicles as the risk surrogate on highD. On the primary evaluation set, our method realizes 1,422 of 1,440 requests within a 0.05 percentile tolerance (98.75%), with mean percentile error 0.00673 and PET-target error 0.00991 seconds. The resulting interface connects context-relative risk specification, physical realization, and evaluation through a common risk scale. Project website and videos of generated scenarios are available at https://hhj233.github.io/CornerPercentile/.

The growing demand for large-scale database search in data-intensive applications, ranging from proteomics to autonomous systems, has exposed fundamental limitations in von Neumann architectures due to memory bandwidth and energy constraints. Hyperdimensional (HD) computing offers a robust and parallelizable framework for such tasks, but its practical implementation remains challenged by high memory demands. Ultra-high-density, energy-efficient ferroelectric NAND (FE-NAND) memory provides a potential solution by enabling in-situ computation. We fabricate quad-level FE-NAND strings with wide memory windows, disturb resilience, and robust retention. Using these planar FE-NAND strings as building blocks, we experimentally demonstrate in-situ multi-level cell (MLC) dot product operations at the single-cell level and, using experimentally calibrated physics-based simulations, demonstrate Hamming similarity calculations between reference and query hypervectors (HV). This platform leverages the inherent error tolerance of HD computing to achieve >90% search accuracy even at high logic levels (TLC, QLC). When benchmarked on Open Modification Search (OMS) tasks in proteomics with TB-scale datasets, our system shows nearly 1,000x speedup and over 10,000x energy efficiency improvement compared to incumbent solutions. These results establish FE-NAND as a viable in-storage compute architecture for large-scale, high-dimensional data processing.

Large language models are increasingly used to generate test scenarios, executable test suites, assertions, interaction sequences, and fuzzing infrastructure. These artifacts serve different purposes and rely on different sources of information about correct behavior. A test that increases implementation coverage, a test that agrees with a reference program, and a test that detects a requirements violation therefore provide distinct kinds of evidence. We present a structured narrative survey that connects the generation process to the evidence used to justify test quality. Drawing on 95 curated source records through 29 September 2026, we organize the field along four dimensions: testing objectives and artifacts, information available during training and generation, generation and learning mechanisms, and evaluation evidence. We synthesize work on unit and requirements-based testing, test oracles, API and GUI testing, fuzzing, and learned test generators. The resulting framework explains how execution feedback can improve executability while also affecting the independence of an oracle, why specification provenance and training-time reference supervision matter for comparisons, and how downstream code-selection gains differ from test correctness. We use these distinctions to organize benchmarks and derive a research agenda for independent oracle assessment, budget-aware generation, repository-scale evaluation, and test maintenance.

Concept erasure is essential for the safe deployment of text-to-image diffusion models, as they may reproduce harmful, copyrighted, or privacy-sensitive content learned from unconstrained large-scale data. Existing methods typically erase unwanted concepts while preserving general generation capability by redirecting target-related text-to-image mappings. However, recent studies show that erased models may still retain visual generative trajectories of target concepts, leaving them vulnerable to adversarial recovery attacks and revealing a fundamental gap between redirecting text-to-image mappings and truly removing visual knowledge. To bridge this gap, we propose VisualErase, a new paradigm that redirects concept-bearing visual generative trajectories toward explicitly defined concept-removed outcomes. To enable this redirection, we use structure-preserving image editing to construct content-aligned, concept-removed counterparts for source images, providing explicit visual endpoints that retain non-target content. We then derive a denoising target from each source-to-counterpart pair and use a dual-branch redirection loss to align both text-conditioned and unconditional predictions with this target, since conditional supervision alone does not explicitly constrain generation without textual guidance. To mitigate the adverse effects of concept erasure on non-target generation, we jointly optimize the redirection loss with a counterpart retention loss that matches denoising predictions from the frozen pretrained model. Across style, celebrity, and nudity erasure, VisualErase limits the maximum attack success rate over seven attacks to 0%, 8%, and 0.1%, respectively, while retaining general generation quality. These results highlight the importance of visual trajectory redirection for robust concept erasure beyond text-to-image mappings alone.

The theoretical understanding of multi-layer neural networks is largely confined to overparameterized settings, which obscure parameter identifiability and incur high sample complexity. Neural tangent kernel (NTK) provides a general theory for wide networks, but does not offer efficient sample-complexity guarantees. Recent feature-learning results go beyond kernel methods for single-neuron, multi-index, and hierarchical targets. However, the analysis is often restricted to shallow or specific architectures and to the overparameterized regime. We break this paradigm to achieve parameter-level recovery of deep target networks, albeit by using active data queries. Specifically, we study $L$-layer polynomial networks with even degree-$k$ monomial activations and nonnegative higher-layer weights. This structure makes the target network input-convex, while the optimization landscape remains highly nonconvex with respect to the parameters. Leveraging input convexity and active queries, we propose ASPIRE (Active SamPling for Iterative Recovery via Eigendirections), a layerwise sampling-based diagonalization algorithm that recovers all network parameters to $\delta$-accuracy with sample complexity $ \widetilde O_{k,L}\left(d^{L^2+O(L)}\delta^{-2e}\right) $ in polynomial time. To our knowledge, this is the first parameter-recovery guarantee for deep target networks whose exponent grows only polynomially with depth, as well as the first justification for the effectiveness of using high-quality data in neural network training, with a remarkably exponential separation.

Emotional expression can influence the safety decisions of large language models (LLMs), offering a potential avenue for improving safety alignment. Existing studies have mainly focused on how emotional expressions facilitate attacks under harmful requests, while overlooking their effects on benign requests. We find that emotional expression can also systematically increase refusal tendencies on benign requests, leading to unnecessary over-refusal. Based on this observation, we propose emotion-guided refusal subspace steering (EmoRSS), an activation-steering method that mitigates emotion-induced over-refusal while preserving refusal behaviour on harmful requests. Specifically, we first identify a refusal-sensitive layer using layer-wise linear probes and construct a refusal subspace from sparse autoencoder (SAE) features aligned with the probe direction. Next, we use paired regular and emotional requests with the same queries to estimate the mean activation shift in the features defining the refusal subspace. Finally, we decode this shift into an activation intervention vector and apply it in the reverse refusal direction during inference, without updating the backbone parameters. Experiments on two LLMs show that, when requests contain emotional expressions, our method achieves a more favourable trade-off between refusing harmful requests and answering benign ones than prior over-refusal mitigation baselines, while better preserving general task performance.

We study offline learning in two-player zero-sum contextual games with public and private information, motivated by strategic settings such as auctions and negotiations with private valuations. We introduce unilateral prescriptive concentrability and show that asymmetric information can change offline coverage through its effect on equilibrium behavior. For finite state-action spaces, we develop a pessimistic algorithm with an $\tilde{O}(1/\sqrt{n})$ exploitability rate, matching the standard sample-size dependence for fully observed minimax games. We further develop a pessimistic policy mirror descent framework, PPA-PMD, for general function approximation and obtain a unified $\tilde{O}(1/\sqrt{n} + 1/\sqrt{T})$ exploitability rate with no-regret actor updates. Together, these results provide the first theoretical framework for offline equilibrium learning under public-private information constraints.

Quantum algorithms enable Heisenberg-limited scaling across many tasks in sensing and learning, achieving precision $\varepsilon$ using $O(1/\varepsilon)$ total evolution time. However, in Hamiltonian property testing, the best known algorithms have $O(1/\varepsilon^2)$ total evolution time, unless it is possible to $\textit{reverse}$ the flow of time under the Hamiltonian. In this work, we prove that time reversal is necessary to close this gap. We establish evolution time lower bounds for Hamiltonian certification and locality testing, for a continuous family of error metrics ranging from average- to worst-case. We also prove matching upper bounds for Hamiltonian certification in all cases, both with and without time reversal, showing that the advantage from time reversal persists. Our results follow from a new Fourier-analytic approach to quantum lower bounds in continuous time, which is sensitive to time reversal and applicable beyond Hamiltonian property testing. This gives an alternative proof that inverses are required for amplitude amplification and estimation, improving the lower bounds of Tang and Wright and extending to fractional queries. We also show that, for unstructured search, uncertainty in the phase of the oracle destroys the quantum speedup without inverse queries, even when the oracle is fixed and coherent. Our work rigorously establishes time reversal as a resource governing both the precision of quantum measurement and the robustness of quantum algorithms.

We propose a protocol for the remote implementation of an arbitrary two-qubit unitary operation using pre-shared entanglement. In our protocol, Bob possesses an unknown two-qubit state, while Alice implements the required operations using calibrated black-box devices. Alice does not need to know the parameters of these operations, and Bob's unknown state never leaves his laboratory. The protocol is based on the Euler decomposition of single-qubit unitary operations and the Cartan decomposition of two-qubit unitary operations. First, we construct protocols for the remote implementation of the single-qubit operation $U = R_z(\gamma)R_x(\beta)R_z(\alpha)$ using pre-shared Bell states. We then construct a protocol for the two-qubit operation $R_{zz}(\theta)$ using a single pre-shared Bell state and extend it to $R_{xx}(\theta)$ and $R_{yy}(\theta)$ through local unitary transformations. These elementary protocols are finally combined to implement an arbitrary two-qubit unitary operation. A third party, Charlie, determines the required decomposition and provides Alice with the corresponding black-box devices before the protocol begins. Thus, neither Alice nor Bob requires the classical description of the unitary operation.

We prove the Baker–Rumely conjecture that the tau invariant of a metrized graph admits a positive lower bound proportional to its total length, with an absolute constant. We also construct simple cubic metrized graphs whose normalized tau invariants tend to $59/7260<1/108$, disproving the proposed universal constant $1/108$. The lower bound is independent of the genus, the number of edges, and the distribution of edge lengths. Its proof combines a second-moment inequality for Euclidean lattices with a partition of the edge coordinates of a cycle lattice into three independent sets. The counterexamples have only two edge lengths and admit an elementary resistance calculation. Through the tropical moment identity, the lower bound also gives a uniform estimate for the non-archimedean terms in height formulas for Jacobians.

For a positive integer $n$, let $$ S_1(n)=\{n\bmod k:1\leq k\leq\lfloor n/2\rfloor\}, \qquad s_1(n)=|S_1(n)|. $$ Baraskar and Vukusic proved $s_1(n)=\Xi n+O(n/(\log n\log\log n))$ for an explicit positive constant $\Xi$, and asked for a sharper error term. Their computations suggested that an $O(n^{1/3})$ bound might hold. We prove that, for every $\alpha>\sqrt2$, $$ \limsup_{n\to\infty} \frac{s_1(n)-\Xi n}{n\exp\{-\alpha\sqrt{\log n\log\log n}\}} =\infty, $$ and the corresponding limit inferior is $-\infty$. In particular, no fixed power saving is possible. We also prove the complementary upper estimate $$ s_1(n)=\Xi n+O\left( n\exp\{-\alpha\sqrt{\log n\log\log n}\} \right) $$ for every fixed $\alpha\in (0,\sqrt{2})$, sharpening the previous $O(n/(\log n\log\log n))$ bound. Thus, the constant $\sqrt2$ is the threshold for bounds of the form $n\exp\{-\alpha\sqrt{\log n\log\log n}\}$.

Detecting malicious smart contracts is essential to safeguarding the Web3.0 ecosystem. However, existing auditing methods based on large language models (LLMs) largely rely on prompt engineering or attack-specific fine-tuning, while the internal representations that support malicious logic detection remain poorly characterized. To address this gap, we propose ContractLens, a framework that localizes and uses latent security knowledge in pretrained LLMs for interpretable malicious smart contract detection. Specifically, we first use layerwise linear probes to identify candidate layers that distinguish malicious from benign contracts, then select critical layers based on their sensitivity to malicious logic removal and stability under behaviour-preserving surface modifications. Next, within these critical layers, we combine activation contrast, malicious-logic specificity, and probe weights into a tri-factor importance score and adaptively select malicious contract-discriminative units (MCDUs) using the knee point of the ranked score curve. Finally, we concatenate the selected units' activations into a compact representation and train a lightweight classifier for detection, keeping the pretrained model frozen throughout. Neuron masking experiments show that masking MCDUs causes larger drops in detection performance than masking randomly selected neurons or neurons matched in activation magnitude. ContractLens achieves higher average detection accuracy than the evaluated methods on all three pretrained models. Further evaluations demonstrate generalization to unseen malicious intents and show that, with the identified MCDUs fixed, a classifier trained on few labelled samples can still achieve strong detection performance.

Large language model (LLM) agents solve complex tasks through multi-step interactions with external tools. These interactions often contain recurring local tool sequences. Treating such sequences as composite "Skills" can shorten tool-use trajectories and reduce repeated low-level decisions. However, when atomic tools and composite skills coexist, skill use becomes a policy problem: the agent must decide whether the current state requires atomic fine control or skill-level abstraction. In this paper, we argue that effective skill use should be studied as adaptive tool granularity selection. The most direct training signal for this problem is to compare the consequences of atomic and skill choices available from the same state. Based on this view, we propose CIPO, a Counterfactual Imagination Policy Optimization framework for adaptive tool granularity. CIPO constructs executable skills through budget-constrained mining of successful tool-use trajectories and trains granularity decisions with counterfactual branch rollouts. For each base rollout, CIPO branches at the first eligible granularity decision and replaces the chosen action with a feasible atomic or skill alternative. The paired outcome difference serves as a supplementary reward for policy optimization. Experiments across multiple benchmarks and model backbones show that CIPO improves task success and decision efficiency over baselines. Further analyses show that CIPO learns effective skill use by improving the choice between atomic tools and composite skills based on the current state, without simply increasing skill frequency.

We study when to buy a share that will later be sold optimally, when the price follows a geometric multi-skew Brownian motion whose skew levels model support and resistance. The reward for buying is the exit premium of the liquidation problem solved in a companion paper. This premium is strictly $r$-subharmonic inside the exit continuation region, so buying is optimal only at levels where the price is pushed upward, and the entry problem reduces to a finite one whose value is an upper concave hull of finitely many points. With one support and one resistance, the optimal rule is to buy exactly at the support, in every exit regime, with an explicit value.

Observing a prediction error at one node can help correct predictions elsewhere, but the benefit depends on the residual dependence between nodes. A graph suggests where that dependence might occur, yet does not establish its sign or strength. We develop a Bayesian model of residual covariance to determine how revealed errors should update a fixed feature-based predictor. For a fixed set of revealed nodes, we derive an exact identity for the change in expected squared error under linear residual propagation. The identity characterizes the optimal linear update and expresses the excess error of any other update as an exact quadratic. Under covariance uncertainty, the Bayes-optimal fixed linear update uses the posterior mean covariance. We average over four positive-semidefinite covariance families representing no dependence, positive dependence, negative dependence, and dependence whose sign can alternate with graph distance. Across eight graph benchmarks, four show posterior support within the distance-profile family for negative covariance between neighbors and alternating signs with distance. Among the evaluated methods, ours achieves the smallest worst-case recall gap to the best method on each graph. The method improves squared error over residual propagation on five of eight graphs, but performs substantially worse on one dense heterophilous graph. In these experiments, selecting the highest-posterior family performs similarly to model averaging.

Latent world models plan by rolling a frozen predictor forward under candidate action sequences and ranking the candidates by the latent distance between their imagined end state and the goal. However, this ranking breaks down when the goal lies several plans away, because the latent distance measures how closely an end state resembles the goal rather than how far it remains from reaching it. To address this, we propose TEMPO, a temporal-distance planning objective that leaves the world model untouched, learns only from the recorded trajectories already used to train it, and adds negligible cost to the planner's search. TEMPO learns a small map of the frozen latent in which the distance between two states of an episode reflects the number of environment steps between them, and blends this distance into the planner's cost. It requires no rewards, policies or success labels and, being a cost rather than a model, applies to frozen world models with one latent vector per state that plan by a latent distance. We evaluate TEMPO on eleven simulated environments (e.g., maze navigation, tabletop pushing, robotic arm control and three-dimensional manipulation) with the LeWM and PLDM planners. With a small MLP that adds at most 0.3% to a plan's arithmetic, TEMPO improves both planners at every goal distance, including the one-plan setting of their evaluations, raises LeWM from 36% to 99% on TwoRoom three plans from the goal, and remains competitive on a broad range of 2D and 3D navigation, reaching and manipulation tasks.

Estimating the volume of a high-dimensional convex body is a fundamental problem in theoretical computer science. Given a quantum membership oracle for a convex body $K\subset\mathbb{R}^d$, we show that estimating $\mathrm{vol}(K)$ to relative error $\varepsilon$ takes $\widetilde O(d^{5/2}+d^{3/2}/\varepsilon)$ queries. This improves the previous quantum upper bound $\widetilde O(d^3+d^{9/4}/\varepsilon)$ of Chakrabarti et al. (ACM TQC 2023) and Cornelissen–Hamoudi (SODA 2023), and achieves a larger quantum speedup over the currently best classical upper bound $\widetilde O(d^{7/2}+d^3/\varepsilon^2)$ of Jia et al. (JACM 2026). Our quantum algorithm combines two new ingredients: an input-dependent effective spectral-gap analysis of quantum walks based on uniform classical warm-start mixing bounds, and a single quantum estimator for products of normalizer ratios in simulated annealing based on classical bridge sampling. We also prove an $\Omega(d)$ quantum query lower bound for constant relative error, improving the previous $\Omega(\sqrt d)$ lower bound.

Disintegrating rocky planets provide a rare opportunity to probe material escaping directly from intensely irradiated planetary surfaces. We present time-resolved, low-resolution ($R\simeq600$) optical spectroscopy and complementary ground-based photometry of BD+05 4868 Ab. The in-transit spectra exhibit an additional unresolved absorption feature near the O I triplet at 7771.94, 7774.17, and 7775.39 Angstroms relative to the out-of-transit baseline. The feature has an equivalent width of $W_\lambda\simeq1.4\pm0.2$ Angstroms, a maximum differential depth of approximately 9%, and an apparent width of 13-16 Angstroms, comparable to the instrumental resolution. We interpret the feature as candidate O I absorption from an oxygen-bearing mineral-vapor outflow. The $3s\,{}^5S^\circ_2$ lower level of the triplet lies 9.146 eV above the ground state and cannot be appreciably populated in LTE at the reported planetary temperature, requiring non-LTE excitation. A one-zone forward model in which photoionization, O$^+$ recombination, radiative cascade, and radiative trapping populate the excited state reproduces the measured equivalent width and instrumentally broadened line profile. The model favors an atomic component with a projected extent larger than the narrow condensed-dust layer and permits gas-to-dust mass-loss ratios of order a few for representative ionizing irradiation. Stellar O I line-profile variations remain the principal alternative at the present spectral resolution. We therefore identify escaping O I as the preferred physical interpretation while retaining the designation candidate O I. High-resolution, phase-resolved spectroscopy can directly test this interpretation by resolving the individual triplet components.

Jev turns natural-language questions into typed answers and probabilities with low latency and cost, enabling applications to route requests and select tools. While this interface allows Jev to integrate naturally into application workflows as a decision layer, the security and privacy implications of this emerging use remain largely unexplored. To address this gap, we conduct the first systematic study of these implications using the official Jev API and NanoJev, a local model with controllable training data and updates, focusing on three research questions: (1) What security threats arise when Jev is deployed as an application decision layer? (2) What private information can Jev reveal despite returning constrained typed outputs? (3) How can Jev's general-purpose decision capability be used for beneficial purposes or misused? Jev's decisions depend on application state and may be influenced by user-provided inputs. We therefore adapt prompt injection and adversarial suffixes to manipulate its decisions. Open-source Jev distribution and updates introduce supply-chain risks, which we examine by implanting backdoors in NanoJev through training data poisoning. Since Jev's outputs reflect both application state and information learned during training, we further adapt membership, private attribute, and internal knowledge inference attacks to recover sensitive information despite its constrained output format. Finally, Jev can serve as a general-purpose decision oracle for defensive and malicious workflows. We examine this dual use through four detection tasks covering prompt injection, jailbreak inputs, harmful content, and AI-generated text, alongside misuse scenarios involving jailbreak and model extraction. Our empirical evaluation shows that Jev remains vulnerable to the examined security and privacy threats, while its decision capability can support beneficial and malicious uses.

Nonlinearity can induce topological phase transitions in periodic systems. It is hence of considerable interest to examine if such transitions persist in the absence of translational symmetry. We address this question directly in real space, using the periodic Su-Schrieffer-Heeger (SSH) chain as a reference for understanding the effect of quasiperiodicity. How nonlinearity drives a chain towards a topological phase transition is found to depend strongly on the strength of the quasiperiodic potential. As the nonlinearity strength increases, we find that weak quasiperiodicity supports a transition from a trivial to a topological phase, whereas the intermediate regime does not have a topological phase transition. Under strong quasiperiodicity, pronounced boundary localization can develop well before a nonlinearity-induced topological transition. Our results show how nonlinearity may interplay with a quasiperiodic potential in inducing topological phase transitions and that boundary localization of nonlinear eigenstates alone may not be taken as a clear signature of topological edge states.

The ideal magnetohydrodynamic (MHD) system that governs a conducting plasma—the Maxwell equations together with the conservation laws of mass, momentum and energy—is the same local balance structure that describes hot accretion flows around black holes. We introduce a uniform four-dimensional spacetime dilation \[ \Phi_L:\;(t',x',y',z')\mapsto (t,x,y,z)=(Lt',Lx',Ly',Lz'),\qquad L>1, \] and show that it is a smooth, orientation-preserving $C^\infty$ diffeomorphism of flat spacetime. Pulling the MHD conservation laws back through $\Phi_L$, we prove that the closed ideal-MHD system is form-invariant (covariant), with the current density rescaling as $\bm j=\bm j'/L$, the magnetic diffusivity as $\eta'=\eta/L$, and external sources rescaled by $L$. The transformation realises three physically distinct effects on the perturbed flow: a linearization to the tangent-space (linearised MHD) system, a movie slow-motion dilation of dissipative and dynamical timescales, and an electron-microscope magnification of small-scale structure. Taking the formal limit $L\to\infty$ localises the dynamics on a neighbourhood of a horizon: resistivity vanishes, the asymptotic time derivative freezes, and all transient multipolar structure of the infalling plasma redshifts away. The remaining stationary, axisymmetric, asymptotically flat electrovacuum is uniquely the Kerr–Newman geometry, so that the exterior of the compact object is characterised only by its mass $M$, angular momentum $J$ and charge $Q$. The result provides a local, conservation-law-based perspective on the black hole no-hair theorem that is complementary to the classical Israel–Carter– Robinson theorems and to contemporary gravitational-wave and Event Horizon Telescope tests.

A new form of the fiscal multiplier suited to data expressed as gross growth rates is proposed, together with a structural vector autoregression that isolates discretionary fiscal policy from regime and rule-based components. Because cumulating growth-rate responses over a horizon involves a product rather than a sum, the proposed multiplier is multiplicative: both numerator and denominator are polynomials in the size of the fiscal impulse, so the multiplier varies nonlinearly with the scale of the expenditure programme – a dependence that conventional additive multipliers cannot capture. The multiplier requires no discounting, no ad hoc conversion factor, and no auxiliary estimate of an average expenditure share of output, and it is computed directly from nominal, non-seasonally-adjusted data, matching the terms in which government budgets are actually expressed. A version of the narrative approach to identification is introduced. It allows to decompose the influence of fiscal policy into a regime component, a standard policy rule embedded in the autoregressive structure, and discretionary shocks. Removing the regime component from the residuals yields smaller identified shocks, and correspondingly smaller multipliers, than found in SVARs not controlling for regime changes – closing much of the gap between SVAR- and DSGE-based estimates. It is found that multipliers vary substantially across countries and across types of government expenditure, that regime volatility itself differs markedly by country, and that several economies display cyclicality in spending and multiplier paths that only becomes visible once responses are cumulated multiplicatively. The resulting multipliers are nominal, derived from unadjusted data, proven to be sensitive to shock size, and purged of regime effects, offering a tool with direct relevance for budget planning and fiscal policy evaluation.

Large language models can generate runnable web applications from natural-language requirements, but many generated applications still fail interactive tasks. Existing generate-execute-repair pipelines execute the generated application and use task outcomes or error messages to guide code revision. However, this feedback often misses the runtime behavior between a browser action and the final task outcome, making interaction-level failures difficult to diagnose. Therefore, we propose TeleGen, an observability-enhanced framework for LLM-based web application generation. TeleGen instruments generated applications, collects runtime telemetry during task execution, and compresses raw telemetry logs into concise briefs for repair. We evaluate TeleGen on WebGen-Bench and Web-Bench. On WebGen-Bench, TeleGen improves task success from 67.7% with repair without telemetry to 76.2%, an increase of 8.5 percentage points. On Web-Bench, it improves cumulative Pass@2 from 21.7% to 29.8%. Ablation results show that runtime telemetry provides a useful diagnostic signal, while telemetry briefs make this signal more effective and less costly to use. Further analysis shows that telemetry is especially helpful for failures involving hidden execution paths, such as navigation, form workflows, and frontend-backend coordination.

Preference alignment has become a standard practice for text-to-image diffusion models. Direct Preference Optimization (DPO) simplifies this process by eliminating explicit reward modeling. Its diffusion variant, Diffusion-DPO, has become a widely adopted baseline. Diffusion-DPO essentially encourages the likelihood of preferred samples while suppressing dispreferred ones. In this paper, we revisit DPO-style alignment methods for diffusion models from the perspective of the manifold hypothesis. Under this view, natural images concentrate near a low-dimensional manifold embedded in the high-dimensional ambient space, whereas DPO directly optimizes preference distributions in the full space without accounting for this geometric structure. This creates a mismatch in the optimization dynamics: it suppresses geometry-preserving tangential updates, while insufficiently restricting hazardous normal-direction updates. This mismatch gradually degrades image quality and diversity. To address this issue, we propose Anisotropic Geometry-Aware Preference Optimization (APO), which replaces the uniform Euclidean treatment of prediction errors with a geometry-aware anisotropic metric derived from the reference model. Concretely, APO adaptively strengthens regularization in directions where the reference denoising function is highly sensitive, while relaxing constraints in directions that permit safe semantic adjustment. This recalibrates preference optimization according to the local manifold geometry, and maintains the original manifold structure. Experiments show that APO achieves strong performance and an average win rate exceeding 60% against various existing alignment methods across diverse benchmarks. It requires significantly fewer training steps than prior methods, and preserves generation diversity throughout training.

Robust tensor completion recovers multidimensional data from partial observations corrupted by sparse gross errors. Existing convolutional low-rank models typically construct translated copies using circular continuation, which introduces artificial wrap-around neighborhoods for finite nonperiodic data. We propose reflective convolution nuclear norm minimization (RCNNM), which replaces circular shifts with endpoint-nonrepeating reflection. The resulting lifting has nonuniform entry multiplicities and satisfies a weighted Gram identity that supports both the recovery analysis and the optimization method. Under random sampling and sparse corruption, we establish high-probability exact recovery of the underlying tensor and sparse errors, together with stability under bounded dense perturbations. We further develop a two-block ADMM with a closed-form entrywise tensor update, while singular-value thresholding is implemented through the smaller right Gram matrix. Experiments on synthetic tensors, BSDS color images, and CAVE multispectral images show that RCNNM consistently improves over its circular-lifting counterpart, with the clearest gains near image boundaries. In particular, average boundary-PSNR improvements reach 3.26 dB while global reconstruction quality remains competitive.

Data selection, curricula and the evaluation of post-training recipes all assume that we can tell, before training, which problems a model will improve on. We test this assumption directly. For two base models, DeepSeek-R1-Distill-Qwen-1.5B and Qwen2.5-Math-1.5B, we evaluate eighteen post-training runs on up to 1532 competition math problems with many samples per problem, and compare signals available before training against a noise ceiling derived from the agreement between disjoint subsets of runs. Three findings hold on both base models. What post-training fixes is shared: independent runs agree on which rarely solved problems improve, with a noise ceiling of about 0.9, yet the two base models agree with each other only at rho=0.25 - the shared component belongs to the base model, not to the problem. A priori signals capture a minority of it: base pass rate, the likelihood of a correct solution, a larger model's pass rate and their combinations explain only 0.30 and 0.24 of the explainable variance in gains. Existing checkpoints are the better predictor: the per-problem gains of a single checkpoint from another family predict a new run better than every a priori signal, alone or combined (0.52 vs. 0.33 and 0.34 vs. 0.17 against the combined signals). The conclusions hold on problems from 2025-2026 competitions and when the baselines on the two sides of every comparison are estimated independently. We propose the noise ceiling as a standard companion to per-problem signals.

Large language models (LLMs) are increasingly used as interactive agents and simulators, yet it remains unclear whether they can recover latent sequential structure beyond surface action frequencies. This distinction is critical for behavioral simulation, where actions are often shaped by prior context rather than marginal frequencies alone. We study this question using controlled two-player Rock–Paper–Scissors interactions and a one-player stochastic n-gram continuation task. Across these experiments, we test whether LLMs can identify latent strategies, follow simple Markov rules, and sustain higher-order conditional dependencies. Our framework separates distribution matching from conditional rule following. Results show that longer context does not improve identification, correct recognition does not ensure faithful simulation, and higher-order dependencies substantially degrade rule recovery. Apparent behavioral fidelity can therefore mask incorrect generative mechanisms.

We evaluate language models' capability to generate novel research ideas through the lens of compression. We aim to lower-bound the potentially tiny probability that a language model generates the essence of a future research idea without any hints. Rather than estimate this probability through expensive repeated sampling, our Priced Guidance framework measures the compression cost: how many additional bits of information are needed to guide the model to recover the target idea. We prove that if the model can recover the target idea with at most K bits of guidance in expectation, then it can generate the idea without any guidance with probability at least $2^{-K}$. In our framework, the language model, called the generator, can pose a sequence of multiple-choice questions and specify a probability distribution over possible answers. A guide, which is a language model with access to the target idea, selects answers. If a selected answer has prior probability p, the generator pays $-\log_2 p$ bits. The generator aims to produce an idea that matches the essence of the target idea with minimal cumulative cost. This cumulative cost equals, up to an additive constant, the number of bits of information sent by the guide. Using this methodology, we evaluate five generator LLMs (Opus 5, Fable 5.1, GPT-6 Astra, GPT-5.6 Sol, and GLM 5.3) on the core ideas in 87 recent high-quality deep learning papers and we use LLM as judge to determine whether the generated idea matches the target in terms of the central research object and defining mechanism. Fable 5.1 achieves the lowest median compression cost at 69.9 bits, substantially lower than gzip's median of 5,712 bits for losslessly compressing the summary of the target idea. A uniform ensemble of Fable 5.1, Opus 5, and Astra further reduces the median compression cost to 55.8 bits and improves the generation probability lower bound by 18,000 times.

Long-horizon tool-using AI agents create subgoals, replan, delegate work, and compose sibling results. Per-tool permission checks cannot establish that a changing goal graph remains within the principal-approved task. We address this authorization gap in a finite structured domain with one principal and one authorization root. Each proposed goal-graph mutation carries a version-bound witness that its continuation traces, resources, obligations, invariants, and closing condition refine the active root contract; every protected effect is rechecked at an atomic commit boundary. Free-form goal text supplies no authority. We prove trace-policy and modeled forbidden-state preservation under explicit mediation, abstraction, freshness, and atomicity assumptions, plus conditional root-success preservation, a separation result for memoryless allowlists, exact finite-domain decidability, and universal-safety monotonicity under sound abstraction refinement. An executable model explores 340 states and 419 transitions. Across 96 matched cases covering 25 structural schemas, the complete mechanism commits zero forbidden states in 48 drifted cases and completes all 48 benign counterparts. Two public upstream runtime paths execute 258 native dispatches across 32 cases, with every case-level decision and receipt chain matching. A frozen host-local study covers 129 synthetic one-factor-at-a-time cells, all matching fixed decisions and reasons. A history-aware continuation comparator blocks every modeled bad trace prefix but commits all operations in 11 cases whose violations lie in typed resources, freshness, or explicit-join evidence outside its trace projection. Within the registered structured domains, runtime authorization preserves useful replanning while preventing self-generated subgoals from becoming a source of new authority.

Large language models (LLMs) have shown strong performance across various tasks, but they still struggle with questions involving ethical judgment. Previous studies have attempted to train LLMs on ethical standards, but the diversity and relativity of ethical norms make them difficult to fully internalize in model parameters. As an alternative, we introduce normative RAG, a retrieval-augmented approach that supports ethical judgment using external normative knowledge. Normative retrieval involves a distinct asymmetry between context rich narrative queries and generalized normative statements. Existing factual retrieval methods rely on query-only expansion into a document-like form, making them insufficient for resolving this asymmetry. Therefore, we propose Intermediate Representation-based Embedding Alignment (IREA), a bidirectional alignment method that maps both text types into a shared situation-behavior representation. This representation captures ethically salient contextual and behavioral information in a normalized form, reducing surface-level discrepancies and improving alignment in the embedding space. Experimental results show that IREA improves normative retrieval and downstream ethical judgment across multiple settings, demonstrating the effectiveness of bidirectional alignment for normative RAG.