arXiv:2606.03938v2 Announce Type: replace
Abstract: Multi-epoch training is becoming the standard now that compute is growing faster than the supply of high-quality text. But pretraining a single model saturates within a few passes, long before the compute budget is exhausted. We argue this calls for a conceptual shift from training a single model toward exploring a population of models and aggregating their predictions. We introduce hyper-epoch pretraining (q0), which turns a multi-epoch budget into a population of diverse models whose combined predictions reach a lower validation loss than a single refined model. q0 reduces to three core primitives. A cyclic schedule with anti-correlated learning rate and weight decay collects diverse models from a few parallel trajectories. Chain distillation trains each model against its predecessor so that model quality compounds across the population. A learned prior, fit on a held out set, selects and weights members for any inference budget. On a 1.8B-parameter model trained on 100M FineWeb tokens, q0 matches a strong 256-epoch ensemble baseline using only ~56 epochs (~4.6x fewer), or ~67 epochs (~3.8x fewer) when matched to the baseline's ensemble size, and continues to improve beyond it. These gains reach cumulative ~12.9x data efficiency under the Slowrun setting and transfer to downstream benchmarks. Crucially, the optimal allocation shifts with the budget, so we give prescriptive recipes for how to spend a given epoch budget to maximize generalization, from a single epoch up to the largest budgets.
Science Journals
arXiv:2606.03972v2 Announce Type: replace
Abstract: We present AAD-1, an Asymmetric Adversarial Distillation framework for One-step autoregressive image-to-video generation. State-of-the-art methods adopt adversarial distillation but suffer from motion collapse and training instability, resulting in static videos. AAD-1 addresses these challenges through two key designs in architecture and training strategy. Our key architectural insight is to break the symmetry between generator and discriminator. While the generator remains causal to preserve autoregressive sampling capability, the discriminator attends bidirectionally over the full spatiotemporal context and produces a single holistic realism score for the entire video sequence. This asymmetric design enables the discriminator to effectively detect global temporal failures and long-range drift that cause motion collapse in autoregressive generation. To stabilize training, we introduce a phased strategy that first uses distribution matching to bootstrap a stable one-step generator, providing a warm-up phase that brings the student distribution closer to the teacher before adversarial distillation begins. Extensive experiments on VBench demonstrate that AAD-1 achieves state-of-the-art performance in one-step autoregressive video generation.
arXiv:2107.01629v3 Announce Type: replace-cross
Abstract: Livestreaming has evolved into a thriving industry where creators can directly monetize and engage with their audiences and followers. In practice, creators and platforms typically concentrate their marketing efforts on the period leading up to the livestream. However, livestreaming events naturally transition into recorded formats once the event concludes, creating potential "residual" opportunities for monetization. This study systematically examines consumer demand for live events throughout the entire livestream life-cycle, using data from a large livestreaming platform that allows consumers to purchase the recorded version of a paid live event after the livestream ends. We find that the demand is surprisingly more price-sensitive during the pre-livestream period compared to the post-period. This is partly driven by two mechanisms: consumer self-selection (infrequent consumers who may have missed the live events exhibit a higher willingness to pay for recorded versions) and quality uncertainty (consumers face higher uncertainty in event quality during the pre-period than in the post-period). Our findings generate implications for the pricing and targeting strategies in livestreaming markets.
arXiv:2606.04244v1 Announce Type: new
Abstract: Multimodal large language models are increasingly capable of complex reasoning, yet their performance often degrades when they must externalize a problem through a tool and then reason over the tool's output, specifically when they rely on visual aids. This gap is especially important because real engineering and scientific workflows often rely on visualization tools for analysis, validation, and decision-making. To study this discrepancy, we introduce VAMPS (Visual-Assisted Mathematical Problem Solving), a benchmark for graph-assisted mathematics. VAMPS contains 1,168 multimodal, bilingual multiple-choice question-answer pairs drawn from Iranian University Entrance Exam algebra and calculus problems and expanded with human-reviewed LLM-generated synthetic variants, all selected so that plotting provides a natural solution strategy by revealing intersections, extrema, asymptotes, etc. Designed for both benchmarking and diagnosis, VAMPS goes beyond prior multimodal benchmarks that primarily evaluate reasoning over fixed visual inputs by testing whether a model can benefit from constructing a useful graph and grounding its answer in the resulting visualization. Overall, we found that across a diverse set of models, direct analytical solving surprisingly outperforms tool-enabled visual solving, even on problems where plotting is a natural strategy.
arXiv:2606.04625v1 Announce Type: new
Abstract: Backside power-delivery-network (BSPDN) schemes require wafer-to-wafer bonding steps that do not leave high order shape changes or localized stresses in the bonded stack to ensure that these don't get transferred to the thinned wafer for further lithographic exposure. However, bonding mechanics involve strong adhesive forces which can inherently create localized distortions that lithography tools must eventually compensate. Some contributors to grid distortion of the target wafer are, method of bond initiation, bond front velocity variations, and lack of symmetry between the wafers. In this work, we evaluate a low-distortion bonding approach in the SUSS XBA tool, where some of these contributors are tackled at the source, namely by initiating the bond without a localized external force, and keeping the wafers compliant & symmetric during bonding. Wafers are bonded with a slight pre-stress due to controlled gas pressure applied over the whole backside of the wafers throughout bonding. Wafer-shape measurements of the bonded stacks are used to perform gradient-based in-plane-displacements (IPD) modelling to estimate bonding-induced grid distortion. Dense scanner metrology is used on patterned-bonded wafers to confirm the location and severity of distortion predictions from patterned wafer geometry (PWG) measurements. On the PWG distortion maps and scanner grid readouts, we perform alignment and CPE modelling with different field layouts. Sub-10 nm levels of residual grid distortion are achievable with relatively low-order correction models, and less than or equal to 3 nm by using advanced CPE models. These results demonstrate that pneumatically warped bonding yields low-distortion bonded stacks, and that simple process tuning can decouple the dominant distortion source from edge-related variability.
arXiv:2606.04110v1 Announce Type: new
Abstract: Online evaluation of ranking and retrieval systems often relies on downstream monetization metrics such as app revenue or creator earnings. These metrics are typically heavy-tailed, with a small fraction of users dominating both mean and variance, leading to low statistical power and unreliable conclusions in A/B experiments -- especially under limited traffic.
We present a practical framework for variance reduction in online experiments by combining post-stratification with CUPED. Our approach leverages pre-experiment covariates to improve the sensitivity of monetization experiments without requiring additional traffic. Deployed at ShareChat across ranking-driven monetization experiments, the method substantially reduces variance and improves decision stability, achieving equivalent statistical confidence with ~45\% less traffic than standard metrics. We further discuss practical design choices, guardrails, and limitations, providing guidance on when post-stratification is appropriate for real-world information retrieval and Recommendation systems.
arXiv:2606.04248v1 Announce Type: new
Abstract: Decentralized rigid formation flocking requires a swarm of autonomous agents to maintain a predetermined geometric configuration while moving, relying solely on local sensing and communication. However, existing decentralized control methods struggle to maintain strict inter-agent distance constraints in cluttered environments, often suffering from local minima deadlocks, high frequency control oscillations, or limited flexibility during obstacle navigation, resulting in low success rate. To address these limitations, we propose Rigid Swarm Control (RSC), a decentralized control framework for large-scale rigid formation flocking. To escape local minima via robust long-term planning while ensuring short-term safety, RSC integrates finite-horizon trajectory predictions with a reactive artificial potential field (APF) safety controller within a hybrid architecture. Furthermore, to accelerate formation reassembly after obstacle traversal without interrupting task execution, RSC introduces an online leader-follower reconfiguration mechanism based on stable role exchange. Extensive evaluations in challenging cluttered environments with 25 UAVs demonstrate that RSC reliably unifies rigid formation maintenance, obstacle avoidance, and target tracking. Under strict success criteria - collision-free operation with a maximum relative edge-length error below 10%, RSC achieves an 83% success rate, significantly outperforming existing heuristic and learning-based baselines that fall below 5%.
arXiv:2606.04260v1 Announce Type: new
Abstract: Majority illusion is an undesirable phenomenon in social networks in which agents incorrectly perceive a minority opinion as dominant. This can severely distort collective behavior and decision-making. We study the fundamental question of detecting whether a social network allows for a majority illusion. Formally, in the $q$-Majority Illusion problem, we ask whether there exists a binary labeling of agents in which at least a $q$-fraction of agents have the majority of neighbors with the minority label. We investigate how various structural properties of the underlying social network influence the tractability of this question, and provide a detailed map of its computational complexity.
arXiv:2605.10246v2 Announce Type: replace
Abstract: AI scientist systems are increasingly deployed for autonomous research, yet their academic integrity has never been systematically evaluated. We introduce SCIINTEGRITY-BENCH, the first benchmark designed around a dilemmatic evaluation paradigm: each of its 33 scenarios across 11 trap categories is constructed so that honest acknowledgment of failure is the only correct response, while task completion requires misconduct. Across 231 evaluation runs spanning 7 state-of-the-art LLMs, the overall integrity problem rate reaches 34.2%, and no model achieves zero failures. Most strikingly, across missing-data scenarios, all seven models generate synthetic data rather than acknowledging infeasibility, differing only in whether they disclose the substitution. A further prompt ablation study separates two drivers: removing explicit completion pressure sharply reduces undisclosed fabrication from 20.6% to 3.2%, while the underlying synthesis rate remains unchanged, revealing an intrinsic completion bias that persists independent of prompt-level instructions. These findings point to the absence of honest refusal as a trained disposition as the primary driver of observed failures. We release SCIINTEGRITY-BENCH at https://github.com/liuxingtong/Sci-Integrity-Bench.
arXiv:2605.10679v3 Announce Type: replace
Abstract: Spiking Neural Networks (SNNs) can reduce energy consumption compared to conventional Artificial Neural Networks (ANNs) when spiking activity is sparse and the neuron model is hardware-friendly. However, biologically faithful models are often too costly to implement on FPGAs, whereas very simple models (e.g., IR/LIF) sacrifice part of the neuronal dynamics. In this work, we present an FPGA accelerator for an SNN using Spiking Recurrent Cell (SRC) neurons, providing a trade-off between biological plausibility and hardware cost. We propose a set of mathematical simplifications that remove costly unary operators (\textit{tanh}, \textit{exp}) and avoid floating-point arithmetic through scaling and piecewise-defined approximations. The complete network is implemented in VHDL and validated using spiking traces derived from the MNIST dataset. The weight matrices computed off-line are stored directly in LUT-registers without any adaptation. This demonstrates the robustness of SRC cells. Experiments were conducted on an Artix-7 XC7A200T clocked at 100 MHz. The reference implementation achieves 96.31\% accuracy with a 220-image spiking trace and a processing time of 1.7424 ms per digit. We then investigate accuracy/energy trade-offs by reducing the spiking trace length and quantizing synaptic weights down to 4 bits, achieving 93.32\% accuracy at 0.55 mJ per digit (55 images, 5-bit weights) and 92.89\% at 0.45 mJ (44 images, 4-bit weights). These results show that SRC-based SNNs can deliver competitive performance with reduced energy consumption, while preserving richer neuronal dynamics than standard LIF/IR models.
arXiv:2606.04254v1 Announce Type: new
Abstract: This study investigates whether behavioral and performance indicators derived from a Moodle-based learning management system are associated with university students' depression and anxiety in two undergraduate Computer Engineering courses. Using a quantitative observational design, LMS event logs, academic records, and self-reported Beck Depression Inventory-II and Beck Anxiety Inventory scores from 97 students were integrated. A broad set of behavioral and performance indicators spanning temporal engagement, session structure, deadline-related behavior, page-refresh patterns, and LMS navigation was extracted from raw event logs and analyzed using descriptive statistics, independent-samples t-tests with Benjamini-Hochberg FDR correction, effect sizes, and Spearman correlations; inventory scores were confirmed invariant by sex and academic year. Several indicators were significantly associated with depression and anxiety. Higher depression was associated with shifted temporal activity patterns, longer session durations, and shorter homework submission lead times, while higher anxiety was associated with concentrated temporal engagement and session-based differences. These findings suggest that routine LMS data can provide meaningful behavioral signals related to student well-being and may support earlier educational awareness of students who experience mental-health-related strain. At the same time, such indicators should be interpreted as contextual and non-diagnostic markers rather than as substitutes for clinical assessment.
arXiv:2605.11130v4 Announce Type: replace
Abstract: Critical events in multivariate time series, from turbine failures to cardiac arrhythmias, demand accurate prediction, yet labeled data is scarce because such events are rare and costly to annotate. We introduce HEPA (Horizon-conditioned Event Predictive Architecture), built on two key principles. First, a causal Transformer encoder is pretrained via a Joint-Embedding Predictive Architecture (JEPA): a horizon-conditioned predictor learns to forecast future representations rather than future values, forcing the encoder to capture predictable temporal dynamics from unlabeled data alone. Second, we freeze the encoder and finetune only the predictor toward the target event, producing a monotonic survival cumulative distribution function (CDF) over horizons. With fixed architecture and optimiser hyperparameters across all benchmarks, HEPA handles water contamination, cyberattack detection, volatility regimes, and eight further event types across 11 domains, exceeding leading time-series architectures including PatchTST, iTransformer, MAE, and Chronos-2 on at least 10 of 14 benchmarks, with an order of magnitude fewer tuned parameters and, on lifecycle datasets, an order of magnitude less labeled data.
arXiv:2605.15118v2 Announce Type: replace
Abstract: We introduce a reusable framework for auditing whether LLM attack benchmarks collectively cover the threat surface: a 4$\times$6 Target $\times$ Technique matrix grounded in STRIDE, constructed from a 507-leaf taxonomy -- 401 data-populated and 106 threat-model-derived leaves -- of inference-time attacks extracted from 932 arXiv security studies (2023--2026). The matrix enables benchmark-external validation -- auditing collective coverage rather than individual benchmark consistency. Applying it to six public benchmarks reveals that the three primary frameworks (HarmBench, InjecAgent, AgentDojo) occupy non-overlapping cells covering at most 25\% of the matrix, while entire STRIDE threat categories (Service Disruption, Model Internals) lack any standardized evaluation, despite published attacks in these categories achieving 46$\times$ token amplification and 96\% attack success rates through mechanisms which no benchmark tests. The corpus of 2,521 unique attack groups further reveals pervasive naming fragmentation (up to 29 surface forms for a single attack) and heavy concentration in Safety \& Alignment Bypass, structural properties invisible at smaller scale. The taxonomy, attack records, and coverage mappings are released as extensible artifacts; as new benchmarks emerge, they can be mapped onto the same matrix, enabling the community to track whether evaluation gaps are closing.
arXiv:2606.04262v1 Announce Type: new
Abstract: Large language models (LLMs) are increasingly used for everyday health questions, including whether a user can safely take another dose of an over-the-counter (OTC) medication. Yet this common safety-relevant setting remains underexplored in existing medical QA evaluations, where correct answers require tracking dose timing, computing rolling 24-hour intake, following product-label constraints, and handling incomplete medication histories. We introduce DOSEBENCH, a focused benchmark of 81 curated OTC dosing scenarios focused on adult acetaminophen and ibuprofen use, with manually annotated gold references. We evaluate four LLMs across repeated runs using metrics for decision correctness, consistency, explanation verifiability, failure types, and confidence-related signals, resulting in 1,620 model responses. Our results show that models frequently struggle with rolling-window reasoning and ambiguity-sensitive cases and that stable or confident-looking responses can still violate dosing constraints. These findings suggest that OTC dosing QA provides a narrow yet practical testbed for evaluating temporal reasoning, constraint following, and safety-relevant uncertainty handling in medical QA.
arXiv:2606.04494v1 Announce Type: new
Abstract: Biomedical agents promise to automate complex biological workflows, yet current systems face two fundamental bottlenecks: bioinformatics tools are highly heterogeneous in interfaces and execution environments, while agent planning still relies on flat prompt-retrieved tool descriptions. As biomedical software ecosystems grow, this coupling between tool coverage and context size leads to tool confusion, unstable planning, and inefficient execution. We introduce BioManus, an MCP-native biomedical agent built on graph-scaffolded planning over structured biological capabilities. BioManus first introduces the BioinfoMCP Compiler, which converts heterogeneous bioinformatics software into standardized MCP servers, yielding a large executable MCP ecosystem. It then organizes this ecosystem as a typed heterogeneous MCP graph over tools, operations, datatypes, and workflow stages. At inference time, BioManus retrieves compact task-specific subgraphs, synthesizes operation-level workflow scaffolds. This design decouples planning complexity from raw tool inventory size, achieving a context compression ratio of Theta(N / (h * m_bar)) under high-recall retrieval, where N is the total tool count, h is the workflow horizon, and m_bar (much smaller than N) is the average number of candidate tools per operation. Experiments on BioAgentBench and LAB-Bench show that BioManus improves execution accuracy, workflow validity, and context efficiency over advanced biomedical agent baselines. This work suggests a paradigm shift: scalable biomedical reasoning requires structured executable capability graphs rather than increasingly larger prompt-level tool retrieval.
arXiv:2605.15152v2 Announce Type: replace
Abstract: LLM quantization has become essential for memory-efficient deployment. Recent work has shown that quantization schemes can pose critical security risks: an adversary may release a model that appears benign in full precision but exhibits malicious behavior once quantized by users. However, existing quantization-conditioned attacks have been limited to relatively simple quantization methods, where the attacker can estimate weight regions that remain invariant under the target quantization. Notably, prior attacks have consistently failed to compromise more popular and sophisticated schemes, limiting their practical impact. In this work, we introduce the first quantization-conditioned attack that consistently induces malicious behavior that can be triggered by a broad range of advanced quantization techniques, including AWQ, GPTQ, and GGUF I-quants. Our attack exploits a simple property shared by many modern quantization methods: large outliers can cause other weights to be rounded to zero. Consequently, by injecting outliers into specific weight blocks, an adversary can induce a targeted, predictable weight collapse in the model. This effect can be used to craft seemingly benign full-precision models that exhibit a wide range of malicious behaviors after quantization. Through extensive evaluation across three attack scenarios and LLMs, we show that our attack achieves high success rates against a broad range of quantization methods on which prior attacks fail. Our results demonstrate, for the first time, that the security risks of quantization are not restricted to simpler schemes but are broadly relevant across complex, widely-used quantization methods.
arXiv:2606.04562v1 Announce Type: new
Abstract: Purpose The WHO's COVID-19 non-pharmaceutical interventions (e.g., lockdowns, vaccinations) effectively curb transmission but impose heavy economic strains. Existing research often neglects individual behaviors and falsely assumes perfect infection tracking and flawless policy execution, failing to account for real-world uncertainties and errors. Methods We propose an integrative approach incorporating uncertainties in both epidemic measurement (infections/hospitalizations) and policy implementation. We built a simulation model of 1,000 individuals making real-time choices regarding mask-wearing, vaccination, and shopping. Concurrently, policymakers deploy interventions (lockdowns, mandates) based on health and economic observations. This framework is driven by hierarchical reinforcement learning agents, utilizing deep Q-networks alongside uncertainty-aware policy gradient variants (DDPG and TD3). Results The simulations effectively managed the epidemic's progression. Masking and vaccinations proved highly effective, significantly reducing both the outbreak's peak height and duration. By integrating individual behaviors, policy uncertainties, and multifaceted interventions, our dynamic control approach successfully mitigated the epidemic's impact. Conclusions Our model overcomes previous research limitations by embedding uncertainty and human behavior into public health policy frameworks. The simulation demonstrates that accounting for individual choices and imperfect data is crucial for designing effective interventions during complex pandemics, with masks and vaccines serving as pivotal tools.
arXiv:2606.04665v1 Announce Type: new
Abstract: Deep unsupervised domain adaptation (Deep UDA) methods successfully leverage rich labeled data in a source domain to boost the performance on related but unlabeled data in a target domain. However, algorithm comparison is cumbersome in Deep UDA due to the absence of accurate and standardized model selection method, posing an obstacle to further advances in the field. Existing model selection methods for Deep UDA are either highly biased, restricted, unstable, or even controversial (requiring labeled target data). To this end, we propose \textit{Deep Embedded Validation} (\textbf{DEV}), which embeds adapted feature representation into the validation procedure to obtain unbiased estimation of the target risk with bounded variance. The variance is further reduced by the technique of control variate. The efficacy of the method has been justified both theoretically and empirically.
arXiv:2606.04381v1 Announce Type: new
Abstract: Recent large language models (LLMs) often appear to exhibit spatial reasoning ability; however, this capability is largely \emph{symbolic}, arising from pattern matching over spatial language rather than true \emph{geometric} reasoning over space. Because LLMs operate on discrete tokens, they lack native support for continuous spatial representations, explicit geometric computation, and structured spatial operators. To address this limitation, we introduce the \emph{Spatial Language Model (SLM)}, the first multimodal LLM that treats location information as a first-class modality and enables geometric spatial reasoning within the model's inference process. SLM directly operates on learned spatial representations rather than textual descriptions of spatial relations. To support effective training, we construct a \emph{Spatial Instruction Dataset} that aligns spatial representations, atomic geometric operations, and natural language instructions. We further propose a new benchmark named \emph{SpatialEval}, which is designed to evaluate spatial reasoning across attributes, distance, topology, and relative-position tasks. Extensive experiments show that SLM significantly outperforms existing LLM-based approaches that rely on symbolic reasoning via prompt engineering or textual abstraction, demonstrating the benefits of integrating geometric spatial representations for robust spatial reasoning.
Our instruction dataset, evaluation benchmark, model training codes, and models' checkpoints can be found at:
\hyperlink{https://github.com/chuchen2017/SLM}{https://github.com/chuchen2017/SLM}.
arXiv:2606.04557v1 Announce Type: new
Abstract: Large Language Models can reason over long contexts, yet prefilling millions of tokens is wasteful as much of the content remains static across queries. Cartridges address this by distilling document collections into reusable key-value (KV) caches that eliminate prefilling while preserving accuracy. A critical limitation of this approach is that cartridges are monolithic and non-compositional: encoding an entire collection into a single KV block does not scale, and naively mixing cartridges trained in isolation collapses performance to near chance. We introduce Cartridges at Scale (CAS), a training framework for scalable multi-cartridge learning with dynamic distractor mixing and a memory-efficient budget manager that rotates hundreds of per-document cartridges between GPU and persistent storage. Our approach scales to collections exceeding a million tokens, improving over a monolithic cartridge by 10-31 points at comparable token budgets. Oracle cartridge accuracy falls within 2-6 points of full in-context learning even at high compression. When paired with retrieval for cartridge selection, CAS matches or exceeds conventional RAG accuracy while consuming 3-4x fewer prompt tokens.
arXiv:2605.15741v2 Announce Type: replace
Abstract: Pixel-space diffusion models bypass the reconstruction bottleneck of Variational Autoencoders (VAEs) but face a fundamental "granularity dilemma": capturing global semantics favors large patch scales, while generating high-fidelity details demands fine-grained inputs. To address this issue, we propose HyperDiT, a unified framework establishing Hyper-Connected Cross-Scale Interactions to bridge the semantic and pixel manifold. Diverging from injecting semantics by AdaLN, HyperDiT utilizes Cross-Attention mechanisms, enabling fine-grained tokens to query multi-level semantic anchors globally. To resolve the spatial mismatch during multi-scale interactions, we introduce Scale-Aware Rotary Position Embedding (SA-RoPE) to ensure precise geometric alignment among tokens of varying patch sizes. Furthermore, we incorporate Registers to learn the dense semantics from a pretrained Visual Foundation Model (VFM), effectively reducing generation hallucination and artifacts. Extensive experiments demonstrate that HyperDiT achieves state-of-the-art (SoTA) FID of $\mathbf{1.56}$ on ImageNet $256\times256$ directly within the pixel space. By combining the fine-grained stream with semantic guidance, HyperDiT offers a superior paradigm for high-fidelity pixel generation.
arXiv:2606.04543v1 Announce Type: new
Abstract: Artificial intelligence in education is evolving from passive chatbots to proactive AI agents capable of initiation and goal-directed interactions. While offering opportunities for personalised learning, this shift risks undermining learner agency and cognitive effort. This paper reviews six pedagogical principles-prior knowledge activation, collaborative learning, problem-based learning, formative assessment, scaffolding, and metacognition-through the lens of agentic AI. We discuss the tension between automation and learning, proposing design recommendations that prioritise intentional friction, dynamic scaffolding, human-in-the-loop oversight, and considered AI utilisation to ensure AI supports rather than supplants human learning.
arXiv:2606.04035v1 Announce Type: new
Abstract: We present a systematic study of domain-dependent safety behavior in open-weight LLMs: 7 standardized experiments across 7 ethical domains, testing 5 models (12B--70B) in 4,200 interactions with dual-judge validation. Using a dual-condition methodology, each scenario tested in both an analytical framing (identify the harm) and an operational framing (help commit the harm), we find compliance rates vary from 14.7% (human trafficking) to 85.7% (surveillance design), a 71-percentage-point span with non-overlapping cluster-bootstrapped 95% CIs. Trustworthy deployment requires predictable safety behavior, yet we find compliance is highly context-dependent: the same model (Mistral Nemo 12B) provides surveillance designs in 100% of requests but assists with trafficking in only 26.7%. This unpredictability is opaque to deployers: the technical framing bypass, where harmful requests reframed as engineering problems override safety training without any external signal that refusal thresholds have shifted. Within-domain heterogeneity reaches 84.4pp, meaning safety behavior cannot be predicted even at the domain level. A replication on five frontier closed models (GPT-4.1/5.2, Claude Haiku/Sonnet/Opus 4.x; n=4,163 responses) accessed via the GitHub Copilot CLI deployed-product surface reproduces the same domain stratification, attenuated in absolute level but identical in shape, with the two low-codification domains (science fraud, surveillance) again the most permissive. These results show that current safety mechanisms lack the transparency and consistency required for trustworthy AI deployment.
arXiv:2606.04628v1 Announce Type: new
Abstract: RAMPART is a compile-time memory model and pure in-RAM block registry for LLM-based agents. Context assembly is a programmable runtime operation where content is compiled from a structured registry under explicit policy for ordering, inclusion, and eviction. Five composable primitives (promote, gate, write, evict, rollback) act on named addressable blocks before compilation at zero prompt-token cost. Provenance tags and non-evictable authorship flags implement a permissioned memory model with block-level ownership. Controlled probes with Qwen3-8B Q4 show that compile-time placement and the structural relationship between blocks and the task query affect task success, with the cliff falling at roughly the seventh block position when the task follows the registry and the twelfth when it precedes. Grouping the critical block with content-adjacent neighbours and promoting the group as a unit lifts task success by tens of percentage points at positions where single-block placement fails. Cross-model replication on Qwen2.5-7B, Llama-3.1-8B, Mistral-7B-v0.3, and Qwen3-14B shows the content-priming effect appears at the same absolute positions across families, with magnitude varying with model strength. Block grouping raises Mistral's mean pass rate roughly fivefold at the hardest registry size, and a smaller model with the intervention can outperform a larger model without it in the mid-registry zone. Relevance gating reduces prompt cost by 67.8\% while recovering 83% of the promoted-condition success rate. Schema eviction produces 0% invocations against 100% with the schema present, a property policy-based approaches cannot guarantee by construction. Shared-registry coordination reduces inter-agent communication to a method call at zero coordination token cost.
arXiv:2606.04049v1 Announce Type: new
Abstract: We develop a general operator-theoretic route that turns Kato-type quasilinear evolution systems on a Banach scale $(Z,X)$ into finite-dimensional interacting approximations. The construction proceeds in two steps. First, one introduces a regularized family $(A_\varepsilon,f_\varepsilon)$ indexed by a scale parameter $\varepsilon>0$, for which the drift $A_\varepsilon[t,z]z+f_\varepsilon[t,z]$ takes values in an output space $Y$ suitable for discretization. Second, one discretizes this regularized dynamics by a sampling-reconstruction pair $(P_N,R_N)$ and obtains an interacting ODE on a finite-dimensional state space $V_N\simeq\R^{dN}$. Our main abstract theorem provides a quantitative estimate of the discrepancy $y_\varepsilon^N-y$ between the lifted discrete solution and the exact one, separating the regularization error $\chi(\varepsilon)$ from the discretization error $(1+L_\varepsilon)N^{-\gamma}$, where $L_\varepsilon$ measures the size of the regularized drift in the output norm. This makes explicit the trade-off between the regularization scale $\varepsilon$, the discretization scale $N$, and the possible deterioration of $L_\varepsilon$ as $\varepsilon\to 0$. As a running example, we focus on quasilinear PDEs on bounded Lipschitz domains with boundary conditions. We show that Burenkov's variable-step mollifiers provide a boundary-compatible kernelization: they regularize differential operators into explicit integral-interaction operators supported inside the domain and preserve boundary traces of sufficiently regular fields. In this setting one can choose an output space $Y$ for which $L_\varepsilon$ remains uniformly bounded, leading to algebraic convergence rates in $N$ for quasi-uniform discretizations.