arXiv:2603.21652v2 Announce Type: replace
Abstract: Phishing attacks remain a persistent cybersecurity threat, and the widespread adoption of TLS certificates has unintentionally enabled malicious websites to appear trustworthy to users. This study examines whether certificate metadata and domain characteristics can help distinguish phishing domains from benign domains within the Danish .dk namespace. A dataset was constructed by combining registry information from Punktum dk with phishing reports and popularity rankings from external sources. TLS certificate attributes were collected using Netlas, while additional domain-based features were derived from DNS records and lexical analysis of domain names. The analysis compares phishing, popular, and less frequently visited domains across several feature categories, including Certificate Authorities (CAs), validity periods, missing certificate fields, SAN structure, registrant geography, hosting providers, and lexical properties of domain names. The results indicate that several features show observable differences between phishing and highly popular domains. However, phishing domains often resemble less popular domains, resulting in substantial overlap across many characteristics. Consequently, no individual feature provides a reliable standalone indicator of phishing activity within the Danish namespace. The findings suggest that certificate and domain attributes may still contribute to detection when combined, while also highlighting the limitations of relying on individual indicators in isolation. This work provides an empirical overview of phishing-related infrastructure patterns in the Danish .dk ecosystem and offers insights that may inform future phishing detection approaches.
Science Journals
arXiv:2604.16429v3 Announce Type: replace
Abstract: We introduce Mosaic, a probabilistic weather forecasting model that addresses three failure modes of spectral degradation in ML-based weather prediction: spectral damping (statistical), high-frequency aliasing (architectural), and residual high-frequency leakage (parametric). Mosaic generates ensemble members through learned functional perturbations and operates on native-resolution grids via mesh-aligned block-sparse attention, a hardware-aligned mechanism that captures long-range dependencies at linear cost by sharing keys and values across spatially adjacent queries. At 1.5{\deg} resolution with 214M parameters, Mosaic matches or outperforms models trained on 6$\times$ finer resolution on key variables and achieves state-of-the-art results among 1.5{\deg} models, producing well-calibrated ensembles whose individual members exhibit near-perfect spectral alignment across all resolved frequencies. A 24-member, 10-day forecast takes under 12s on a single H100~GPU. Code is available at https://github.com/maxxxzdn/mosaic.
arXiv:2603.21071v2 Announce Type: replace
Abstract: As one of the most important underwater sensing technologies, forward-looking sonar exhibits unique imaging characteristics. Sonar images are often affected by severe speckle noise, low texture contrast, acoustic shadows, and geometric distortions. These factors make it difficult for traditional teacher-student frameworks to achieve satisfactory performance in sonar semantic segmentation tasks under extremely limited labeled data conditions. To address this issue, we propose a Collaborative Teacher Semantic Segmentation Framework for forward-looking sonar images. This framework introduces a multi-teacher collaborative mechanism composed of one general teacher and multiple sonar-specific teachers. By adopting a multi-teacher alternating guidance strategy, the student model can learn general semantic representations while simultaneously capturing the unique characteristics of sonar images, thereby achieving more comprehensive and robust feature modeling. Considering the challenges of sonar images, which can lead teachers to generate a large number of noisy pseudo-labels, we further design a cross-teacher reliability assessment mechanism. This mechanism dynamically quantifies the reliability of pseudo-labels by evaluating the consistency and stability of predictions across multiple views and multiple teachers, thereby mitigating the negative impact caused by noisy pseudo-labels. Notably, on the FLSMD dataset, when only 2% of the data is labeled, our method achieves a 5.08% improvement in mIoU compared to other state-of-the-art approaches.
arXiv:2603.20216v2 Announce Type: replace
Abstract: Diffusion language models (DLMs) have emerged as a promising alternative to autoregressive (AR) models, offering sub-linear generation latency and bidirectional capabilities that are particularly appealing for code generation and editing. Achieving sub-linear latency in discrete DLMs requires predicting multiple tokens in parallel. However, standard DLMs sample tokens independently from conditional marginal distributions, failing to capture the joint dependencies among concurrently generated tokens. As a result, they often lead to syntactic inconsistencies and break multi-token structures. In this work, we introduce CoDiLA (Coherent Diffusion with Local Autoregression), a method that reconciles parallel sampling with local dependency modeling. Rather than forcing the DLM to resolve fine-grained syntax, CoDiLA delegates local decoding to a small, auxiliary AR model operating on the diffusion latents. This design allows for parallel generation while ensuring sequential validity within a block and maintaining core DLM capabilities, including bidirectional modeling across blocks. We demonstrate that using a highly compact auxiliary AR model (e.g., 0.6B parameters) effectively eliminates coherence artifacts, establishing a new Pareto frontier for accuracy and speed in code generation benchmarks.
arXiv:2603.19538v2 Announce Type: replace
Abstract: Monocular 3D object understanding has largely been cast as a 2D RoI-to-3D box lifting problem. However, emerging downstream applications require image-plane geometry (e.g., projected 3D box corners) which cannot be easily obtained without known intrinsics, a problem for object detection in the wild. We introduce MoCA3D, a Monocular, Class-Agnostic 3D model that predicts projected 3D bounding box corners and per-corner depths without requiring camera intrinsics at inference time. MoCA3D formulates pixel-space localization and depth assignment as dense prediction via corner heatmaps and depth maps. To evaluate image-plane geometric fidelity, we propose Pixel-Aligned Geometry (PAG), which directly measures image-plane corner and depth consistency. Extensive experiments demonstrate that MoCA3D achieves state-of-the-art performance, improving image-plane corner PAG by 22.8% while remaining comparable on 3D IoU, using up to 57 times fewer trainable parameters. Finally, we apply MoCA3D to downstream tasks which were previously impractical under unknown intrinsics, highlighting its utility beyond standard baseline models.
arXiv:2603.19470v3 Announce Type: replace
Abstract: Off-policy problems such as policy staleness and training--inference mismatch have become a major bottleneck for training stability and further exploration in LLM RL. The distribution gap between the inference and updated policies grows because of the techniques to enhance inference efficiency, leading to heavy-tailed importance ratios. Heavy-tailed ratios arise when the policy is locally sharp, which further inflates gradients and can push updates outside the trust region. To address this, we propose Adaptive Layerwise Perturbation (ALP), which injects small learnable perturbations into the input hidden states of each layer during updates and uses the resulting perturbed policy as the numerator of the importance ratio against the unchanged inference policy in the objective. Intuitively, by adding controlled noise to intermediate representations, ALP prevents the updated policy from deviating too sharply from the inference policy and enlarges the policy family to cover inference-time mismatch noise. Hence, the flattened distribution can naturally tighten the gap between the updated and inference policies and reduce the tail of importance ratios, thus maintaining training stability. This is further validated empirically. Experiments on single-turn math and multi-turn tool-integrated reasoning tasks show that ALP not only improves final performance, but also avoids blow-up in the importance-ratio tail and KL spikes during iterative training, along with boosted exploration. Ablations show that representation-level perturbations across all layers are most effective, substantially outperforming partial-layer and logits-only variants.
arXiv:2603.18702v4 Announce Type: replace
Abstract: We study off-policy learning (OPL) in contextual bandits, which plays a key role in a wide range of real-world applications such as recommendation systems and online advertising. Typical OPL in contextual bandits assumes an unconstrained environment where a policy can select the same item infinitely. However, in many practical applications, including coupon allocation and e-commerce, limited supply constrains items through budget limits on distributed coupons or inventory restrictions on products. In these settings, greedily selecting the item with the highest expected reward for the current user may lead to early depletion of that item, making it unavailable for future users who could potentially generate higher expected rewards. As a result, OPL methods that are optimal in unconstrained settings may become suboptimal in limited supply settings. To address the issue, we provide a theoretical analysis showing that conventional greedy OPL approaches may fail to maximize the policy performance, and demonstrate that policies with superior performance must exist in limited supply settings. Based on this insight, we introduce a novel method called Off-Policy learning with Limited Supply (OPLS). Rather than simply selecting the item with the highest expected reward, OPLS focuses on items with relatively higher expected rewards compared to the other users, enabling more efficient allocation of items with limited supply. Our empirical results on both synthetic and real-world datasets show that OPLS outperforms existing OPL methods in contextual bandit problems with limited supply.
arXiv:2603.18178v2 Announce Type: replace
Abstract: The rapid growth of ego-centric dashcam footage presents a major challenge for detecting safety-critical events such as collisions and near-collisions, scenarios that are brief, rare, and difficult for generic vision models to capture. While multimodal large language models (MLLMs) demonstrate strong general reasoning ability, they underperform in driving contexts due to domain and temporal misalignment.
We introduce VLM-AutoDrive, a modular post-training framework for adapting pretrained Vision-Language Models (VLMs) to high-fidelity anomaly detection. The framework integrates metadata-derived captions, LLM-generated descriptions, visual question answering (VQA) pairs, and chain-of-thought (CoT) reasoning supervision to enable domain-aligned and interpretable learning. Off-the-shelf VLMs such as NVIDIA's Cosmos-Reason1 7B (CR1) exhibit near-zero Collision recall in zero-shot settings; fine-tuning with VLM-AutoDrive improves Collision F1 from 0.00 to 0.69 and overall accuracy from 35.35% to 77.27%.
VLM-AutoDrive offers a scalable recipe for adapting general-purpose VLMs to safety-critical, temporally localized perception tasks. Evaluated on real-world Nexar dashcam videos, it achieves substantial gains in Collision and Near-Collision detection while producing interpretable reasoning traces, bridging the gap between perception, causality, and decision reasoning in autonomous driving.
arXiv:2603.18053v2 Announce Type: replace
Abstract: Online social platforms increasingly rely on crowd-sourced systems to label misleading content at scale, but these systems must both aggregate users' evaluations and decide whose evaluations to trust. To address the latter, many platforms audit users by rewarding agreement with the final aggregate outcome, a design we term consensus-based auditing. We analyze the consequences of this design in X's Community Notes, which in September 2022 adopted consensus-based auditing that ties users' eligibility for participation to agreement with the eventual platform outcome. We find evidence of strategic conformity: minority contributors' evaluations drift toward the majority and their participation share falls on controversial topics, where independent signals matter most. We formalize this mechanism in a behavioral model in which contributors trade off private beliefs against anticipated penalties for disagreement. Motivated by these findings, we propose a two-stage auditing and aggregation algorithm that weights contributors by the stability of their past residuals rather than by agreement with the majority. The method first accounts for differences across content and contributors, and then measures how predictable each contributor's evaluations are relative to the latent-factor model. Contributors whose evaluations are consistently informative receive greater influence in aggregation, even when they disagree with the prevailing consensus. In the Community Notes data, this approach improves out-of-sample predictive performance while avoiding penalization of disagreement.
arXiv:2603.17751v4 Announce Type: replace
Abstract: In the emerging mixed traffic environments, Connected and Autonomous Vehicles (CAVs) have to interact with surrounding human-driven vehicles (HDVs). This paper introduces MSH-MCCT (Multi-Source Human-in-the-Loop Mixed Cloud Control Testbed), a novel CAV testbed that captures complex interactions between various CAVs and HDVs. Utilizing the Mixed Digital Twin concept, which combines Mixed Reality with Digital Twin, MSH-MCCT integrates physical, virtual, and mixed platforms, along with multi-source control inputs. Bridged by the mixed platform, MSH-MCCT allows human drivers and CAV algorithms to operate both physical and virtual vehicles within multiple fields of view. Particularly, this testbed facilitates the coexistence and real-time interaction of physical and virtual CAVs \& HDVs, significantly enhancing the experimental flexibility and scalability. Experiments on vehicle platooning in mixed traffic showcase the potential of MSH-MCCT to conduct CAV testing with multi-source real human drivers in the loop through driving simulators of diverse fidelity. The videos for the experiments are available at our project website: https://dongjh20.github.io/MSH-MCCT.
arXiv:2504.16481v4 Announce Type: replace
Abstract: We study discounted random walks in directed graphs. In each step, the walk either terminates with a constant probability $\alpha$, or proceeds to a random out-neighbor. Our goal is to estimate the probability $\pi(s, t)$ that a discounted random walk starting from $s$ terminates at $t$. This probability is also known as the Personalized PageRank (PPR) score, which measures the relevance of $t$ to $s$, for instance, when $s$ and $t$ are web pages on the Internet. We aim to estimate $\pi(s, t)$ within a constant relative error with constant probability.
A variety of algorithms have been developed for several problem variants, such as single-pair, single-source, single-target, and single-node estimation, under both worst-case and average-case settings, and for different combinations of allowed graph queries. However, in many important cases, there remain polynomial gaps between known upper and lower bounds.
In this paper, we establish tight upper and lower bounds (up to logarithmic factors of $n$) for all problem variants and query combinations, closing all existing gaps in both the worst-case and average-case settings. Below we give some examples for the worst-case settings.
As an upper-bound example, the classic power method estimates $\pi(s,t)$ if it is above a threshold $\delta$ in time $O(m\log(1/\delta))$ but $\pi(s,t)$ can be as small as $1/n^{\Theta(n)}$. For contrast, we propose algorithms that deterministically estimate arbitrarily small $\pi(s,t)$ in $O(m\log n)$ time.
As a lower-bound example, we improve the lower bound for the single-pair problem from $\Omega(\min\{n,1/\delta\})$ to $\Omega(\min\{m,1/\delta\})$, which is optimal (up to logarithmic factors) since a simple Monte Carlo estimate takes $O(1/\delta)$ time.
arXiv:2605.18305v1 Announce Type: new
Abstract: It is a widespread belief among musicians that a violin's sound ``opens up'' or improves through regular playing. However, physical evidence for this ``playing-in'' effect remains elusive. This study revisited the phenomenon by testing two hypotheses: (1) the instrument undergoes physical evolution, or (2) the player undergoes behavioral adaptation. We conducted a longitudinal study centered on a seldom-played test violin played daily by a professional soloist for six months, alongside two stored control violins and a control group of ten violinists (N=10). All three violins had been rarely played prior to the experiment. Assessments performed at the beginning (``Before'' phase) and end (``After'' phase) of the period included input admittance measurements via laser vibrometry, standardized sound recordings, and acquisition of the soloist's bowing gestures via motion capture. Results revealed no acoustical evolution in the played violin exceeding the environmental drift observed in the control instruments. Furthermore, kinematic analysis showed no significant drift in the soloist's bowing strategy (bow position, velocity, contact point). These findings suggest that the ``playing-in'' phenomenon is driven neither by macroscopic physical changes in the instrument nor by a fundamental reorganization of the player's technique.
arXiv:2605.18302v1 Announce Type: new
Abstract: Synthetic participants represent a methodologically concerning concept that threatens the integrity of UX research. Findings from previous experiments specify how AI outputs are misaligned with the behaviors and thoughts of real humans in various ways. However, industry voices keep underestimating their severity, advocating for practical compromises where good-enough data does not need to be perfect, and all issues will be solved by future tuning. Our study tackles the lack of systematic understanding of the practical issues that arise with synthetic behavior and its use for steering decisions within real contexts. Within twelve diverse first click tests (n = 3431) obtained from real UX practice, we examine the ability of GPT to predict where humans click and how they reason about their behavior. Results (e.g., significantly different distribution from real data in 53% of tasks) demonstrate critical failures to reflect the patterns in which users click on visual elements and the underlying cognitive processes. Participant personas, chain-of-thought reasoning in GPT, and different sampling parameters fail to create sensible fidelity improvements apart from inflating believability. We expose a multitude of nuanced distortions in synthetic responses that reduce their overall analytical usefulness as a decision-making resource, compared with real data. Observed distortions can be theoretically linked to the properties categorically inherent to LLMs: their statistical nature and encoding of semantic heuristics dependent on their training on linguistic data.
arXiv:2605.18299v1 Announce Type: new
Abstract: Search-augmented reasoning agents interleave internal reasoning with calls to an external retriever, and their performance relies on the quality of each issued query. However, under outcome-reward reinforcement learning, every search decision in a rollout shares the same trajectory-level reward, leaving individual queries without step-specific credit. Recent process-supervision approaches address this gap by drawing step-level signals from outside the policy, relying either on a much larger teacher model, or on sub-question annotations produced by a stronger external system. In contrast, we propose SD-Search, which derives step-level supervision from the policy itself through on-policy hindsight self-distillation, requiring neither an external teacher nor additional annotations. In SD-Search, a single model plays two roles that differ only in conditioning: a student that sees only the context available at inference time, and a teacher that additionally conditions on a compact hindsight block summarizing the search queries and final outcomes of a group of rollouts sampled from the same question. Since the teacher knows how each rollout unfolded and which ones succeeded, its query distribution implicitly marks which decisions were worth making, and the student is trained to recover this behavior by minimizing the token-level Jensen--Shannon divergence to the teacher at search-query positions. This layers a dense, step-level signal on top of GRPO's coarse trajectory reward. Crucially, this signal is produced by the policy itself within the standard RL training loop, without external model inference, auxiliary annotation pipeline, or additional training stage.
arXiv:2605.18298v1 Announce Type: new
Abstract: Foundation models pre-trained through masked reconstruction on large-scale EEG data have emerged as a promising paradigm for learning generalizable neural representations across diverse brain-computer interface applications. However, a critical yet overlooked challenge is that EEG encoders must learn representations invariant to incomplete observations-when different masked views of the same signal have minimal overlap, existing methods fail to constrain them to a consistent latent subspace, leading to degraded transferability. To address this, we propose DARE-EEG, a self-supervised foundation model that explicitly enforces the mask-invariance property through dual-aligned representation learning during pre-training. Specifically, we introduce mask alignment that constrains representations from multiple masked views of the same EEG sample via contrastive learning, complementing anchor alignment that aligns masked representations to momentum-updated complete features for semantic stability. Additionally, we propose conv-linear-probing, a parameter-efficient strategy that adapts pre-trained representations to heterogeneous electrode configurations and sampling rates through decoupled spectro-spatial projections. Extensive experiments across diverse EEG benchmarks demonstrate that DARE-EEG consistently achieves state-of-the-art in accuracy performance while maintaining relatively low parameter complexity and superior cross-dataset portability compared to existing methods. Furthermore, DARE-EEG contributes to effectively discovering and utilizing the rich potential representations in EEG.
arXiv:2605.18059v1 Announce Type: new
Abstract: Robustness is a critical requirement for deploying autonomous driving systems in the real world. Existing robustness benchmarks for autonomous driving have made important progress in studying the effects of image-level corruptions, such as adverse weather or camera degradation, on perception modules and open-loop planning outputs. However, deployment can also involve system-level imperfections, such as inference latency and ego-state estimation errors, which remain less studied in closed-loop E2E-AD evaluation. These imperfections can accumulate through the feedback loop and destabilize control. In this work, we present Bench2Drive-Robust, to our knowledge the first device-centric robustness benchmark for closed-loop end-to-end autonomous driving under realistic deployment perturbations. We systematically evaluate deployment-oriented perturbations arising from three major sources: camera-stream failures (frame drop, partial observation), ego-state estimation errors (GPS noise, and speed or odometry errors), and compute-induced control delay (model inference delay). We evaluate representative end-to-end driving methods and analyze their robustness under different perturbation severities. Our results show that these deployment-related perturbations can substantially degrade closed-loop driving performance, revealing robustness challenges that are not fully captured by conventional image-level corruption evaluations. By establishing a closed-loop evaluation protocol and demonstrating the substantial impact of these deployment-oriented perturbations, Bench2Drive-Robust defines practical robustness problems for end-to-end autonomous driving and encourages further research on deployment-aware robust driving systems.
arXiv:2507.05688v2 Announce Type: replace-cross
Abstract: Diffusion models have shown strong performance in speech enhancement, but their real-time applicability has been limited by multi-step iterative sampling. Consistency distillation has recently emerged as a promising alternative by distilling a one-step consistency model from a multi-step diffusion-based teacher model. However, distilled consistency models are inherently biased towards the sampling trajectory of the teacher model, making them less robust to noise and prone to inheriting inaccuracies from the teacher model. To address this limitation, we propose ROSE-CD: Robust One-step Speech Enhancement via Consistency Distillation, a novel approach for distilling a one-step consistency model. Specifically, we introduce a randomized learning trajectory to improve the model's robustness to noise. Furthermore, we jointly optimize the one-step model with two time-domain auxiliary losses, enabling it to recover from teacher-induced errors and surpass the teacher model in overall performance. This is the first pure one-step consistency distillation model for diffusion-based speech enhancement, achieving 54 times faster inference speed and superior performance compared to its 30-step teacher model. Experiments on the VoiceBank-DEMAND dataset demonstrate that the proposed model achieves state-of-the-art performance in terms of speech quality. Moreover, its generalization ability is validated on both an out-of-domain dataset and real-world noisy recordings.
arXiv:2603.14942v3 Announce Type: replace
Abstract: The Hawkes process models self-exciting event streams, requiring a strictly non-negative and stable stochastic intensity. Standard identification methods enforce these properties using non-negative causal bases, yielding conservative parameter constraints and severely ill-conditioned least-squares Gram matrices at higher model orders. To overcome this, we introduce a system-theoretic identification framework utilizing the sign-indefinite orthonormal Laguerre basis, which guarantees a well-conditioned asymptotic Gram matrix independent of model order. We formulate a constrained least-squares problem enforcing the necessary and sufficient conditions for positivity and stability. By constructing the empirical Gram matrix via a Lyapunov equation and representing the constraints through a sum-of-squares trace equivalence, the proposed estimator is efficiently computed via semidefinite programming.
arXiv:2605.18285v1 Announce Type: new
Abstract: A key requirement on any well-behaved process language is its compositionality: behavioural equivalence of processes should be respected by the constructors of the language. Turi and Plotkin's abstract GSOS provides an elegant bialgebraic framework for modelling rule formats that guarantee compositionality from the outset. Their original results, however, are restricted to compositionality of strong bisimilarity, a rather fine-grained notion of process equivalence. In the present paper, we demonstrate that Turi and Plotkin's approach also applies to trace equivalence, which only observes external actions of processes. To this end, we revisit the general compositionality result of their original theory and present it in a refined form with regard to the required naturality conditions. This step makes abstract GSOS applicable over Kleisli categories and thereby enables reasoning about compositionality in the setting of coalgebraic trace semantics. As our main contribution, we introduce De Simone laws, a type of GSOS laws over Kleisli categories, and prove that their operational models are compositional for coalgebraic trace equivalence. This result recovers and explains compositionality of the well-known De Simone rule format for labelled transition systems in a natural categorical setting. As a further application, we derive from our general framework a novel De Simone-type format for probabilistic systems, compositional for probabilistic trace equivalence.
arXiv:2605.17999v1 Announce Type: new
Abstract: This paper proposes a Shared Backbone Proximal Policy Optimization (Shared Backbone PPO) algorithm. By sharing the base module between the Actor and Critic networks, the algorithm achieves efficient training and improved performance. The algorithm is implemented in a connectivity-preserving multi-UAV swarm communication coverage task and compared with the standard PPO algorithm. Experimental results demonstrate that the proposed method achieves superior performance. Furthermore, a graph information aggregation module is incorporated into the model architecture to accommodate the communication conditions among agents. With the integration of this module, the algorithm remains effective, and the trained agent swarm exhibits a higher level of cooperation.
arXiv:2603.20562v3 Announce Type: replace
Abstract: Large language models (LLMs) are now widely used as judges, yet their decisions can change under presentation choices that should be irrelevant. We study one such source of instability: candidate-order sensitivity in listwise factuality evaluation, where several answers can look similarly polished while differing substantially in hallucination risk. We introduce PCFJudge, an inference-time method that reruns the same factuality-first listwise prompt over multiple orderings of the same candidate set and aggregates the resulting scores, ranks, and uncertainty signals into a single consensus decision. On RewardBench 2 Factuality, the final seven-permutation aggregate (K=7) improves top-1 selection accuracy from 86.00% to 91.33% with GPT-5.4 and from 86.33% to 89.67% with Claude Sonnet 4.6. These results suggest that candidate order can be a meaningful source of factuality-judging error and that marginalizing over this nuisance variation can improve the reliability of LLM evaluation.
arXiv:2605.17745v1 Announce Type: cross
Abstract: Quantization is essential for reducing the computational cost and memory usage of deep neural networks, enabling efficient inference on low-precision hardware. Despite the growing adoption of uniform and floating-point quantization schemes, selecting optimal quantization parameters remains a key challenge, particularly for diverse data distributions encountered during training and inference. This work presents a novel statistical error analysis framework for uniform and floating-point quantization, providing theoretical insight into error behavior across quantization configurations. Building on this analysis, we propose iterative quantizers designed for arbitrary data distributions and analytic quantizers tailored for Gaussian-like weight distributions. These methods enable efficient, low-error quantization suitable for both activations and weights. We incorporate our quantizers into quantization-aware training and evaluate them across integer and floating-point formats. Experiments demonstrate improved accuracy and stability, highlighting the effectiveness of our approach for training low-precision neural networks.
arXiv:2605.18281v1 Announce Type: new
Abstract: Modern deep learning science often assumes that neural networks learn from a fixed data distribution. However, many practically important learning problems involve data distributions that change throughout training. How does such non-stationarity impact the inductive biases of deep learning towards models with different structural, generalisation, and safety properties? A fruitful testbed for studying inductive bias is in-context linear regression sequence modelling, where small transformers display strikingly different generalisation patterns depending on the diversity of the (fixed) training task distribution. In this paper, we explore the effect of diversifying the task distribution across training time, finding that such temporal diversity leads to an increased bias towards generalisation over memorisation.
arXiv:2605.18284v1 Announce Type: new
Abstract: Software repositories accumulate large amounts of unstructured knowledge in commit messages, pull-request discussions, and issue threads, but developers and AI coding assistants rarely reuse this history effectively. Recent work on typed-memory architectures for LLM agents (MemGPT, generative agents, and the PlugMem module of Yang et al.) argues that agent memory should be distilled, typed knowledge rather than raw interaction text. We adapt that stance to a software repository's own git history under a constrained regime: deterministic, dependency-free, local-only, no embeddings. We present CommitDistill, an open-source Python prototype that mines a local git history into typed knowledge units (Facts, Skills, Patterns) using deterministic regex and surfaces them through a TF-IDF retriever with a calibrated silence threshold (theta = 2.5) that abstains on out-of-distribution queries. The artefact is a trust-instrumented memory substrate: deterministic, no external service, inspectable plain-JSON store, tunable abstention. A case study on five public repositories spanning Python, JavaScript, C, and Java (25,000 commits, 1,167 extracted units) reports useful-precision 0.525 at Cohen's kappa = 0.633 on 40 dual-annotated Python units. The decisive finding is budget-constrained retrieval: at a 256-character per-query budget, CommitDistill reaches 0.750 hit-rate on a 12-query benchmark against BM25's 0.333 and git log --grep's 0.083. On a four-arm paired LLM-as-judge evaluation (n=200 time-travel bug-fixes, two judges) covering control, CommitDistill, a body-budget-matched CD-Hybrid, and BM25, no condition produces a statistically detectable lift over control on the headline mean and CD-Hybrid is indistinguishable from BM25 head-to-head. Extraction over 10,000 commits completes in under 4 seconds on a laptop. Source, annotations, baselines, and a reproducibility script accompany this paper.
arXiv:2605.18239v1 Announce Type: new
Abstract: Large Language Models (LLMs) remain vulnerable to jailbreak attempts that circumvent safety guardrails. We investigate whether multi-turn conversations using low-resource African languages (Afrikaans, Kiswahili, isiXhosa, and isiZulu) can bypass safety mechanisms across commercial LLMs. We translated prompts from existing datasets and evaluated ChatGPT, Claude, DeepSeek, Gemini, and Grok through automated testing and human red-teaming with native speakers. Single-turn translation attacks proved ineffective, while multi-turn conversations achieved English harmful response rates from 52.7% (Claude 3.5 Haiku) to 83.6% (GPT-4o-mini), Afrikaans from 60.0% (Claude 3.5 Haiku) to 78.2% (GPT-4o-mini), and Kiswahili from 41.8% (Claude 3.5 Haiku) to 70.9% (DeepSeek). Human red-teaming increased jailbreak rates compared to automated methods. Over all evaluated languages, the average jailbreak rate increased from 59.8% to 75.8%, with improvements of +20.0% (Afrikaans), +12.7% (isiZulu), +12.3% (isiXhosa), and +1% (Kiswahili), demonstrating that poor translation quality limits jailbreak success. These findings suggest that vulnerabilities in LLMs persist in multilingual contexts and that translation quality is the critical factor determining jailbreak success in low-resource languages.