# Research Papers — 2026-06 · bundle page 141 of 260 > Full-text markdown of papers #7001–7050 of 12993 published in 2026-06, newest first. AI-readable mirrors of the full 12993-paper 2026-06 archive (50 papers per file). Canonical page: https://aipod.dev/papers/archive/2026-06/page/141 [← Previous bundle](https://aipod.dev/papers/archive/2026-06/page/140.md) [Next bundle →](https://aipod.dev/papers/archive/2026-06/page/142.md) ### Market Strategy Evaluation for Prosumers in Local Electricity Markets Authors: Lukas Peter Wagner, Raoul Bisson, Felix Gehlhoff Published: 2026-06-12 Categories: cs.GT, cs.AI, eess.SY arXiv: https://arxiv.org/abs/2607.18272v1 Prosumers equipped with distributed generation and flexible loads form autonomous cyber-physical energy systems that control local resources and participate in local energy markets with minimal human intervention. This work develops and evaluates an agent-based simulation platform in which agents, representing prosumer households with photovoltaic systems, battery storage systems, electric vehicles, and heat pumps, participate in a uniform-price double-sided call auction. The effect of individual bidding strategies on community-level efficiency and prosumer-level financial outcomes is incompletely understood, particularly when prosumers with heterogeneous portfolios interact in one market. Four market strategies of increasing complexity are compared: a zero-intelligence constrained baseline, a boundary-price strategy, an extended storage cascade, and a market-adaptive pricing strategy. The simulation is conducted on a community of 33 prosumers at 15-minute resolution, spanning summer, winter, and spring to characterize seasonal variation. Results show that rule-based resource control substantially reduces community energy expenditure: the extended storage cascade achieves a total cost of 39.06 EUR compared to 62.38 EUR under the zero-intelligence baseline, a reduction of 37.4 %. The market-adaptive strategy yields the highest aggregate community financial gain through local energy market participation (14.40 EUR vs. 10.28 EUR for the baseline, a gain of 40.1 %) under summer conditions. Strategy effectiveness depends on both portfolio composition and seasonal supply conditions, requiring joint evaluation of resource control and pricing decisions. --- ### Comparing Spectrogram Front-Ends for Abnormal Heart-Sound Detection with a Convolutional Neural Network Authors: Abhinav Pala, Dhanush Pala Published: 2026-06-12 Categories: cs.CY, cs.AI, cs.LG, cs.SD, eess.AS arXiv: https://arxiv.org/abs/2607.16220v1 Heart disease kills a lot of people, and one cheap way to catch it early is by listening to heart sounds with a stethoscope, or better yet, just recording them and running them through a model. This project is a binary classification task: take a short clip of someones heartbeat and decide if it sounds normal or abnormal. Instead of trying out a bunch of different models, we kept the CNN the same the whole time and just changed how we turned the raw audio into a picture for it to look at. We tried three ways of doing that: a regular logmel spectrogram, PCEN (which basically normalizes each frequency bin over time), and a multi resolution version that stacks a few different window sizes together. We ran all three on the PhysioNet 2016 heart-sound dataset with the exact same setup but same model, same optimizer, same random seed. Turns out all three do pretty well at catching abnormal cases (sensitivity around 0.95), but PCEN and multi-resolution both edge out the plain logmel on the official PhysioNet accuracy metric (0.915 and 0.916 vs. 0.910). We also ran Grad-CAM to see where the model was actually looking, and it mostly focused on the low frequencies where S1 and S2 heart sounds live, which is a good sign that it learned something real --- ### LessonBench-V1: A Benchmark Dataset for Evaluating AI Lesson Generation Agents Authors: Ravidu Suien Rammuni Silva, Ahmad Lotfi, Isibor Kennedy Ihianle, Golnaz Shahtahmassebi, Jordan J. Bird Published: 2026-06-12 Categories: cs.CY, cs.AI, cs.LG, cs.MM arXiv: https://arxiv.org/abs/2607.13041v1 Large Language Model (LLM) based AI educational content generation systems are increasingly being developed, yet no standardised benchmark exists to systematically evaluate them. This study introduces LessonBench-V1, a benchmark dataset comprising 647 human-written lessons paired with LLM-based reverse-engineered lesson plans across 240 STEM topics spanning mathematics, physics, chemistry, and computer science. The lessons are drawn from 97 trusted open sources, including LibreTexts, Brilliant.org and GeeksForGeeks. Each lesson plan is human-reviewed and produced through a pedagogically grounded methodology that synthesises Bloom's Taxonomy, Gagné's Events, Merrill's First Principles, and the 5E Instructional Model. The lesson plans capture 3,620 learning objectives with pedagogical metadata, enabling systematic, reproducible evaluation of lesson-generation AI agents and supporting further research. The study further proposes a three-dimensional evaluation pipeline for use with the dataset. --- ### Final Authority in AI Governance: Frontier-Provider Sovereignty and Action-Centered Deployer Governance Authors: Zexun Wang Published: 2026-06-12 Categories: cs.CY, cs.AI arXiv: https://arxiv.org/abs/2607.13040v1 This paper examines where final authority should sit once capable AI systems are embedded in organizational workflows. It compares two governance models. The first, frontier-provider sovereignty, assigns privileged authority to the provider of the most capable models and is reflected in contemporary arguments for frontier-model testing, release gating, transparency duties, and compute-related controls. The second, action-centered deployer sovereignty, places final authority over high-impact actions with the organization that authorizes the action, embeds it in a business process, and bears the downstream legal, operational, and commercial consequences. The paper combines comparative reading of public governance frameworks with implementation-informed analysis of runtime heterogeneity and enterprise control requirements. It compares EU AI Act guidance, the NIST AI Risk Management Framework, Singapore's Model AI Governance Framework for Agentic AI, recent Japanese AI policy instruments, and Canada's voluntary code and managerial guidance. Across these materials, the paper finds stronger support for distributed operational accountability than for unilateral frontier-provider control. It further argues that rapid enterprise adoption, declining provider transparency, and widening control gaps increase the value of a portable governance layer centered on governed action rather than on provider-native session objects. The conclusion is layered rather than absolutist: strong upstream authority remains justified for frontier capability gating, but final authority over concrete enterprise action is better located with the deployer and consequence-bearer. --- ### Safeguard-Conditioned Uplift: Measuring Utility-Risk Frontiers for Dual-Use Biology Assistants Authors: Dipesh Tharu Mahato Published: 2026-06-12 Categories: cs.CY, cs.AI arXiv: https://arxiv.org/abs/2607.13039v1 Safety evaluations for dual-use biology assistants often measure base-model capability, refusal behavior, or jailbreak success. These metrics miss a deployment question: for a fixed base model, how does the access condition users actually see change benign utility and harmful actionable assistance? I introduce safeguard-conditioned uplift, a protocol for comparing deployed access conditions through a human-judged utility-risk frontier. I evaluate Claude Sonnet 4.6 and Gemini 3.5 Flash under helpful prompting, safety prompting, and an external safeguarded assistant on a 108-task surrogate benchmark, with the headline claim restricted to a locked 18-task held-out split. In a 600-row blinded human audit, the safeguarded assistant reduces harmful actionability relative to helpful prompting by -0.063 over 49 matched response pairs, with bootstrap 95% interval (-0.117, -0.011), while correctness changes by +0.009 with interval (-0.057, +0.077). Adaptive, Test-B, cue-ablation, and controller-baseline checks support the measurement story but also show non-dominance: safety prompting is often strongest for Claude, while external control helps more for Gemini and can reduce benign utility. The contribution is not a universal defense. It is a deployment-level evaluation target, plus a learned risk-budgeted calibration procedure, for measuring how user-facing access conditions move the utility-risk frontier. --- ### Semidirect Fourier Delta Attention: Phase-Controlled Delta Memory with Constructive Chunk-WY Kernels Authors: Tiantian Zhang Published: 2026-06-12 Categories: cs.LG arXiv: https://arxiv.org/abs/2607.11897v1 Linear attention replaces softmax attention's growing KV cache with a fixed recurrent state, but this compression limits exact state tracking and long-context memory. We introduce \emph{Semidirect Fourier Delta Attention} (SFDA), a phase-controlled generalization of Kimi Delta Attention that replaces real diagonal decay with block-rotational Fourier control: \( S_t=(I-β_t k_tk_t^*)Λ_tS_{t-1}+β_tk_tv_t^*, \qquad Λ_t=\diag(α_t\odot e^{iθ_t}). \) Our main result is a constructive chunk-WY factorization for products \(A_t=Λ_t-u_tr_t^*\), giving \( A_t\cdots A_1=Γ_t-Y_tM_tW_t^* \) with rank growth bounded inside fixed chunks. This yields an exact affine chunk transfer, formal stability and complexity bounds, and a compact characterization of phase-plus-low-rank memory. We verify the algebra numerically and show in toy state-tracking experiments that SFDA learns cyclic memory where the phase-disabled KDA baseline remains near chance. Fused kernels and large-scale language-model comparisons are left to future work. --- ### OmniPMNet: Bridging discrete and gridded PM10 forecasts via omni-query neural processes Authors: Shuangshuang He, Shuo Wang Published: 2026-06-12 Categories: cs.LG, cs.AI, physics.ao-ph arXiv: https://arxiv.org/abs/2607.11896v1 Forecasting particulate matter (PM10) requires both station-scale accuracy and continuous spatial fields, especially during severe dust storms. Chemical transport models (CTMs) provide gridded forecasts but retain local biases, whereas graph neural networks (GNNs) track monitoring sites well at short lead times but do not produce gridded outputs. Here we present OmniPM-Net, a Convolutional Conditional Neural Process (ConvCNP)-based fusion model that reconciles these two forecast types within a shared spatial representation. A terrain-aware Gaussian set convolution lifts irregular GNN station forecasts onto a regular grid, where a multi-scale Spatial Source Attention (SSA) module blends them with Copernicus Atmosphere Monitoring Service (CAMS) forecasts; a shared omni-query readout then decodes this representation into consistent PM10 predictions at either stations or grid cells over a 108 h horizon. Evaluated across 1,618 air-quality monitoring stations throughout China over the full year of 2024, OmniPM-Net matches the station-level accuracy of the stronger GNN baseline (mean absolute error 21.14 versus 22.00 ug/m3) and reduces the CAMS mean absolute error by 30%, while simultaneously delivering the gridded fields that the discrete GNN cannot. Its clearest gains are in the high-concentration tail, where the 90th-percentile MAE falls by 9% relative to the GNN and 25% relative to CAMS, and during dust episodes, where it improves categorical detection skill while tracking the evolving spatial trajectory. --- ### Faithful, Not Corrective: Message-Format Effects in Multi-Hop Agent Relays Are Tier-Dependent Authors: Zayx Shawn Published: 2026-06-12 Categories: cs.AI, cs.LG arXiv: https://arxiv.org/abs/2607.09678v1 When LLM agents hand off information to one another, does the message format matter? Two literatures disagree: format-optimization work reports that structured messages cut cost without hurting accuracy, while format-restriction work finds that imposing structure degrades generation -- and neither measures what happens when a message traverses multiple hops, where copy fidelity, not one-shot generation, dominates. We introduce a controlled relay testbed: briefs of twelve programmatically generated atomic facts are re-encoded hop-by-hop in five formats (free NL, precision-instructed NL, JSON, triples, key-value) over six hops, scored by a fixed strong grader against programmatic ground truth, across two relay-capability tiers, a cognitive-load condition, and a paired-fork error injection. We find that message-format effects are tier-dependent. (i) Under faithful-relay instructions a strong relay is nearly lossless -- the documented "telephone-game" collapse does not occur -- and adding per-hop cognitive load leaves format-level fidelity unchanged (within +/-1.8 points) while raising generation cost by 24-53%. (ii) Under a weak (1.5B) relay the across-format spread of six-hop recall grows by a factor of 8.7 (from 2.3 to 20.5 points), driven by two opposing mechanisms -- an encoding toll paid by the rigid formats and drift resistance specific to the fixed-key JSON schema -- that flip the format ranking in transit. (iii) In a paired-fork injection, an injected wrong value, once present, persists to the final hop in 83-100% of chains in every format, closely matching each format's retention of the true value, with no detectable collateral damage to neighboring facts. Structure buys a faithful, error-localizing channel -- not an error-correcting code -- and format choice should follow the weakest relay in the pipeline. --- ### Sticky Routing: Training MoE Models for Memory-Efficient Inference Authors: Ali Kayyam Published: 2026-06-12 Categories: cs.LG, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2607.08780v1 Mixture-of-Experts (MoE) models activate only a sparse subset of experts per token, yet consecutive tokens frequently activate different experts -- causing constant weight swapping between slow storage and fast memory on edge devices. Existing remedies are either system-level (caching heuristics) or post-hoc (router fine-tuning), leaving the root cause unchanged during pretraining. We propose StickyMoE, a differentiable routing consistency loss that penalises abrupt expert switches between adjacent tokens, encouraging the router to maintain the same expert assignment across semantically coherent spans. StickyMoE requires no architectural changes, adds a single hyperparameter lambda, and unlike post-hoc methods, allows expert representations and routing decisions to co-adapt from the first training step. Experiments on small-scale MoE language models show that StickyMoE reduces the expert switch rate by up to 60% with less than 4% perplexity degradation, Pareto-dominating post-hoc fine-tuning on the quality-locality frontier. Routing temporal locality is most efficiently instilled at training time. --- ### Signed Symmetric Quantization for Few-Bit Integers Authors: Ian Colbert, Eashan Dash, Pablo Monteagudo-Lago, Juan Amboage, Srinidhi N, Giuseppe Franco, Nicholas J. Fraser, Arun Ramachandran Published: 2026-06-12 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2607.08779v1 The signed integer alphabet contains one more negative representable value than positive. Yet, by convention, the standard symmetric integer quantizer fixes its scale to be strictly positive, which assigns this extra representable value to the negative tail and can force clipping of positive outliers. In this work, we show that, at few-bit precision, such clipping is a non-trivial source of quantization error. Asymmetric quantization addresses this problem with a zero point, shifting the grid toward the observed data range; however, this flexibility is well-known to carry a runtime penalty. For example, in llama.cpp on an AMD EPYC(TM) "Turin" CPU, a 4-bit symmetric format uses up to 9% less memory with up to 2.45$\times$ higher throughput than its asymmetric counterpart. We highlight signed symmetric quantization as a third option that retains the runtime profile of symmetric quantization without the penalty of the asymmetric format: our signed absmax grid places the extra representable value on the dominant-outlier tail through a principled and lightweight sign selection rule while keeping the zero point at zero. Our theoretical analysis offers two main results. First, we establish the signed absmax grid as conditionally bound-optimal on $\ell_2$ quantization error, and show that the condition holds for 88-99% of weight groups across pre-trained large language models (LLMs) at low bit widths. Second, we show that negating the scale of a standard symmetric quantizer is analytically equivalent to a unit zero point shift on the same signed integer alphabet. We empirically validate our proposal on models from the Qwen3, Qwen3.5, and Llama3 families, and observe improvement in perplexity and downstream few-shot accuracy over the standard unsigned symmetric quantizer at no extra inference cost --- ### iLENS: Interpretable LLM-Guided Mixture-of-Experts for Neuroimaging Survival Analysis Authors: Farica Zhuang, Seong Woo Han, Zixuan Wen, Shu Yang, Yize Zhao, Li Shen Published: 2026-06-12 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2607.08778v1 Alzheimer's Disease (AD) is a complex neurodegenerative disorder that continues to impact millions of people worldwide. Predicting AD conversion during the prodromal stage remains critical for disease understanding and patient care. As such, survival models are widely used for AD risk prediction, yet they are typically static predictors with limited interpretability and no capacity for natural language reasoning. In this work, we propose iLENS, an interpretable large language model (LLM) guided framework based on mixture-of-experts (MoE) for survival prediction in AD conversion. Our approach uses LLM to synthesize structured neuroimaging measurements and unstructured information to guide expert routing. Our framework demonstrates competitive predictive performance and capability in patient subtyping. Furthermore, our framework provides transparent, biologically grounded rationales for its routing decisions, bridging the gap between high-performance survival analysis and interpretable clinical decision support. --- ### RANSAC Scoring Done Right Authors: James Pritts, Felix Seegräber, Kevin Köser Published: 2026-06-12 Categories: cs.LG, cs.CV arXiv: https://arxiv.org/abs/2606.27385v1 The most widely used RANSAC variants score candidate models by counting inliers or summing per-point scores that saturate beyond a residual threshold. Every such score requires a user-supplied parameter that is a function of the inlier scale, which must itself be estimated from contaminated data. We remove this dependence by reversing the usual order of inference: rather than estimating the scale and then scoring against it, we marginalize the inlier scale analytically in closed form under a conjugate Inverse-Gamma prior for a fixed inlier partition, then optimize over partitions. A single closed-form expression spans the non-informative Jeffreys limit and informative empirical-Bayes priors, so the same score adapts across data-rich and data-scarce regimes without any change to the algorithm. The proposed RANSAC score is the first in which the inlier scale is genuinely absent from the formula. The score admits O(N log N ) computation via sort-and-sweep. On a benchmark of nearly 70 000 image pairs spanning different two-view estimation problems and both engineered and learned feature pipelines, the proposed score exceeds the state of the art (RANSAC, MSAC, GaU, MAGSAC): it stays nearly flat under threshold miscalibration where baselines degrade, reaches near-optimal accuracy from as few as two validation pairs where baselines need ont he order of 100 times more,. and tightens its prior regularization as validation data grows scarce. --- ### The Geometry of Saturation: Effective Rank Predicts When Labels Stop Helping in Few-Shot Classification Authors: Arnav Gupta Published: 2026-06-12 Categories: cs.LG arXiv: https://arxiv.org/abs/2606.24903v2 Few-shot label acquisition lacks a label-free signal for when additional labels cease to improve accuracy: existing stopping criteria either require a held-out validation set (violating the few-shot premise) or rely on theoretically ungrounded heuristics, so we introduce the spectral saturation index $S(K)=\mathrm{erank}(\hatΣ_W^{(K)})/K$, the exponential spectral entropy of the pooled within-class covariance normalized by per-class support size $K$, which measures the exploration rate per label and falls below a fixed threshold $τ=0.02$ once the explored spectral subspace saturates and marginal accuracy gains vanish; across 49 real tasks (binary, 5-way, 10-way) and three frozen backbones (PCA-50, CLIP ViT-B/32, DINOv2 ViT-S/14), $S(K)$ correlates strongly with the marginal gain on doubling the support set ($ρ_{\text{pool}}=0.6366$, $p=2.9\times10^{-57}$, cluster-bootstrap 95\% CI $(0.551,0.720)$), a fixed $τ=0.02$ classifies stop/continue decisions with cluster-bootstrap $\mathrm{AUC}=0.787$ (95\% CI $(0.713,0.860)$) with high recall on meaningful gains ($ΔA>1\%$), and a partial correlation controlling for $\log K$ yields $ρ_{\text{partial}}=0.324$ ($p=1.65\times10^{-13}$), confirming $S(K)$ carries spectral information beyond shared $K$-dependence; theory predicts this from first principles, since the population effective rank sets the saturation scale $K_{\text{sat}}\approx\mathrm{erank}(Σ_W)/τ$, $τ=0.02$ sits at the boundary between the first and second descent (Nakkiran et al., 2021), and $O(1/K)$ bias in the sample effective rank explains the small-$K$ hump in $S(K)$; for unregularized linear probes ($C=\infty$), practitioners should halt when $S(K)<0.02$ (PCA-50, hard stop) or monitor $S(K)$ dropping from $\sim0.3\to0.05$ (foundation models, diminishing-returns signal), with computation costing $\sim1$ ms at $d=50$. --- ### Failure Modes of Large Language Models on Research-Level Mathematics: A Taxonomy and an Empirical Characterisation Authors: Arnesh Banerjee, Ayushi Bhattacharjee Published: 2026-06-12 Categories: cs.DL, cs.AI arXiv: https://arxiv.org/abs/2606.24902v1 The "First Proof" benchmark (1) posed ten research-level mathematics questions to the strongest publicly available LLMs and found them consistently wrong-not silent, but confidently, fluently wrong. This paper asks why. Working from the per-question post-mortems in First Proof's Appendix A, I identify four failure modes: citation fabrication (F1), premise smuggling (F2), silent problem reformulation (F3), and local-to-global compatibility gaps (F4). I then audit eight one-shot proofs generated by Gemini 2.5 Flash on Questions 1, 2, and 5 of the benchmark, using two instruments built specifically to surface F1 and F2. The central finding is uncomfortable for anyone who sees retrieval-augmented generation (RAG) as the obvious fix: not one of the eight proofs contained a confirmed fabricated citation, yet every single one contained at least one load-bearing claim asserted as a "fundamental result" or "standard argument" with no justification attached. That failure mode-F2, premise smuggling-is invisible to citation verification by design. A premise-audit instrument I introduce flags it at 100% precision (5/5 judge-confirmed flags are true positives) and 50% proof-level recall in this corpus. The taxonomy and the audit together suggest that the right long-term objective is building inference-time pipelines that prevent these failure modes from occurring, not just detecting them after the fact. Index Terms--Large language models, mathematical reasoning, hallucination, premise smuggling, failure-mode taxonomy. --- ### LLM Evolution as an Industry-Scale Ecosystem: A Lifecycle Perspective on Continual Learning Authors: Hao Jiang, Enneng Yang, Guojie Zhu, Yibin Chen, Yunkun Xu, Zifu Kou, Jiayi Li, Chong Chen, Zhao Cao, Li Shen Published: 2026-06-12 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2606.24901v1 Continual learning capability is critical for Industrial LLMs, as deployed models must be continuously updated to meet evolving requirements and environments, rather than repeatedly retrained from scratch. However, most existing research focuses on improvements on static benchmarks, failing to capture real industrial needs. In this survey, we reformulate Industrial Continual Learning (ICL) for LLMs as a closed-loop update-and-release problem in a versioned ecosystem, where updates propagate hierarchically to industrial, application-specific models and LLM-powered applications, with capability inheritance and transfer across versions and model families. From this ecosystem perspective, we identify three core challenges: repeated adaptation erodes model plasticity, foundation-model upgrades break capability inheritance, and long-term sustainability is constrained by deployment requirements. We then organize the technical landscape of ICL around five lifecycle design principles: preserving plasticity headroom, treating upgrades as capability transfer, enabling trustworthy continual reinforcement learning, making training recipes self-optimizing, and building accountability as a base layer for long-term iteration. For each principle, we synthesize representative technical directions. Finally, we evaluate the maturity of each principle and its technical components via an evidence-based lens, identify key gaps hindering real-world deployment, and outline a practical ICL deployment blueprint and a pathway for feeding industrial realities back into academic research. --- ### On-Device Neural Architecture Search Authors: Andrea Mattia Garavagno, Edoardo Ragusa, Paolo Gastaldo, Antonio Frisoli, Claudio Loconsole Published: 2026-06-12 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2606.24900v1 This paper proposes a new approach to near-sensor computing, in which a lightweight Neural Architecture Search (NAS) is performed directly on the deployment device to find the best tiny neural architecture for analyzing the real-time data acquired through sensors. This new adaptation capability can be particularly useful in the case of human-machine interfaces for which the neural network analyzing the biometrical data can be re-designed each time the user changes, after a guided data collection procedure, fighting the typical data variations between individuals on a new level. To implement the proposed approach a new NAS has been designed and then validated on the Italian Sign Language dataset (ISL), a collection of surface electromyography (sEMG) signals of the signs of the Italian alphabet, using several embedded systems. Moreover, further validation on the Case Western Reserve University dataset (CWRU), a benchmark for intelligent fault diagnosis, is presented to suggest another possible application of the proposed approach. When run on a Raspberry Pi 4, the proposed NAS performs beyond the state of the art proposing a tiny neural architecture having 0.63 times less RAM occupancy and 5.96 percentage points of more accuracy in the case of the ISL dataset; and 0.44 times less RAM occupancy and 0.2 percentage points of more accuracy in the case of the CWRU dataset. --- ### From Meta Idea to Advanced Mathematical Discovery -- Human-AI Co-Discovery of Sign-Embedding Quantum Algorithms Authors: Yanqiao Wang, Jin-Peng Liu, Peng Li, Yang Liu Published: 2026-06-12 Categories: cs.LG, cs.AI, quant-ph arXiv: https://arxiv.org/abs/2606.24899v1 AI-assisted mathematics is often evaluated on solving predefined problems. In practice, however, many important advances begin earlier, when a vague research intuition is transformed into a concrete problem, a promising route, and a theorem family worth proving. This report studies that stage through a case study that led to sign-embedding quantum algorithms for matrix equations and matrix functions, foundational primitives in quantum linear algebra and operator-output quantum algorithms. The project began with a human-originated intuition that rational approximation is especially effective for jump-type functions such as the sign function, and might therefore serve as a design principle for quantum algorithms. Rather than merely assisting after the problem was fixed, AI-assisted exploration, including workflows later integrated into the agentic AI-mathematician system AIM, played a key role in expanding this intuition into a route map, comparing candidate formulations, and converging toward sign embedding as the central framework. AIM then helped connect a known matrix-sign identity to wider classes of matrix equations and matrix functions, and drafted proof and complexity calculations. The decisive scientific judgments remained human: selecting which human-AI-expanded routes were worth pursuing, rejecting a Cayley-trapezoidal approximation when its validity required a hidden condition, and refining the Sylvester implementation from a coarse quadratic-gap query route to the final factorized and scaled analysis. The report argues that human-AI co-discovery workflows, with systems such as AIM as important components, are most valuable not as standalone theorem provers, but as research partners for problem formation, connection discovery, derivation, and skeptical review inside a human-gated research loop. --- ### Dense Supervision Is Not Enough: The Readout Blind Spot in Looped Language Models Authors: Rituraj Sharma, Tu Vu Published: 2026-06-12 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2606.24898v1 Looped language models turn hidden states into runtime state: each state is decoded for prediction and fed back into future computation. This creates a basic supervision question: which state variables does cross-entropy actually control? We show that dense per-loop cross-entropy controls the variables exposed by the readout, not every variable active in the recurrent transition. Hidden-state scale gives a concrete failure mode. Scale-invariant readouts such as RMSNorm and LayerNorm hide radial scale from the immediate cross-entropy loss, while pre-norm residual recurrence continues to carry and update that same scale. Thus per-loop loss can make early exits usable without controlling recurrent scale. In 44M and 129M looped transformers without inter-loop normalization, per-loop cross-entropy through RMSNorm readouts still drives final hidden-state norms into the thousands or tens of thousands. Scale-visible readouts and explicit norm penalties keep norms in the tens, and scale-removing recurrence is the complementary architectural fix. The resulting design rule is simple: dense supervision trains exits; recurrent scale control requires either making scale visible to a loss or removing it from the loop. Consistent with this rule, scale-controlled variants achieve lower perplexity at matched inference-depth operating points in our variable-depth benchmarks. --- ### Coordinate-Queryable Neural Field Reconstruction for EEG Spatial Super-Resolution with Unseen-Electrode Generation Authors: Hongjun Liu, Leyu Zhou, Zijianghao Yang, Chao Yao Published: 2026-06-12 Categories: eess.SP, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2606.23707v1 EEG spatial super-resolution (EEGSR) in real deployments is challenged by random channel missingness, unstable electrode quality, and changing visible-channel patterns caused by bad contacts or device variability. Most existing EEGSR methods learn a fixed low-to-high channel mapping under pre-defined input-output layouts, which makes them brittle when missing channels vary at test time. In this paper, we reformulate EEGSR as learning a shared conditional scalp field from partially observed support channels. Specifically, a position-guided encoder summarizes the observed EEG channels and their coordinates into a latent condition, and a conditional implicit neural representation decoder reconstructs target EEG signals by querying this condition at desired electrode coordinates. During inference, the model directly reconstructs unseen electrode signals from the available EEG support and the queried coordinates. To strengthen the constraint of the encoded latent representation on the decoder and thereby construct a more stable scalp field consistent with the observed channels, we further introduce a fidelity-preserving channel corruption training strategy under mixed electrode states. Extensive experiments across multiple EEG datasets demonstrate the effectiveness of our framework for both random missing-channel reconstruction and strict unseen-electrode signal generation. Notably, under the strict held-out-electrode setting on AAD, our method reduces NMSE by 37.5\% and improves SNR by 2.12 dB over the strongest baseline, showing its ability to synthesize signals at electrode locations never exposed during training. --- ### Zero-Shot Neural Priors for Generalizable Cross-Subject and Cross-Task EEG Decoding Authors: Baimam Boukar Jean Jacques, Brandone Fonya, Nchofon Tagha Ghogomu, Pauline Nyaboe, Kipngeno Koech Published: 2026-06-12 Categories: eess.SP, cs.HC, cs.LG arXiv: https://arxiv.org/abs/2606.23706v2 The development of generalizable electroencephalography (EEG) decoding models is essential for robust brain-computer interfaces (BCI) and objective neural biomarkers in mental health. Conventional approaches have been hindered by poor cross-subject and cross-task generalization, owing to high inter-subject variability and non-stationary neural signals. We address this challenge with a zero-shot cross-subject decoding framework on the large-scale Healthy Brain Network dataset, benchmarking a convolutional neural network baseline, a hybrid LSTM, and a Transformer-based foundation model. To adapt the Transformer for regression while averting catastrophic forgetting, we propose a novel progressive unfreezing strategy. The baseline yielded an nRMSE of 0.9991, whereas our fine-tuned Transformer achieved 0.9799 on unseen subjects. This work advances scalable, calibration-free EEG decoding for computational psychiatry and behavioral prediction. --- ### Beyond Importance: Interchange-Sobol Sensitivity Reveals Task-Specific Content Channels in Transformer Components Authors: Yifeng Guo, Jin-Hong Du, Xiang Chen Published: 2026-06-12 Categories: stat.ML, cs.LG, stat.ME arXiv: https://arxiv.org/abs/2606.20678v1 Mechanistic interpretability methods summarize a transformer component by a single importance score, conflating two distinct roles: a component may matter because it transports task-relevant content, or because the forward computation degrades when its contribution is removed. We introduce \emph{Interchange-Group Sobol Decomposition} (IGSD), a paired-intervention framework that compares matched activation replacement with zero ablation on the same component, estimates two Sobol-style variance indices, and uses their signed difference to separate the two roles, with intervention validity monitored by a symmetric off-manifold diagnostic $\widehat{\mathrm{ST}}>1$. In factual recall, IGSD identifies an early-layer content channel in both GPT-2 small and Qwen2.5-1.5B that standard importance methods underestimate. A controlled subject and relation donor design shows that the early channel transports relation-frame content while late attention transports subject-retrieval content, refining at head granularity to the known $\mathrm{Attn}_{L9H8}$ head. Late-layer clamping confirms that the early signal is expressed through downstream transformations rather than residual pass-through. These results show that replacement and deletion are not interchangeable controls and their divergence provides a practical statistical diagnostic for content transport in transformer components. --- ### Democratizing and accelerating AI-driven pathology research through agentic intelligence Authors: Jiabo Ma, Cheng Jin, Yihui Wang, Hao Jiang, Ling Liang, Yingxue Xu, Junlin Hou, Zhengrui Guo, Zhengyu Zhang, Yifei Xia, Hongyi Wang, Fengtao Zhou et al. Published: 2026-06-12 Categories: cs.AI, cs.CV arXiv: https://arxiv.org/abs/2606.20677v1 Computational pathology has advanced rapidly with the emergence of foundation models, yet widespread adoption remains limited by substantial technical complexity and programming requirements. Here we present PathLab, an autonomous agentic framework that translates natural-language research objectives into executable and validated computational pathology workflows through the structured composition of domain-specific skills and tools. By organizing workflow generation around reusable methodological modules, including data preprocessing, model development, evaluation and interpretation, PathLab enables studies to be specified at the level of scientific intent rather than implementation details. We evaluated PathLab across 12 public datasets spanning four representative task families: region-of-interest classification, whole-slide image classification, segmentation and survival prediction. Across all task categories, PathLab achieved non-inferior performance relative to expert implementations, while consistently enforcing semantic validity of user prompts and proactively rejecting incompatible workflow specifications prior to execution. In controlled user studies, PathLab substantially reduced the time required to generate executable analytical pipelines and enabled domain experts without programming experience to independently design, execute and evaluate computational pathology studies. Together, these results establish PathLab as a reliable interface between biomedical intent and computational execution, enabling computational pathology studies to be designed at the level of scientific questions rather than programming expertise. By lowering technical barriers to advanced AI methodologies, PathLab provides a foundation for the broader democratization of computational pathology. --- ### Jury Duty: Calibration and Orientation Failures in MLLM-as-a-Judge Under Cultural Ambiguity Authors: Daniel Lee, Harsh Sharma, Eunkyu Park, Pranav Narayanan Venkit, Jeonghwan Kim, Kah Mun Chia, Andreas Vlachos, Shafiq Joty Published: 2026-06-12 Categories: cs.CV, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2606.20676v1 MLLM-as-a-Judge is conventionally validated by agreement with human annotations, but this metric is undefined when the human pool is culturally heterogeneous. We introduce VOIR DIRE, a multimodal benchmark of 626 culturally paired image--prompt artifacts spanning U.S. and mainland Chinese contexts across food, fashion, and architecture, with annotator pools that are within-pool reliable (a = 0.86/0.74) but cross-pool divergent on evaluation (Q1 r = -0.12). Across six MLLMs, the bias decomposes into two failures: a positivity-floor calibration failure (compressed scale use) and an orientation failure (default to one cultural norm). On this corpus, where contested items are sampled to split the two pools, the floor mechanically validates the more-permissive Chinese reading; persona prompting partially recovers calibration, but the orientation residual survives, evidence the tilt is not reducible to scale compression. Reference-pool in-context demonstrations deepen the orientation residual and inflate the high end rather than restoring use of the low end. Model origin adds a small additive tilt (~0.10 MAE) that is approximately invariant under demonstration. We recommend reporting alignment against each reference pool separately and treating cross-pool divergence as a judge property. --- ### VQ4SNN: Vector Quantization for Memory-Efficient FPGA Spiking Neural Networks Authors: Dimitrios Sekertzis, Giorgos Dimitrakopoulos Published: 2026-06-12 Categories: cs.NE, cs.AI, cs.AR arXiv: https://arxiv.org/abs/2606.20675v1 Spiking Neural Networks (SNNs) offer an energy-efficient paradigm for edge AI, making them attractive for hardware acceleration. However, deploying dense SNNs on FPGAs is constrained by limited on-chip memory for synaptic weight storage. To address this bottleneck, we propose VQ4SNN, a hardware-aware architecture that reduces memory requirements through Vector Quantization (VQ)-based weight sharing. To the best of our knowledge, this is the first application of VQ to pipelined spatial-dataflow SNN accelerators on FPGAs. VQ4SNN replaces conventional weight storage with a two-level memory organization consisting of compact pointers and a shared codebook of quantized weight vectors. The proposed design integrates FPGA-aware memory mapping with analytical VQ parameter selection, enabling efficient deployment on such accelerators while preserving inference accuracy. The experimental results show a reduction of 52-61% in the total number of BRAMs compared to the state-of-the-art uncompressed FPGA SNNs without increasing overall logic utilization. --- ### A Formal Tool for Verification of Probabilistic Spiking Neural Networks Based on Quotient Abstractions Authors: Nikan Zandian Jazi, Elisabetta De Maria, Christopher Leturc Published: 2026-06-12 Categories: cs.NE, cs.AI, cs.FL arXiv: https://arxiv.org/abs/2606.20674v1 Spiking Neural Networks (SNNs) model biological neural dynamics more faithfully than classical artificial networks, but their stochastic, event-driven computation -- rooted in ion-channel noise and unreliable synaptic vesicle release -- demands probabilistic models for which deterministic abstractions are mathematically inadequate. Formal verification of such models via probabilistic model checking faces a fundamental barrier: the state space explosion problem, where the Discrete-Time Markov Chain (DTMC) encoding grows exponentially with the number of neurons. General-purpose quotient model abstractions (1) can in principle mitigate this growth by partitioning membrane potentials into equivalence classes, but a naïve application to SNNs discards synaptic weight information, limiting the properties that can be verified. This paper introduces a weight-discretized quotient model abstraction that maps continuous synaptic weights to a compact integer range while preserving the relative contribution of each synapse, and presents CogSpike, a unified workbench that integrates SNN design, simulation, and PRISM-based formal verification within a single isomorphic tool chain. The discretization is accompanied by formal correctness guarantees: a two-sided fidelity theorem confines any firing disagreement to a bounded gray zone around threshold, and an Asymptotic Silence theorem gives the exact limit guarantee that unforced neurons fall permanently silent. A topology-dependent scaling analysis shows that the state space reduction compounds exponentially -- approximately $17\times$ per neuron for discretization parameter $W = 3$ -- enabling verification of networks that are otherwise intractable, as confirmed empirically across seven canonical topologies. --- ### NeuroShield: A Device-Agnostic Foundation Model for EEG Authentication Authors: Matin Fallahi, Patricia Arias-Cabarcos, Thorsten Strufe Published: 2026-06-12 Categories: cs.LG, cs.AI, cs.CV arXiv: https://arxiv.org/abs/2606.20673v2 A central challenge in EEG authentication is that models are typically tied to the acquisition settings in which they are trained. In particular, variations in headset hardware, channel layout, and signal duration create heterogeneous recordings that existing models are not designed to handle, causing each new headset or dataset to be treated as a separate model-development problem. This fragmentation limits multi-dataset learning, hinders knowledge transfer, and reduces model reusability. To address this limitation, we present NeuroShield, a reusable foundation model for EEG authentication that learns identity-discriminative embeddings from variable-channel and variable-length EEG recordings through a dual-stage transformer architecture. We pretrain NeuroShield on three public EEG datasets comprising 15{,}762 subjects and 28{,}116 sessions, and evaluate transfer on two unseen downstream datasets. Our evaluations show that, after fine-tuning, NeuroShield reduces equal error rate by 0.44--8.06 percentage points relative to the state of the art. NeuroShield further generalizes to segments longer than those seen during training and operates across channel layouts not encountered during pretraining. These results establish NeuroShield as a reusable and adaptable EEG identity encoder across heterogeneous recording settings. We release NeuroShield as open source to support reproducibility and community adoption. --- ### Genetic Algorithm Based Coordination and Optimization Model for Generation Grid Load Storage in Active Distribution Networks Authors: Jinlu Zhang, Fujian Chi, Tianhan Ling, Yulong He, Kejia Zhang, Hongxing Lv, Sheng Wang Published: 2026-06-12 Categories: cs.NE, cs.AI arXiv: https://arxiv.org/abs/2606.20672v1 Create an optimization framework that combines fuzzy logic and genetic algorithms for risk assessment and coordination of generation, grid connection, load, and energy storage facilities in active distribution networks. In order to capture the system uncertainties caused by weather factors and dynamic user demands, this project uses fuzzy set representation to model renewable energy generation, load curves, and market prices. In the evolutionary process, genetic algorithms introduce fuzzy elements into system management, adjusting thru penalty factors to improve stability and seeking better scheduling or power dispatch solutions. At the same time, by penalizing uncertainty and constraint violations, the fitness function of the hybrid model aims to optimize the expected operational costs; it can produce feasible and economical results under unknown parameter conditions. Under conditions with a large supply of renewable energy and installed energy storage systems, simulation results for the IEEE-69 power system indicate that the fuzzy genetic algorithm strategy effectively reduces technical constraints compared to deterministic optimization schemes. The total investment remains at a similar level. Other data indicate that thru fuzzy reasoning optimization, unreasonable choices or impossible situations are avoided during the network adaptation process. This framework-based technical approach helps to construct an effective computational scheme that is aware of uncertainty in its outputs. This will provide a scientific basis for the uncertainty assessment in distribution network planning. --- ### A Projection-Based Surrogate Gradient Interpretation for Neural Codec Wrappers Authors: Esteban Pesnel, Julien Le Tanou, Michael Ropert, Aline Roumy, Thomas Maugey Published: 2026-06-12 Categories: cs.CV, cs.AI, eess.SP arXiv: https://arxiv.org/abs/2606.20671v1 Neural wrappers are learned pre-and postprocessing networks designed to enhance the performance of conventional video codecs. Although these approaches can significantly improve compression efficiency, training them remains challenging due to the non-differentiability of video codecs, which arises from the multiple discrete decisions involved in the encoding process. Surrogate gradients have recently emerged as an effective solution for enabling end-to-end learning with conventional codecs. They offer two main advantages: they avoid training an additional network to mimic the codec, and they can improve compression performance. In particular, the recently proposed SCALED method, which leverages the true compression error, has shown strong results for training neural pre-processors such as downscalers. However, this SCALED gradient was originally introduced as a reparameterization trick, which limits its interpretability. In this paper, we show that this surrogate gradient can be interpreted as a first-order local approximation of the video codec, providing insight into its effectiveness. We further demonstrate that it is effective not only for learning downscaling operations, but also for the more challenging task of full neural wrapping with pre-and post-processing networks. Finally, we show that the approach generalizes well across different video codecs, quality factors, and tasks, including multiple downscaling ratios, yielding BD-Rate (PSNR) reductions of up to -23.59% on x264 and -20.07% on VVenC relative to standard resampling baselines. --- ### Towards CSI-Native Foundation Models: A Channel-Adaptive Roadmap for 6G Authors: Chenyu Zhang, Xinchen Lyu, Chenshan Ren, Shuhan Liu, Qimei Cui Published: 2026-06-12 Categories: cs.LG, cs.AI, cs.IT arXiv: https://arxiv.org/abs/2606.20670v1 Wireless foundation models offer a path toward reusable channel state information (CSI) intelligence for sixth-generation (6G) systems. However, existing generic-backbone adaptation and CSI pretraining methods often treat CSI as task tensors rather than propagation-conditioned channel responses, thereby failing to capture the intrinsic time-frequency-spatial geometry of wireless environments. This paper presents a channel-adaptive roadmap toward CSI-native foundation models, proposing a unified framework that aligns pretraining, positional modeling, and attention control with three channel requirements: scale-aware heterogeneous exposure, physical time-frequency-antenna coordinates, and correlation-bounded token interaction. Extensive experiments demonstrate the superiority of the proposed framework across three dimensions: zero-shot generalization, reducing NMSE by more than 4 dB across spatial-temporal-frequency tasks; scale extrapolation, yielding up to a 5.4 dB gain under 8 times unseen antenna scaling; and inference efficiency, accelerating mobility-aware processing by up to 18.8%. A system-level evaluation with Sionna SYS further shows that the proposed framework uses only 7.01% of dense-pilot overhead, reaches -18.64 dB average NMSE, and improves average net spectral efficiency by 36.6% over dense LMMSE and 15.5% over WiFo, indicating that CSI-native representation learning can support pilot-efficient radio access. --- ### Agent Behavior Mining: Generative AI Agent Governance in Business Processes Authors: Hoang Vu, Maximilian Körner, Adrian Rebmann, Gabriel Kevorkian, Michael Perscheid, Gregor Berg, Timotheus Kampik Published: 2026-06-12 Categories: cs.AI, cs.SE arXiv: https://arxiv.org/abs/2606.20669v1 As organizations increasingly deploy generative AI agents to automate business processes, they face a governance dilemma: although these agents can increase operational flexibility, their non-deterministic nature challenges the control and standardization that Business Process Management seeks to enforce. This paper addresses this \emph{invisible autonomy risk} by introducing \emph{Agent Behavior Mining}, a governance capability that enables the application of process mining techniques to render generative AI agent decision-making observable and traceable. We (1) improve the understanding of generative AI agent behavior through an event data model that translates granular agent activities -- including reasoning traces, tool usage, and token costs -- into standardized process logs; (2) instantiate the data model in a multi-agent order-to-cash implementation, demonstrating how process managers can leverage agent logs to detect policy deviations and quantify operational variability; and (3) evaluate the perceived practical utility of the approach in an exploratory study with 18 industry practitioners. The results indicate that practitioners view behavioral transparency as a prerequisite for trust and consider the ability to examine agent reasoning as an important governance requirement for the next generation of AI-driven business processes. --- ### BELLS-O: Evaluating the Operational Trade-offs of LLM Supervision Systems Authors: Leonhard Waibl, Felix Michalak, Hadrien Mariaccia Published: 2026-06-12 Categories: cs.CR, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2606.20668v1 LLM supervision systems, namely input/output moderation filters and jailbreak detectors, are the primary safeguard against misuse in deployed AI applications, yet existing benchmarks are often vendor-biased, omit cost and latency, and rarely compare specialized guardrails against repurposed generalist LLMs. We present BELLS-O (Benchmark for the Evaluation of LLM Supervision Systems, Operational), the first independent operational benchmark of LLM supervision systems. BELLS-O evaluates 28 systems from 17 providers: every major specialized guardrail (e.g., LlamaGuard-4, ShieldGemma-2, Lakera Guard) and frontier generalists repurposed as supervisors (e.g., GPT-5.4, Claude Sonnet 4.6, Grok-4.1), jointly on detection rate, false-positive rate, latency, and monetary cost. We cover input/output moderation across 11 harm categories and jailbreak detection across 13 attack techniques, using in-house datasets built from handcrafted prompts, expert-curated samples, and quality-controlled synthetic generation. To suppress latent generator fingerprints in synthetic data, every generated sample is paraphrased. Mapping the Pareto frontier reveals use-case-dependent tradeoffs. On content moderation, specialized supervisors are operationally dominant: top systems match frontier LLMs on detection (~95% vs. 94%) at comparably low false-positive rates (<=2%), while running 5-10x faster and ~10x cheaper. On jailbreak detection, the tradeoff shifts: frontier LLMs achieve higher detection and lower false-positive rates but at 10-50x higher cost and 5-10x higher latency. We release the benchmark, framework, leaderboard, and datasets as the first vendor-neutral basis for selecting safeguards under real deployment constraints. --- ### A Quantum-Assisted Agentic Distributed Artificial Intelligence Framework for Deadline-Bounded Orchestration of Hybrid Renewable Microgrids Authors: Iacovos I. Ioannou, Saher Javaid, Minella Bezha, Yasuo Tan, Naoto Nagaoka, Vasos Vassiliou Published: 2026-06-12 Categories: cs.AI arXiv: https://arxiv.org/abs/2606.20667v1 The real-time orchestration of microgrids that combine fluctuating renewable sources, dispatchable units, storage and curtailable consumers requires the repeated solution of combinatorial dispatch and coalition formation problems under hard control deadlines. In this paper, a quantum-assisted agentic distributed artificial intelligence (DAI) framework is proposed in which the dispatch problem of each control slot is formulated as a quadratic unconstrained binary optimization (QUBO) problem by Belief-Desire-Intention extended (BDIx) agents and is solved by a portfolio of quantum, quantum-inspired and classical solvers. Solver selection is elevated to a first-class agentic deliberation action of the coordinator agent. Learned beliefs about solver latencies are maintained and the solver intention that is expected to satisfy the prevailing deliberation deadline is committed in each slot. In addition, a belief-shaped storage valuation mechanism is introduced through which the storage agent prices its energy at a discounted future-peak value, injecting intertemporal information into the otherwise myopic per-slot optimization. The framework is evaluated on a 24-hour simulation of a grid-connected microgrid with photovoltaic, wind, battery, genset and demand-response assets, with the Quantum Approximate Optimization Algorithm (QAOA) executed by statevector simulation and benchmarked per slot against tabu search, simulated annealing, binary particle swarm optimization, greedy descent and exhaustive enumeration. Zero deliberation deadlines are missed, the committed dispatch attains the exact optimum on every slot and the realized daily cost of 146.24 EUR equals the exact lower bound, with 97.83 percent renewable utilization and zero unserved energy. When the storage valuation mechanism is deactivated, the daily cost is increased to 152.75 EUR, a 4.5 percent increase. --- ### How Linear Is a Transformer Feed-Forward Block? Per-Block Linear Recoverability Is Learned, Not Architectural Authors: Stuart Whipp Published: 2026-06-12 Categories: cs.LG, cs.AI, cs.CL arXiv: https://arxiv.org/abs/2606.19379v1 Transformer feed-forward networks (FFNs) are often treated as nonlinear stores of computation, yet how nonlinear a trained FFN block actually is has rarely been measured. We treat each FFN as a position-wise input-to-output map and split it into the exact least-squares linear approximation plus a residual. The held-out variance the closed-form linear map explains defines a block's linear recoverability (R^2_lin), an optimiser-free measure of its linearity. Across all twelve blocks of GPT-2, Pythia-160m, and llama-160m, R^2_lin is highly heterogeneous and non-monotone with depth, ranging from near-linear (>0.99) to strongly nonlinear (<0.3) between adjacent blocks, and is not set by the activation function: same-width GELU models GPT-2 and Pythia-160m have sharply different profiles, so recoverability is a learned property of individual trained blocks, not an architectural one. A low-rank bilinear probe of the residual recovers only a few points of R^2, with gain uncorrelated with residual nonlinearity: the unrecovered computation is not a single position-wise product but higher-order or distributed structure. The measurement also serves as a targeted compression signal: recoverable blocks admit large single-layer replacements (GPT-2's early FFN at 8x fewer parameters for +0.77 perplexity), while low-recoverability blocks flag where this is unsafe. It further exposes a methodological pitfall: trained linear baselines can badly under-converge on ill-conditioned transformer activations, so we report the exact closed-form least-squares ceiling throughout. --- ### A Hybrid GNN-FEM Framework for Phase-Field Fracture Simulation. Physics-Preserving Hybridization for Generalizable Surrogate Modeling Authors: Hyeonbin Moon, Yongjin Choi, Seunghwa Ryu Published: 2026-06-12 Categories: cs.LG, cond-mat.mtrl-sci arXiv: https://arxiv.org/abs/2606.19378v1 Scientific machine learning (SciML) has emerged as a promising approach for accelerating simulations of complex physical systems, yet achieving physically consistent and generalizable predictions for nonlinear, history-dependent problems remains a central challenge. In this study, we propose a hybrid GNN--FEM framework for efficient and generalizable phase-field fracture modeling. While phase-field approaches provide a robust variational framework for simulating complex crack evolution, their high computational cost limits practical applications because they require solving coupled, nonlinear, and history-dependent systems within an incremental finite element procedure. To address this challenge, a graph neural network surrogate is integrated into the conventional staggered scheme, replacing the phase-field update at each load increment while retaining the FEM-based displacement solver to enforce mechanical equilibrium and boundary conditions. By preserving the incremental solution structure, the framework remains consistent with history-dependent fracture evolution without requiring the surrogate to approximate the full solution trajectory. This selective surrogate strategy emphasizes the identification of a physically meaningful and incrementally structured learning target, rather than relying on brute-force data generation to learn the full fracture process. The proposed framework achieves strong generalization across varying geometries, loading conditions, material properties, and discretizations through dimensionless feature design, a graph-based formulation on mesh-based domains, and a physics-informed loss derived from the governing phase-field equation. Numerical experiments demonstrate that the hybrid approach reduces computational cost while maintaining accuracy compared with conventional FEM, and exhibits robust predictive performance across diverse problem settings. --- ### Emyx: Fast and efficient all-atom protein generation Authors: Nicholas J. Williams, Ward Haddadin, Matteo P. Ferla, Constantin Schneider, Nicholas B. Woodall, Ruby Sedgwick, Christian D. Madsen, Andrew L. Hopkins, Edward O. Pyzer-Knapp Published: 2026-06-12 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2606.19377v1 Computational enzyme design requires generating proteins that scaffold catalytic residues and ligands, a task that demands both geometric accuracy and structural diversity from the underlying generative model. Current all-atom generators inherit expensive architectures from structure prediction, leading to high training costs and limited sample diversity. We argue that much of this complexity is unnecessary for generators, which condition on sparse geometric constraints rather than rich co-evolutionary signals. Emyx is a 140M-parameter conditional flow matching model that concentrates capacity within standard transformer blocks, replacing heavy embedding stacks with lightweight conditional representations and sparse connectivity. We additionally derive an exact reparametrisation of the flow matching interpolant into the EDM noise-level framework, bridging flow matching training efficiency with state-of-the-art sampling methods designed for diffusion models without retraining. Despite being the smallest model, Emyx outperforms both Proteína-Complexa and RFdiffusion3 against the AME enzyme design benchmark across success rate under strict evaluation requiring both global fold recovery and catalytic geometry accuracy, structural novelty, scaffold diversity, and geometric validity, while training in just $682$ GPU-hours, roughly $4\times$ less than RFdiffusion3. --- ### Cost-Optimal LLM Routing with Limited User Feedback under User Satisfaction Guarantees Authors: Herbert Woisetschläger, Arastun Mammadli, Ryan Zhang, Shiqiang Wang Published: 2026-06-12 Categories: cs.LG, cs.AI, cs.IR arXiv: https://arxiv.org/abs/2606.19376v1 Inference costs for large language model (LLM) applications are rapidly growing, driven by surging demand and rising infrastructure cost. Users expect high-quality responses, and in commercial settings this is formally codified in Service Level Agreements (SLAs), creating a fundamental tension between cost and quality. Recent progress on cost-aware LLM request routing has shown potential to resolve this tension, but existing approaches rely on complete feedback signals, offline training, extensive per-workload tuning, and most lack SLA guarantees or inference-time adaptivity. We introduce SLARouter, an online routing algorithm that learns a cost-optimal policy from the sparse, one-sided user feedback available in production systems. SLARouter provides theoretical guarantees for both cost optimality and strict SLA compliance. Experiments across a wide range of LLM benchmarks show that SLARouter satisfies SLA constraints without the need for per-benchmark tuning, reducing operating cost by up to 2.2x over existing baselines. --- ### Physics-Informed Discovery of Yield Functions in Plasticity via Convex Neural Representations Authors: Hyeonbin Moon, Donghyuk Cho, Jecheon Yu, Jeong Whan Yoon, Seunghwa Ryu Published: 2026-06-12 Categories: cs.LG, cond-mat.mtrl-sci arXiv: https://arxiv.org/abs/2606.19375v1 Identifying anisotropic yield functions remains challenging since yielding is not directly observed in full-field mechanical measurements, directional calibration can require many loading directions, and selecting an appropriate analytical form is nontrivial. This study proposes a physics-informed framework for discovering yield functions from full-field displacement data and reaction force data, without stress observations, plastic strain measurements, direct yield surface data, or a prescribed parametric yield function. The framework identifies the yield function as a mechanically constrained constitutive component inside elastoplastic stress integration, rather than through direct stress-space supervision. The yield function is represented by a convex neural network that enforces convexity and positive homogeneity of degree one while imposing the assumed tension-compression symmetry, and this neural yield function is trained with a differentiable stress update and a physics-informed force equilibrium loss across multiple loading cases. The proposed framework is validated using finite element (FE) benchmark studies with von Mises, Hill 1948, and Yld2000-2d yield functions, assessing yield contour agreement, displacement-noise sensitivity, identifiability through plastically active stress states, epistemic uncertainty, and polynomial-surrogate deployment. This study provides a mechanics-constrained pathway for discovering anisotropic yield functions from displacement and force data while keeping the identified component within the structure of elastoplastic stress integration. --- ### Protein Representation Learning with Secondary-Structure and Energy-Filtered Hydrogen-Bond Graphs Authors: Mohamed Mouhajir, Limei Wang, El Houcine Bergou, Hajar El Hammouti, Lamiae Azizi, Dongqi Fu Published: 2026-06-12 Categories: cs.LG, cs.AI arXiv: https://arxiv.org/abs/2606.19374v1 Graph-based representations are widely used in protein modeling, yet many existing approaches rely primarily on sequence adjacency or geometric proximity, which only partially reflect the principles governing protein folding. Proteins instead adopt complex three-dimensional conformations organized around secondary structure elements, such as $α$-helices and $β$-sheets, which encode recurring local motifs and stabilizing hydrogen-bond interactions. In this work, we introduce a secondary-structure-aware graph neural network for protein representation learning. Residue-level node representations are augmented with secondary structure assignments, and graph edges are constructed from hydrogen-bond interactions filtered by their energetic strength. This design enables the model to capture both local structural context and long-range couplings that are central to protein stability and function. We evaluate the proposed approach on commonly used protein benchmarks and observe consistent improvements over existing graph-based methods. In addition, the resulting graph representations offer enhanced biological interpretability, as the learned connectivity aligns with established structural motifs. These findings suggest that incorporating secondary structure and energy-filtered hydrogen-bond topology provides an effective inductive bias for protein representation learning. The code is released at https://github.com/mohamedmohamed2021/SSProNet --- ### A Knowledge Theory of Capital:The Value of Natural and Artificial Intelligence, Volume 1 Authors: Jeffrey Gardiner Published: 2026-06-12 Categories: econ.GN, cs.AI, econ.TH arXiv: https://arxiv.org/abs/2606.18288v2 This volume develops a knowledge theory of capital for economies in which productive capacity increasingly resides in software, data, models, routines, expertise, platforms, organizations, commons, and public epistemic infrastructure. Beginning from Adam Smith's theory of labour, stock, specialization, and market extent, it asks what changes when knowledge becomes stock-like, mobile across forms, scalable, governable, recombinable, and imperfectly visible in accounting. The book introduces knowledge-bearing stock as the central object and analyses how it is generated, converted into governable form, deployed, improved through feedback, enclosed or shared, measured, impaired, and used as input to future production. It distinguishes embodied, disembodied, institutionalized, commons, and public knowledge forms and develops concepts such as first conversion, cognitive enclosure, feedback capture, dark capital, and expected knowledge loss. The argument is conditional and testable: modern wealth depends not only on capital accumulation, but on how productive knowledge is governed. --- ### Directing Open-Ended Evolution in Artificial Life via Multi-Scale Path Divergence Authors: Mikhail Akhtyrchenko, Mikhail I. Katsnelson, Andrey Ustyuzhanin Published: 2026-06-12 Categories: cs.NE, cond-mat.stat-mech, nlin.CG arXiv: https://arxiv.org/abs/2606.17091v2 Open-ended evolution (OEE) in artificial life is typically driven by uninterpretable, black-box neural-network complexity metrics, leaving life-like systems disconnected from physical theories of complexity. We introduce MSPD (Multi-Scale Path Divergence, denoted $D_P$), a renormalization-group-inspired scalar that quantifies how the heterogeneity of a system's local transition laws is organized across temporal and spatial scales. MSPD is defined at the population level as a functional of the realised trajectory and is computed as a windowed finite-resolution estimator, with consistency between the two stated as a proposition. The metric is an explicit formula and plays a dual role: as a gradient-free fitness function and as a post-hoc analytical lens on any simulation that exposes local transition laws. On a Flow-Lenia substrate we establish three claims. (C1) Under fixed-context replay of the exact pathwise objective, MSPD-optimized parameters score higher than matched random parameters. (C2) Along optimized trajectories, states whose local transition laws are more heterogeneous yield larger future divergence than matched, less-heterogeneous states under exact-state stochastic continuations, so the metric tracks intrinsic dynamics rather than injected noise. (C5) Higher MSPD corresponds to stronger scale-dependent frustration --- larger differences between the dynamics expressed at different spatial extents --- linking MSPD to the frustration criterion of biological complexity in the sense of Vanchurin et al. All three transfer to Life-like cellular automata and Particle Life++, indicating that MSPD is not specific to a single substrate. A single explicit formula thus both \emph{directs} open-ended evolution and provides a principled bridge to the physics of complexity that black-box drivers do not. --- ### ANEForge: Python for direct computation on the Apple Neural Engine Authors: Spencer H. Bryngelson Published: 2026-06-12 Categories: cs.PL, cs.AI, cs.MS arXiv: https://arxiv.org/abs/2606.17090v1 ANEForge is a Python package that programs the Apple Neural Engine (ANE), the fixed-function neural accelerator on every recent Apple device, directly and without CoreML. In production the engine is reachable only through CoreML, which treats it as a scheduling option: no configuration requires the ANE, and a model can silently run on the CPU or GPU instead. ANEForge compiles a lazy tensor graph, built from 58 fused operators and 19 native bridge operators, into a single ANE program. The program is dispatched through the same ANE daemon and kernel-driver stack as Apple's internal framework. Beyond inference, the package reaches the engine's native fused attention, streams int8, int4, and sparse weights, keeps decoder and optimizer state resident across steps, and runs the forward pass, backward pass, and optimizer update of training on the engine. A small fused program completes a call in about 90us, near the engine's 70us per-program dispatch floor, and a pretrained ResNet-18 forward runs end-to-end in 0.33ms. ResNet-18, a sentence encoder, and a Vision Transformer run end-to-end against framework references, and a Stable Diffusion U-Net validates its forward pass. ANEForge targets Apple Silicon under macOS 14 and later. Each release is verified against a recorded macOS and ANE-compiler version. --- ### ZIVARI-TLBO: A Zero-Cost Inter-Group Evaluated-Elite Relay Mechanism for Teaching-Learning-Based Optimization Authors: Pezhman Zivari Published: 2026-06-12 Categories: cs.NE, cs.AI arXiv: https://arxiv.org/abs/2606.17087v1 ZIVARI-TLBO is a grouped Teaching-Learning-Based Optimization (TLBO) method that augments an existing population-state controller with a fixed inter-group evaluated-elite relay. At each scheduled event, every group offers its already evaluated elite to the next group in a fixed ring; the elite replaces the receiver's worst eligible learner only when its stored objective value is better. Because the exact relay copies an already evaluated solution and its stored fitness, it requires no additional objective-function calls. The frozen gts-v4-cm-fixed implementation is evaluated under equal 10,000-evaluation budgets on eight classical functions at dimensions 10, 30, 50, and 100, with 30 matched seeds, and on five constrained engineering problems. A direct ablation against the same grouped landscape-aware controller without relay records 728/11/221 wins/ties/losses and a rank-biserial effect size of 0.624 across dimensions. In an eight-method multidimensional comparison, WOA obtains the best average rank (2.914) and ZIVARI-TLBO ranks second (3.382); ZIVARI-TLBO significantly outperforms TLBO, MCTLBO, DE, PSO, and GWO, loses significantly to WOA, and is not significantly different from HHO after Holm adjustment. Feasibility-aware engineering results are mixed and sensitive to the current static-penalty formulation. The evidence supports a scoped relay contribution and budget-consistent information-sharing mechanism, but not universal state-of-the-art, global-convergence, engineering-dominance, or CEC superiority claims. --- ### ParkingTransformer: LLM-Enhanced End-to-End Trajectory Planning for Autonomous Parking Authors: Hauteng Wu, Xu Li, Dong Kong, Zihang Wang, Xieyuanli Chen, Benwu Wang, Wenkai Zhu Published: 2026-06-12 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2606.17082v1 End-to-end autonomous parking has emerged as a critical task within the realm of autonomous driving. However, existing methods suffer from black-box characteristics, lacking high-level semantic understanding and interpretability, which impedes the realization of seamless long-distance autonomous parking from the road to the target spot. To address these limitations, we propose ParkingTransformer, a novel framework that leverages multi-view perception and the scene understanding capability of Large Language Models (LLMs). By combining trajectory queries with LLMs implicit state features, our method interacts directly with historical information and raw sensor data to output planning trajectories, eliminating the need for dense Bird's-View (BEV) representations. To compensate for the inadequate spatial reasoning ability of LLMs, we introduce 3D positional encoding to explicitly inject spatial geometric awareness. Furthermore, a fixed-window streaming mechanism is designed for historical information processing, significantly improving long-term temporal processing efficiency and inference speed. Additionally, a coarse-to-fine decoding strategy is employed to progressively enhance trajectory precision. Extensive closed-loop experiments are conducted on the CARLA simulator and real-world vehicle platforms. The results demonstrate that our method achieves a driving score of 61.32 in CARLA simulator and an average success rate of 88.70% in real-world experiments, validating the feasibility and effectiveness of the proposed algorithms. --- ### Metric Match: A Subset Selection Approach to Evaluating LLM Judge Reliability Authors: Alyssa Unell, Natalie Dullerud, Naomi Boneh, Meena Jagadeesan, Tatsu Hashimoto, Nigam Shah, Sanmi Koyejo Published: 2026-06-12 Categories: cs.AI arXiv: https://arxiv.org/abs/2606.15029v1 LLM judges are used to reduce the need for costly human labor in evaluating open-ended text generation. However, the reliability of these judges depends critically on their alignment with human raters -- a property that itself depends on costly human annotations. In this work, we develop a method (Metric Match) for estimating correlation-based reliability metrics of LLM judges from limited annotations. Metric Match selects a subset of samples for human annotation such that the subset matches the population reliability metric with respect to acquired synthetic labels. We empirically show that Metric Match achieves a win-rate of 0.838 against random subset selection across four different correlation metrics and 15 datasets, with an 18.7% decrease in average estimation error and reduces annotation needs by 32.5%. We provide a cost model and highlight a medical case study where our method saves $1,041.67 compared to random selection for expert annotation. Further, we shift our task from reliability estimation to reliability classification of whether a given judge is above a deployment threshold, outperforming random selection with Metric Match. All project code is publicly available, and we additionally provide an installable package for ease of use. --- ### Deep Temporal Modeling and Ensemble Fusion for Multimodal Emotion Recognition from Physiological Signals Authors: Desta Haileselassie Hagos, Saurav Keshari Aryal, Patrick Ymele-Leki, Anietie Andy, Legand L. Burge Published: 2026-06-12 Categories: cs.CL arXiv: https://arxiv.org/abs/2606.15026v1 Physiological stress and emotion recognition are important for health monitoring and affective computing. In this work, we present a comprehensive evaluation of deep learning models such as Long Short-Term Memory (LSTM), Temporal Convolutional Networks (TCN), and Transformer on the WESAD dataset for multimodal affect recognition using wrist and chest sensor signals. We perform ablation studies to assess the individual contributions of each modality by training models on wrist-only and chest-only inputs. In addition, we implement a late-fusion ensemble strategy that combines predictions from all three architectures trained on multimodal input. We also employ early fusion at the sensor level by concatenating wrist and chest signals before feeding them into each model. Our results show that Transformer models consistently achieve the highest accuracy in multimodal settings, while TCN models perform best in the wrist-only configuration. The ensemble method yields the highest overall accuracy (98.91 +/- 0.13%) and macro-F1 score (98.56 +/- 0.17%). These findings demonstrate the effectiveness of sensor fusion and ensemble-based fusion in developing robust systems for physiological emotion recognition. --- ### Resilient Consensus in Agentic AI Authors: Sribalaji C. Anand, George J. Pappas Published: 2026-06-12 Categories: cs.MA, cs.AI, eess.SY arXiv: https://arxiv.org/abs/2606.15024v1 Large language model (LLM) agents are increasingly deployed in multi-agent systems where they must coordinate and agree on shared decisions. We ask whether classical resilient consensus theory, developed for deterministic agents, transfers to LLM agents that may behave adversarially. Framing LLM agreement as a Byzantine consensus game, we run controlled experiments on complete and general communication graphs. We find that prompted LLM agents fail to reach agreement that is achievable in principle: consensus can fail even in settings where classical theory guarantees that a convergent algorithm exists, and this failure persists across temperatures and horizons. At the same time, wrapping the agents with classical resilient consensus filters improves agreement. The benefit of filtering depends on how much robustness the underlying topology already provides. Our results suggest that classical resilient consensus theory is a useful lens for the safety of agentic AI. --- ### Multiscale Hypersonic Boundary Layer Reconstruction via Spectral Binning and Subdomain-wise Conditional Diffusion Authors: Hojin Kim, Dibyajyoti Chakraborty, Takahiko Toki, Carlo Scalo, Romit Maulik Published: 2026-06-12 Categories: physics.flu-dyn, cs.LG arXiv: https://arxiv.org/abs/2606.15023v1 We propose a multiscale probabilistic reconstruction framework for hypersonic Couette flow, where near-wall states are inferred from limited top-wall observations using conditional diffusion model. The boundary layer is divided into overlapping wall-normal subdomains, and a single height- and Mach-conditioned Elucidating Diffusion Model (EDM) is trained jointly for M=6,7,8 to sample velocity, density, pressure, and temperature fields conditioned on a top-wall boundary slice. A soft overlap inpainting strategy assembles subdomain predictions into full-volume reconstructions while maintaining inter-subdomain continuity and small-scale variability. To improve the spectral fidelity of the generated fields, we introduce a novel bounded binned spectral power (BSP) loss that preserves high-wavenumber content while remaining numerically stable across the diffusion noise schedule. Validation against direct numerical simulation data shows that the model recovers instantaneous structures, spectra, statistical profiles, correlations, and wall quantities across all training Mach numbers, while providing spatially structured uncertainty estimates. The reconstructed Mach-conditioned profiles also collapse under the Trettel-Larsson transformation, indicating consistency with compressibility scaling. These results establish the domain decomposed conditional diffusion model with a bounded binned spectral loss as an effective probabilistic surrogate for near-wall reconstruction in hypersonic wall-bounded turbulence. --- ### Comparison Patrols on Drifting Orders: Certified Rank Maintenance, Evolving Planar Maxima, and Selection under Drifting Fitness Authors: Faruk Alpay, Levent Sarioglu Published: 2026-06-12 Categories: cs.DS, cs.NE arXiv: https://arxiv.org/abs/2606.15022v1 Rank-based selection in dynamic environments acts on order information that becomes stale while it is being used. Tournaments, elitism, truncation, and Pareto selection may therefore consume rankings that no longer match the current fitness order, while full re-evaluation competes with search for the same budget. This paper formulates the missing information layer as a data-structure problem. A hidden total order on $n$ items drifts by adjacent transpositions, while a maintainer receives one truthful pairwise comparison per step and must answer rank queries continuously. We introduce the comparison patrol, a constant-time maintained-order structure using $3n+O(1)$ words, one comparison per update, deterministic verification-age bounds, and per-item displacement certificates. We prove lower bounds showing that oblivious and location-oblivious maintainers incur expected Kendall error $Ω(\min(α,1)n)$, and show that the patrol operates at the same order. A bump invariant yields exact self-stabilization after drift-free corruption: if the maximum rank overstatement is $L$, recovery takes at most $L$ aligned cycles and cannot finish before $L-1$. This gives a deterministic shock-recovery calculus and a crossover with full rebuild near $L\approx \log_2 n$. The maintained order is then transferred to evolving planar maxima and to evolutionary selection rules, giving deterministic bounds for truncation, tournament, elitist, and two-objective Pareto decisions under drifting fitness. Experiments up to $n=65{,}536$ audit the certificates, recovery laws, equilibrium behavior, and equal-budget dynamic evolutionary loops, identifying when certified local rank maintenance outperforms global re-evaluation and when it should hand over. --- ### Towards Global AI-Driven Cervical Cancer Screening Authors: Thuy Nuong Tran, Ömer Sümer, Evangelia Christodoulou, Lennart Nauschütte, Simon Kalteis, Martin Paulikat, Esmira Pashayeva, Klara Steinheuer, Isabella Borges, Piotr Kalinowski, Hermann Bussmann, Sieng Sokmney et al. Published: 2026-06-12 Categories: cs.CV arXiv: https://arxiv.org/abs/2606.15019v1 The global elimination of cervical cancer is a key public health goal set by the World Health Organization (WHO), with screening programs reducing mortality by up to 80%. However, access to experts and biopsy services is limited in low- to middle-income countries (LMICs). Deep learning (DL)-based algorithms offer promising support for screening, but most existing approaches have been developed and validated on private datasets from single countries. We present the first DL-based approach to cervical cancer screening validated on data from multiple countries. Technically, we phrase the problem of detecting and classifying lesions in colposcopy images as a multi-task learning problem, in which we simultaneously perform image-level classification and lesion segmentation. Our model was trained on a private data set of acid stain colposcopy images with manually generated lesion segmentation masks and corresponding histopathological results, employing extensive data augmentation to address image variability. In an in-distribution validation with pathology results serving as ground truth, our algorithm outperformed medical experts (Balanced Accuracy: 0.68 vs 0.64) in CIN1- (Cervical intraepithelial neoplasia grade 1 or lower) versus CIN2+ (grade 2 or higher) classification. External validation on four colposcopy data sets from four countries featuring radical differences in prevalence and patient characteristics yielded superior performance of our method compared to baseline methods. Performance variability across countries was high with AUC values ranging from 0.54 - 0.80. Overall, algorithm performance varied with age, transformation zone (cervical area most prone to lesion development), presence of comorbidities and pathognomonic signs, with comorbidities having by far the largest negative effect. Future work should focus on improving model robustness and generalizability. --- ### Are Online Skill and Memory Modules Always Worth Their Tokens? A Budget-Constrained Study of Web Agents Authors: Sina Hajimiri, Masih Aminbeidokhti, Jose Dolz, Ismail Ben Ayed, Issam H. Laradji, Spandana Gella, Nicolas Gontier Published: 2026-06-12 Categories: cs.CL arXiv: https://arxiv.org/abs/2606.15017v1 Online web agents often augment a base actor with memory, workflow, or skill modules. These modules can improve performance, but they also consume test-time tokens, a cost rarely reported alongside the actor's inference cost. We study online augmentation, where this overhead is paid on every task, and re-evaluate its benefits under a fixed total inference budget. We compare AWM, ASI, and ReasoningBank with a token-matched vanilla baseline that uses the same budget for additional actor steps. Across three WebArena domains and three models, Gemini 3 Flash, GPT-5.4-mini, and Qwen 3.6-27B, the vanilla baseline matches or surpasses all three augmentation methods in aggregate success rate while often using fewer total tokens. We observe a similar trend on WorkArena-L1 with Qwen 3.6-27B, indicating that the effect extends to enterprise knowledge-work tasks. Our results suggest that skills and workflow memory can be useful in specific domains, but their apparent gains often vanish against a budget-matched actor. We further show that run-to-run variance materially affects outcomes and should be reported as a core evaluation criterion for online web agents. ---