arXiv:1909.08210v4 Announce Type: replace
Abstract: A restricted Boltzmann machine (RBM) is a two-layer neural network with shared weights and has been extensively studied for dimensionality reduction, data representation and recommendation systems in the literature. The traditional RBM requires a probabilistic interpretation of the values on both layers and a Markov chain Monte Carlo (MCMC) procedure to generate samples during the training. The contrastive divergence (CD) is efficient to train the RBM but its convergence has not been proved mathematically. In this paper, using a maximum a posteriori (MAP) estimate and the expectation maximization (EM) algorithm, we show that the CD algorithm without MCMC is convergent for the conditional likelihood object function. Another key contribution in this paper is the reformulation of the RBM into a deterministic model. Within the reformulated RBM, the CD algorithm without MCMC approximates the gradient descent (GD) method. This reformulated RBM can take the continuous scalar and vector variables on the nodes with flexibility in choosing the activation functions. Numerical experiments show its capability in both linear and nonlinear dimensionality reduction, and, for the nonlinear dimensionality reduction, the reformulated RBM can outperform principal component analysis (PCA) by choosing the proper activation functions. Finally, we demonstrate its application to vector-valued nodes for the CIFAR-10 dataset (color images) and the multivariate sequence data, which cannot be configured naturally with the traditional RBM. This work not only provides theoretical insights regarding the traditional RBM but also unifies the linear and nonlinear dimensionality reduction for scalar and vector variables.
Science Journals
arXiv:2605.26911v1 Announce Type: new
Abstract: LLM-generated peer reviews are increasingly common at major venues, yet their deficiencies are hard to detect because they are uniformly fluent and well-structured. Existing work either classifies authorship without judging quality, or scores quality with features designed for human-written reviews; no prior system detects deficiencies in LLM-generated reviews at the level of individual defect types. To bridge the gap, we introduce TADDLE, a Tool-Augmented Agent for Detecting Deficient LLM-Generated Peer Reviews, together with the first expert-annotated benchmark for this task. Our benchmark comprises 1,800 reviews on 50 ICLR 2025 papers, multi-label-annotated by 18 domain experts against a taxonomy of six defect categories (plus a non-deficient label). TADDLE decomposes detection into four specialized analysis tools -- Verify, Correct, Complete, and Transform -- orchestrated by an agent; an integrator synthesizes their outputs into binary and multi-label classifications via two-stage semi-supervised learning. Extensive experiments show that TADDLE performs strongly on both binary detection and the multi-label classification task. We release the benchmark and code at https://github.com/AquariusAQ/TADDLE.
arXiv:2605.26860v1 Announce Type: new
Abstract: The continuous spontaneous localization (CSL) model is one of the most promising approaches to address the wave function collapse problem in the measurement process of standard quantum mechanics. In this work, the effect of the CSL model on a Bragg atom interferometer was investigated. A Bragg interferometer achieving high fringe contrast of 99$\%$ has been demonstrated, maintaining this performance level at interrogation time up to $T=60~\mathrm{ms}$. The primary factors responsible for fringe contrast loss in the atom interferometer were systematically analyzed and corrected. This improvement established a new upper limit of $\lambda_{\rm CSL}=1.27\times10^{-5}~\mathrm{s}^{-1}$ at $r_C=10^{-5}~\mathrm{m}$ for the CSL collapse rate, representing approximately 4 times enhancement over previous atom-interferometric constraints.
arXiv:2605.27358v1 Announce Type: new
Abstract: Mixture-of-Experts (MoE) has become the de facto architecture for hundred-billion-parameter language models, yet its advantages at sub-billion scales for on-device deployment remain largely unexplored. To close this gap, we present MobileMoE, a family of on-device MoE language models with sub-billion active parameters (0.3-0.9B active and 1.3-5.3B total) that establish a new Pareto frontier for on-device LLMs. We first formulate an on-device MoE scaling law that jointly optimizes MoE architecture under mobile memory and compute constraints, identifying an on-device sweet spot - moderate sparsity with fine-grained and shared experts - that is simultaneously memory and compute-optimal. Building on the derived architectures, we train MobileMoE with a four-stage recipe covering pre-training, mid-training, instruction fine-tuning, and quantization-aware training, all on open-source datasets. Across 14 benchmarks, MobileMoE matches or exceeds leading on-device dense LLMs with 2-4$\times$ fewer inference FLOPs, and matches or surpasses the state-of-the-art MoE OLMoE-1B-7B with up to 60% fewer parameters. To bridge the last mile to mobile deployment, we provide the first efficient MoE inference on commodity smartphones with comprehensive on-device profiling. At comparable INT4 weight memory, MobileMoE-S delivers $1.8$-$3.8\times$ faster prefill and $2.2$-$3.4\times$ faster decode than the dense baseline MobileLLM-Pro.
arXiv:2605.26196v1 Announce Type: new
Abstract: Trusted Execution Environments (TEEs) protect confidentiality and integrity of trusted applications by creating an isolated environment for executing code. Prior work has shown that users may feel more comfortable sharing data when they know it will be protected by a TEE, especially if they understand what a TEE is. In this study, we evaluated text-based explanations introducing TEEs to non-experts. We analyzed existing TEE explanations to develop candidate explanations and evaluated them via vignette scenarios with 966 crowdworkers. The explanations that enhanced understanding most were non-technical ones that highlighted specific threats that can be prevented by a TEE. Surprisingly, even the explanations that enhanced understanding had little effect on willingness to use the TEE-enhanced technology. These results provide insights into ways to communicate technical security concepts more effectively but also suggest that explaining security technology might not be enough to address users' privacy concerns.
arXiv:2605.26796v1 Announce Type: new
Abstract: Incentive-based load curtailment unlocks critical demand-side flexibility but is hindered by the limited knowledge of private user parameters and the inherent nonsmoothness of responses due to physical device constraints. We address this via a constrained bilevel optimization framework and propose the Bi-ZOL (Bilevel Zeroth-Order Learning) algorithm. Unlike conventional black-box methods, Bi-ZOL exploits the bilevel structure to decompose the hypergradient, integrating the exact analytical information of the SO's objective with a zeroth-order estimate of the unknown response sensitivity. This structural decomposition-based learning method mathematically smoothes the nonsmooth response landscape and reduces hypergradient estimation error. We provide theoretical convergence guarantees to an approximate stationary point and demonstrate through simulations that Bi-ZOL achieves near-optimal performance.
arXiv:2605.26415v1 Announce Type: new
Abstract: Deploying Vision-Language Models on resource-constrained hardware typically requires INT8 quantization, but in joint-embedding architectures such as CLIP this introduces a failure mode distinct from quantized CNN classifiers: activation noise accumulated across transformer blocks perturbs the direction of the multimodal embedding, eroding the cosine alignment on which zero-shot retrieval depends. We characterize this as Quantization-Induced Representation Collapse (QIRC) and quantify it on INT8 CLIP ViT-B/32, where the layer-wise noise-to-signal ratio grows from below 10% in shallow blocks to 52% at Layer 11. We propose LRA-EE (Layer-wise Representation-Aware Early Exit), which bypasses noise-saturated deep layers via Spatio-Semantic Aggregation (replacing the immature shallow [CLS] with a global patch-token average), a learned multi-feature gate (confidence, top-2 margin, spatial-activation variance), and Layer-adaptive Confidence Thresholding calibrated to each layer's Information-to-Noise Ratio. On ImageNet-1K zero-shot classification, LRA-EE reduces FLOPs by 13.4% and improves Top-1 accuracy by +2.44%p (58.72% -> 61.16%) over the INT8 baseline. A four-quadrant decomposition isolates the Rescue Effect: 9.5% of samples are correctly classified at shallow exits but lost to noise at full depth, against only 7.1% suffering the inverse.
arXiv:2605.26414v1 Announce Type: new
Abstract: Large Language Models (LLMs) achieve impressive accuracy on mathematical reasoning benchmarks, yet their performance drops when problems are modified with simple changes like different names or numbers. Code execution methods, which let models generate and run Python code instead of reasoning in natural language, have been proposed as a solution, but their effect on reasoning robustness (the ability to maintain accuracy across problem variations) has not been systematically tested. This study evaluates three approaches on 1,000 problems from the GSM-Symbolic dataset: pure reasoning using chain-of-thought (CoT) prompting, single-shot code execution using Program-Aided Language models (PAL), and iterative code execution using Step-by-Step Coding (SBSC). All three were run on paired original and modified problems using Claude Haiku 4.5. CoT was the most robust method, with an accuracy drop of 1.3 percentage points and 1.8% of problems breaking under perturbation. PAL was the least robust at 1.7 percentage points and 3.1% broke, with SBSC falling in between. Although these differences were not statistically significant ($p = .096$), the directional trend was consistent across all measures, suggesting that code execution, whether single-shot or iterative, does not improve reasoning robustness on grade-school-level problem variations.
arXiv:2510.24354v2 Announce Type: replace
Abstract: Despite extensive research, the mechanisms through which online platforms shape extremism and polarization remain poorly understood. We identify and test a mechanism, grounded in empirical evidence, that explains how ranking algorithms can amplify both phenomena. This mechanism is based on well-documented assumptions: (i) users exhibit position bias and tend to prefer items displayed higher in the ranking, (ii) users prefer like-minded content, (iii) users with more extreme views are more likely to engage actively, and (iv) ranking algorithms are popularity-based, assigning higher positions to items that attract more clicks. Under these conditions, when platforms additionally reward \emph{active} engagement and implement \emph{personalized} rankings, users are inevitably driven toward more extremist and polarized news consumption. We formalize this mechanism in a dynamical model, which we evaluate by means of simulations and interactive experiments with hundreds of human participants, where the rankings are updated dynamically in response to user activity.
arXiv:2605.26380v1 Announce Type: new
Abstract: Frontier multimodal large language models (MLLMs) have been reported to achieve over 90% accuracy on fine-grained perception benchmarks. However, such scores do not necessarily imply faithful use of visual evidence. Prior studies have identified three shortcuts that inflate benchmark performance. First, linguistic priors and lexical cues in questions often enable models to infer plausible answers without seeing the image. Second, coarse global semantics from the visual encoder can bypass fine-grained local details. Third, in some ``think-with-images'' benchmarks, corrupting the intermediate images returned by visual tools barely affects the final answer. These findings suggest that higher input resolution or larger question pools alone do not elicit genuine active visual search. To address this, we introduce VisualNeedle, a challenging, information-dense, and fine-grained benchmark for scenes where critical evidence is spatially constrained to minute regions and not discernible at a glance. We further propose a counterfactual crop-black setting, which replaces crops returned by tools with black images of the same size, to test whether tool-enabled performance truly relies on intermediate visual evidence. We evaluate 9 promninent MLLMs across three settings: no-tool, standard tool-enabled, and crop-black. No-tool accuracy stays below 20\%, and the best tool-enabled model reaches only 56.01\%, still trailing the 63.00% human majority-vote accuracy. These results reveal persistent limitations in fine-grained visual search, while the crop-black ablation confirms that success on VisualNeedle hinges on genuine intermediate visual evidence.
arXiv:2510.22623v3 Announce Type: replace
Abstract: Carbon nanotube (CNT) films are widely considered as prospective building blocks for advanced electronic and nanostructured materials. In particular, electrical transport in high-density CNT films results from a complex interplay between network morphology and CNT connectivity, which remains challenging to characterize quantitatively. To identify the structural parameters that govern the electrical current in CNT films, we employed coarse-grained molecular dynamics to construct dense mesoscale CNT film models that include CNTs with different chiralities and lengths. The effects of CNT geometrical features on the film morphologies were quantified by devising a set of structural descriptors and analyzing their mutual correlations. The impact of varying the concentration of amorphous carbon (aC) inclusions on the film structure was assessed. Finally, we employed a nodal analysis framework to compute the electrical current across the networks and correlate the charge transport characteristics to the underlying structural descriptors. The current is found to be enhanced in films that exhibit high curvature and buckling, low bundling, and strong connectivity. We discuss how the presence of aC inclusions modifies these morphological and current characteristics. This work provides a mesoscale modeling framework for modeling structure-transport relationships in dense CNT films and highlights the role of morphological descriptors in guiding the interpretation of electrical transport in complex nanostructured networks.
arXiv:2605.26350v1 Announce Type: new
Abstract: In-context learning (ICL) is often motivated by the intuition that demonstrations help because they provide correct input-output examples. However, we reveal a counterintuitive phenomenon: correctness does not guarantee exemplar utility, and some correct demonstrations can even reduce ICL accuracy. To study this correctness-utility gap, we introduce task-preserving perturbations, where only the exemplar input is changed, while the example remains a correct instance of the same task. Concretely, each perturbed exemplar is assigned the target induced by the task mapping. This framework covers both label-updating perturbations, where task-relevant semantics change and targets are recomputed, and stricter target-preserving perturbations, where the original target remains valid. We formalize the resulting failure mode as contextual evidence shift: task-preserving perturbations can change the effective mixture of evidence used by the model for contextual inference, thereby separating exemplar correctness from exemplar utility. Across sentiment classification, logical reasoning, and math word problems, we find that task-preserving perturbed demonstrations can substantially degrade ICL performance, especially for smaller models, harder tasks, and higher perturbation ratios. Our results show that robust ICL requires evaluating not only whether demonstrations are correct, but also how they influence contextual inference. Code is available at https://github.com/Chenghao-Qiu/Task-Preserving-ICL.
arXiv:2605.27261v1 Announce Type: new
Abstract: We present a pilot study on the collection and synchronisation of multimodal data for player experience investigation. We collected game telemetry, self-reported surveys, biometrics, and cued-retrospective think-aloud (C-RTA) data from 19 participants playing three Atari 2600 games. The study then uses the data to investigate difficulty in PX, showcasing a protocol for future multimodal research.
The dataset obtained from the experiment, which is publicly available, shows potential as a rich, transformative source that can be used to investigate dynamic difficulty adjustment algorithms, game balancing strategies or broader explorations of games user research. The study findings suggest that the experimental approach holds strong potential for generalisation in future player experience studies.
arXiv:2605.26273v1 Announce Type: new
Abstract: Semantic segmentation in complex environments such as urban driving scenes remains challenging under adverse lighting conditions, where RGB images alone provide insufficient information. RGB-Thermal fusion leverages the complementary strengths of visible and infrared imagery to improve scene understanding; however, effectively integrating these heterogeneous modalities at varying levels of feature abstraction remains an open problem. In this paper, we propose a multi-modal fusion architecture built upon dual ConvNeXt V2 backbones that employs stage-wise, modality-adaptive fusion strategies. For early-stage features, we introduce a Frequency-Based Fusion Module that decomposes infrared features into low- and high-frequency components via Gaussian filtering, applies dual-branch spatial attention to selectively emphasize thermal patterns and fine-grained boundaries, and integrates them with RGB features through a confidence-gated residual mechanism. For late-stage features, we design a semantic fusion module with cross-modal attention and multi-scale depthwise convolutions to capture semantic correspondences across modalities. The fused features are decoded via a PANet-style bidirectional decoder with deep supervision. Experiments on MFNet and PST900 demonstrate that our lightest variant achieves 61.73\% and 86.24\% mIoU, respectively, with only 35.43M parameters, outperforming recent methods while using substantially fewer parameters and lower computational cost. Code is available at https://github.com/ismailemrecntz/VISIBLE-INFRARED-SENSOR-FUSION
arXiv:2507.18412v4 Announce Type: replace-cross
Abstract: Toda lattice or FPUT chain-like dynamics have been regarded as the prerequisite condition to explain the length dependency of high thermal conductivity of low-dimensional systems at the nanoscale. In this paper, a hypothetical condition is introduced that establishes a theoretical connection between the thermal conductivity of a nanoscale low-dimensional system in nonequilibrium steady states(NESS) and the canonical motion of the equation in the Toda lattice in equilibrium. The hypothesis relies on a numerically driven coarse grained molecular dynamics(CGMD) system acquired from the trajectory data of nonequilibrium molecular dynamics(NEMD) simulation. It models the macroscopic motion from longitudinal and flexural modulation observed in NEMD as a separate Hamiltonian in CGMD with a perturbation term governed by an overdamping process, which is assumed to be dominant during heat transfer. The Smoluchowski equation for the perturbation, which is derived from the cross correlated states between two degrees of freedom, suggests that the potential energy function induced from NESS is identical to that of the Toda Lattice under the specific condition in the partition function for coarse grained particles. The restrictions derived from the model are well confirmed by the data from the numerically driven coarse grained model.
arXiv:2605.26619v1 Announce Type: new
Abstract: Reconstructing continuous state trajectories of chaotic dynamical systems from sparse, noisy observations remains a fundamental open problem in nonlinear science. We introduce the Physics-Informed Diffusion Model with Dormand-Prince Integration (PIDM-DP), which embeds a fully differentiable 5th-order Dormand-Prince (DP-RK45) ODE integrator directly into the reverse sampling loop of a Denoising Diffusion Probabilistic Model (DDPM). At each denoising step, physics residuals are back-propagated via automatic differentiation, constraining every generated trajectory to satisfy the system's governing equations to 5th-order accuracy. A linear-scheduled guidance mechanism that ramps the physics weight from zero at high noise levels to its full value near the clean-data limit prevents the gradient explosions that cause naive physics-informed approaches to fail on stiff systems with Jacobian eigenvalues of order $O(10^3)$. Evaluated across five benchmark systems of increasing complexity 3D Lorenz, 3D R\"ossler, 5D Hyperchaotic, 20D Lorenz-96, and the stiff 3D Rabinovich-Fabrikant at 10% observation density with additive Gaussian noise ($\sigma=0.05$), PIDM-DP achieves reconstruction RMSE improvements of up to $15.4\times$ over an unconstrained diffusion baseline and decisively outperforms the Ensemble Kalman Filter on stiff systems where ensemble covariance collapses. On the Rabinovich-Fabrikant out-of-distribution benchmark, PIDM-DP attains RMSE $0.1097 \pm 0.0269$ versus $0.9443 \pm 0.5288$ (unconstrained diffusion, $8.6\times$ worse) and $0.3561 \pm 0.3040$ (EnKF, $3.2\times$ worse), with $p<0.001$ in paired Wilcoxon tests ($N = 30$). Topological validation via the Rosenstein Lyapunov estimator confirms that PIDM-DP preserves the chaotic invariant measure.
arXiv:2605.26618v1 Announce Type: new
Abstract: The Abbe diffraction limit, tied to the fundamental spatial bandwidth constraint imposed by any physical aperture, remains the primary barrier to achieving ultimate far-field optical resolution and precise light-matter interactions. However, current efforts to engineer structured light fields beyond this limit often come at the cost of massive sacrifices in energy efficiency. In this work, we mathematically complete the family of non-zero azimuthal-order Circular Prolate Spheroidal Wave Functions (CPSWFs), introducing them as a complete class of band-limited superoscillatory optical vortices carrying helical phase. Compared with classical Laguerre-Gaussian (LG) beams, we rigorously prove that these eigenmodes achieve the theoretical upper bound for extreme energy concentration under strict band-limited constraints. At the scale of light-matter interactions, this optimal concentration directly amplifies the intensity gradients and angular momentum densities that govern optical forces. This advantage translates directly into a 29.9% reduction in the trapping power threshold and a 2.3-fold increase in the subdiffraction orbital rotation speed of nanoparticles. Looking forward, this fundamental physical framework not only establishes strict mathematical boundaries for structured light fields but also serves as an absolute theoretical benchmark for deep-learning inverse design, and next-generation extreme optical micro-manipulation systems.
arXiv:2605.27098v1 Announce Type: new
Abstract: We consider the problem of dividing a set of indivisible goods among agents with additive valuations. This problem has been studied under various objectives in both the computer science and the operations research literature. Our main contribution is a novel dictator test using this problem, which can separate a dictator from any function sufficiently far from a dictator. We use this test to prove the following hardness results (assuming the unique games conjecture is true):
(1) We show that it is NP-hard to approximate the max Nash welfare by a factor better than $\sqrt[3]{\frac{81}{65}} - \varepsilon \approx 1.0761$. This improves on the previous best known inapproximability factor of $\sqrt{\frac87} - \varepsilon \approx 1.069$.
(2) We show that it is NP-hard to approximate the maximum budgeted allocation by a factor better than $\frac{243}{227} - \varepsilon \approx 1.07$. This improves on the previous best known inapproximability factor of $\frac{16}{15} - \varepsilon \approx 1.067$.
(3) We show that it is NP-hard to approximate the max generalized assignment problem (GAP) by a factor better than $\frac{145}{129} - \varepsilon \approx 1.124$. This improves on the previous best known inapproximability factor of $\frac{11}{10} - \varepsilon \approx 1.10$.
arXiv:2605.26609v1 Announce Type: new
Abstract: Software stack upgrades are a routine part of software maintenance and evolution, typically motivated by improved performance, stability, or functionality. Yet their impact on energy consumption - a growing concern for organizations pursuing sustainability - remains poorly understood. This paper presents a systematic method for measuring how different versions of core software stack components, such as Spring Boot and the Java Virtual Machine (JVM), influence the energy consumption of applications. Our approach evaluates combinations of framework versions, runtime versions, and execution platforms through automated benchmarking. Using a case study based on the Spring Petclinic REST application, we show that energy consumption varies substantially across Spring Boot and JVM versions, in some cases producing unexpected regressions. Notably, newer JVM releases and virtual threads (introduced in Java 21 and 23) yielded significant energy savings without requiring application changes. These results demonstrate that software upgrades can meaningfully affect energy usage and that measuring energy consumption provides valuable evidence for decision making in software maintenance and evolution.
arXiv:2605.26365v1 Announce Type: new
Abstract: Large Language Models (LLMs) often exhibit homogenized cultural perspectives. While the World Values Survey (WVS) provides a gold standard for mapping human values, traditional direct prompting of LLMs on WVS often fails to access the model's latent cultural depth, leading to safety-aligned refusals or neutral responses. Here, we propose a generalizable framework for cultural evaluation and intervention that transitions from abstract queries to scenario-based behavioral probing. By extracting implicit token probabilities across 300 situational dilemmas, we bypass surface-level alignment to map the latent coordinates of LLMs cultural value. We further introduce activation steering to shift these internal alignments during the forward pass without retraining. Across multiple LLMs, we find substantial variation in adaptability and uncover a consistent phenomenon of latent entanglement, where interventions along one cultural dimension induce shifts along another. These results suggest that cultural values are encoded as coupled structures, limiting precise alignment. This work establishes a computationally efficient framework for cultural steering, highlighting the structural complexities when navigating global value with LLMs.
arXiv:2605.26975v1 Announce Type: new
Abstract: Nonlinear reformulations of the spectral clustering method have gained a lot of recent attention due to their increased numerical benefits and their solid mathematical background. However, the estimation of the multiple nonlinear eigenvectors is associated with an increased computational cost. We present an implementation of a direct multiway spectral clustering algorithm in the $p$-norm, for $p\in(1,2]$, using a novel C++ GraphBLAS API. The key operations are expressed in linear algebraic terms and are executed over the resulting sparse matrices and dense vectors, parameterized in the algebra pertinent to the computation. We demonstrate the effectiveness and accuracy of our shared-memory algorithm on several artificial test cases. Our numerical examples and comparative results against competitive methods indicate that the proposed implementation attains high quality clusters in terms of the balanced graph cut metric. The strong scaling capabilities of our algorithm are showcased on a range of datasets with up to $8$ million nodes and $48$ million edges.
arXiv:2605.26384v1 Announce Type: new
Abstract: At global scale, data-center electricity demand is growing faster than the grids that supply it, while system operators increasingly require large flexible loads that can adjust power within seconds to absorb variable wind and solar generation. For multi-megawatt AI/HPC facilities, the key unresolved question is practical and measurable: how quickly can the software stack translate a grid request into a real change in GPU power at the facility meter, where commitments are settled? We answer this on real hardware with GridPilot, a three-tier predictive controller operating across milliseconds, seconds, and hours, augmented by a deterministic safety-island bypass for fast response. On a three-GPU NVIDIA V100 testbed, GridPilot achieves a measured end-to-end trigger-to-target response of 97.2 ms, which is 6.9x faster than the 700 ms requirement of Nordic Fast Frequency Reserve. We further incorporate an instantaneous Power Usage Effectiveness (PUE) correction so dispatched commitments remain robust at meter level rather than only at IT load level. In replay experiments across six representative European grids (from Sweden to Poland), the PUE-aware controller closes 2.5-5.8 percentage points of cooling-overhead drag. GridPilot is released as open source and serves as a proof of concept that MW-scale AI/HPC demand can be engineered as controllable, grid-responsive flexibility by design.
arXiv:2605.26385v1 Announce Type: new
Abstract: Large-scale search, recommendation, and retrieval-augmented generation (RAG) systems typically employ a two-stage architecture: an early-stage ranker (ESR) generates a candidate set, which is subsequently re-ranked by a late-stage ranker (LSR). While there are many reinforcement learning (RL) methods for training the LSR, end-to-end training of the ESR has proven challenging. In particular, naive application of "vanilla" policy gradient (V-PG) is not scalable for candidate-set sizes relevant for practical use due to exploding variance. This issue arises because V-PG propagates the gradient to the joint probability of the candidate sets, ignoring the contribution of each specific item in the candidate set to the reward. To mitigate this issue, we propose a novel "credit-assigned" policy gradient (CA-PG), which computes gradients with respect to the probability that the target item is chosen in any candidate set, i.e. marginalizing over all candidate sets that contain it. Our theoretical analysis reveals that CA-PG significantly reduces the variance of V-PG by marginalizing over the specific composition of the candidate set, while preserving the ability to learn the correct ranking of items under a reasonably aligned LSR policy. Experiments on both synthetic and real-world data demonstrate that CA-PG improves the convergence speed and training stability for ESRs utilizing the canonical Plackett-Luce model, especially when the candidate-set size is large.
arXiv:2605.02958v2 Announce Type: replace
Abstract: Representation Engineering analyses often characterize refusal using static directions extracted from terminal or pooled representations. We ask whether this view misses how refusal is constructed across layer-token positions. Using causal tracing, we identify a \textit{Refusal Trajectory}: a sparse upstream activation pattern that often persists even when attacks such as GCG suppress terminal refusal signals. Based on this observation, we propose SALO (Sparse Activation Localization Operator), a lightweight white-box detector that operates on raw hidden-state volumes from a selected layer window. Across Qwen, Llama, and Mistral models, SALO improves jailbreak detection on several attack families under a fixed XSTest-calibrated operating point. We further analyze static RepE-style baselines, ROI sensitivity, adaptive GCG attacks, and encoded-input boundary cases, clarifying both the promise and limitations of refusal-trajectory monitoring.
arXiv:2605.26966v1 Announce Type: new
Abstract: Understanding student difficulties in programming is a complex challenge due to the wide range of topics and the abundant varieties of misconceptions and errors. This paper presents the design and development of a fine-grained taxonomy that categorizes novice programmers' difficulties specifically related to reading and understanding the control flow constructs selection and iteration. Building upon prior research and our own empirical data from quizzes and interviews with students, the taxonomy is constructed through the iterative methodology of the Extended Taxonomy Design Process (ETDP). Key contributions include clear distinctions between different student difficulties and a detailed analysis of common student misunderstandings concerning conditional statements and loops. The taxonomy aims to aid computing education researchers by providing a harmonized framework to classify and analyze student errors, fostering deeper theoretical insights and informing pedagogical strategies. Future work will involve applying the taxonomy to novel student data and evaluating its usability among educators and researchers.