Computer Science and Mathematics

Abstract: As humans and AI agents work together across long-horizon tasks and accumulate interaction history, how cognitive work is organized on one task can shape both their later performance and later capability. Within a Human–AI system, the next step may involve further reasoning, retrieval from prior experience, verification, a tool call, another model, human judgment, or stopping. The cognitive operation used also shapes who gets what experience: whether the human continues to practice a skill, whether the AI receives demonstrations or corrections, which failures are exposed for supervision, and which trajectories become reusable. The way cognitive work is organized today therefore helps construct the Human–AI system that will act tomorrow. This survey connects human cognition, Human–Automation and HCI, and modern AI agents to explain how present cognitive organization shapes future Human–AI capability. Human cognition research explains how limited attention, computation, and control are organized, and how experience can restructure later cognition. HumanAutomation and HCI research shows how functions, authority, and initiative are divided across people and machines, and how that division changes human readiness and skill. In modern AI agents, many of these cognitive operations are explicit and programmable, while interaction trajectories can be retained and transformed into memories, workflows, skills, and policies that shape later behavior. Read together, these literatures expose a causal sequence: organizing cognitive work determines the distribution of practice, supervision, feedback, and trajectories; that experience changes human skill and beliefs, machine memory and policy, and the coordination between them; and the updated joint system then organizes later work differently. We argue that Human–AI collaboration should therefore be studied as a process of joint cognitive development. The design problem is to organize current cognition while accounting for the Human–AI system that the current choices are shaping. The way humans and AI work together today changes what each can do tomorrow.

Abstract: How well can you predict production performance from development process data alone? Modern DevOps pipelines record 164 metrics across eight SDLC phases, yet no public dataset links these signals to runtime outcomes. We built PRESTO and a Gaussian copula-based synthetic data generator, released as a public reproducibility package (Zenodo DOI: 10.5281/zenodo.22005439). During this revision, an audit of the generator’s own code, prompted by peer review, found three implementation bugs (48 of 84 specified metric correlations were silently never applied; domain-specific calibration overrides were never wired in; one metric’s documented distribution family was silently substituted for a different one). We fixed all three, regenerated all three synthetic domains, and reran the full downstream pipeline; every synthetic-domain number below reflects the corrected code, and Section 3.1.3 discloses exactly what changed. Across three enterprise-domain synthetic datasets, the best model achieved R² = 0.11 to 0.49 using only SDLC process features; a target-correlation ablation shows this largely reflects recovery of the generator’s specified structure, so the synthetic domains alone are not evidence of genuine signal. Adding historical uptime features raised the primary domain’s R² to 0.91, and roughly 88% of that gain is autoregressive persistence rather than process signal. Tree-based ensembles remained the best untuned model on all three domains, though Gradient Boosting now collapses badly on the primary domain (R² = −1.49) while Random Forest does not; once linear models are properly tuned, a tuned model edges out Random Forest on that same domain, but by a margin smaller than the estimation noise, so this is a near-tie, not a linear-model win. Build and testing infrastructure availability were the strongest synthetic-domain predictors; build-phase metrics now outweigh code-quality metrics roughly 6.6:1 in aggregate importance, and testing metrics roughly 4.3:1, both smaller ratios than an earlier, bug-affected 11:1 figure; real-world TravisTorrent data is directionally consistent (Test-phase features rank #1 in 5 of 7 projects), but a robustness check excluding a definitionally-circular test-duration feature drops every project’s holdout R² to negative, so this should be read as a weak, feature-ranking-only signal, not confirmation. We introduce Phase-Aware Recursive Feature Elimination, which guarantees cross-phase coverage and, under the corrected data, avoids collapse on two of three domains, though it fails badly on the primary domain. Four independent real-world datasets, unaffected by the generator fixes since they use real external data, still with bootstrap confidence intervals: TravisTorrent (R² = 0.475 best of seven, median 0.005), Mozilla Perfherder (R² = 0.17 to 0.91, autoregressive persistence), GHALogs (R² ≈ 0.11, zero leakage risk, significant for all five models, n = 28,443), and SQuaD(R² = 0.402 defect-fix, our strongest point estimate but not significant; enriching with real static-analysis code-quality features raises it to 0.483 and narrows the CI to a near-miss [−0.037,0.738], still not significant; R² = 0.245 CVE-count, weaker but significant). GHALogs and SQuaD’s CVE-count result are what actually license a confirmatory claim: SDLC process signals carry genuine, transferable predictive value, though not uniformly across every real-world dataset examined.

Abstract: Stock movement prediction remains challenging because financial time series are nonstationary, noisy, and interdependent across heterogeneous information sources. Conventional multimodal forecasting methods often integrate price, text, and inter-firm relationships into a single representation, making it difficult to capture modality-pair-specific interactions and their varying relevance across market conditions. To address this issue, this study proposes Pairwise Cross-Modal Attention Fusion with Mixture-of-Experts (PaCMoE), which explicitly decomposes multimodal interactions into three modality pairs— price-text, price-graph, and text-graph pair. Each pair is modeled by an independent cross-modal attention expert, and a Mixture-of-Experts router adaptively integrates the resulting representations according to the input. On the ACL18 (StockNet) benchmark, PaCMoE is evaluated using chronological data splitting, validation-based model selection, and ten random seeds. The proposed method achieves an average accuracy of 60.69% and an MCC of 0.2200, outperforming the compared methods evaluated with multiple runs or random seeds in both metrics. These results demonstrate the effectiveness of explicitly modeling pairwise cross-modal interactions and adaptively integrating their contributions for robust stock movement prediction.

Abstract: Large language models (LLMs) suffer from hallucinations—confidently generating false information when uncertain. Conventional token prediction provides no dedicated output action for abstention: uncertainty can be represented only indirectly through the output distribution, while the model is nevertheless trained to select ordinary vocabulary tokens. We investigate whether an explicit abstention state, combined with scalable supervision for when abstention is appropriate, improves selective prediction. We add a single ABSTAIN token to the vocabulary and train models to predict it using corruption augmentation—a scalable data augmentation technique where corrupted inputs are mapped back to the abstain token. In a simple feedforward network tasked with single-token prediction, this approach reduced the hallucination rate on unseen data from 95% to 0% while maintaining perfect accuracy on known examples. The same principle also scaled to a real question-answering (QA) model: a distilled Bidirectional Encoder Representations from Transformers (BERT) model fine-tuned on SQuAD abstained on 89% (±7.8) of nonsense questions at corruption level p = 0.10, rising to 94.5% (±7.8) at p = 0.20, while retaining approximately 95% of its baseline accuracy at both levels; this abstention transferred to natural text corruption (OCR errors, encoding artifacts, character damage) without any noise-specific training, and mining the model's own confident errors provided complementary supervision, extending abstention to fluent-but-unanswerable questions that mechanical corruption misses; combining the two supervision sources yielded broader abstention coverage across all evaluated failure modes.

Abstract: Deep learning now supports many tasks in medical imaging, but most systems assume a fixed data source and a single training event. Clinical practice is not static as new scanners arrive, protocols change, patient populations shift, and new diagnostic questions appear. A model that trains once and stays frozen tends to lose accuracy over time, while a model that retrains on new data alone tends to forget what it already knew, a failure known as catastrophic forgetting. Continual learning studies how a model can learn from a stream of data and tasks while it keeps prior knowledge. This survey reviews continual learning in medical imaging with a focus on recent progress. We report a reproducible review methodology, formalize the three learning scenarios, and organize the methods into five families: regularization, replay, dynamic architectures, foundation-model adaptation, and hybrids. We give particular attention to the shift toward frozen foundation models with lightweight adapters. We review applications across tasks and modalities, provide a critical analysis of study quality that distinguishes peer-reviewed work from preprints, and connect methods to the current regulatory framework for model updates. We close with open problems, aiming to provide researchers and clinical teams with a clear and up-to-date reference and to set out what remains to be shown before such systems can be updated safely in practice.

Abstract: Metric positional encoding supplies physical distances to Transformers, but its contribution to segmentation performance and spacing robustness remains uncertain. We evaluated whether metric coordinates improve performance, produce a more favorable response to resampling, and affect predictions through supplied spacing. A factorial UNETR comparison combined token-index or millimeter-based coordinates with additive Fourier encoding or 3D RoPE. KiTS23 evaluation used three-fold cross-validation on 391 patients; 117 eligible volumes were resampled to through-plane spacings of 3, 4, and 5 mm. Controls altered supplied spacing independently of image content. The RoPE comparison was repeated after retraining on AMOS22. In KiTS23, metric variants responded to supplied spacing but did not yield a more favorable Surface Dice change under resampling (contrast: −0.0029; 95% CI: −0.0104 to 0.0043). Among 38 eligible AMOS22 cases, metric RoPE retained higher absolute scores at every tested spacing, although its advantage narrowed from 0.0512 at native spacing to 0.0216 at 5 mm. Its baseline-relative response was less favorable (contrast: −0.0241; 95% CI: −0.0296 to −0.0186). These findings support evaluating absolute scores alongside changes from baseline and controls of spacing sensitivity. A less favorable change can coexist with higher absolute performance, conditional on the trained models and retrospective perturbations.

Abstract: Hallucination detectors commonly read their evidence from natural-language inference (NLI) heads softmax-normalized over {entailment, neutral, contradiction}. When a detector’s only evidence features are the entailment probability T and the contradiction probability F of one such call, T + F ≤ 1 holds identically, so the state in which evidence both supports and refutes a claim has no coordinate of its own (Theorem 1). The bound concerns coordinates: it implies no detection ceiling and no error rate. We test a dual-NLI protocol that reads T and F from separate calls on 20 conflicting-evidence items and 90 controls in six families, with two public NLI models on CPU. Holistic scoring flags 4/20 conflicting items, 3/20 under a margin rule; with oracle span roles, evidence-decomposed scoring reaches 15/20 at 2/80 false positives under a model-free lexical gate. Controls delimit the result: the escape from the simplex follows from scoring two spans in separate calls, not from using two models; requiring the support channel active, not the diagonal T + F > 1, removes the pure-contradiction false positives; dropping the oracle roles keeps 15/20 at 4/90 false positives, whereas collapsing that gated pair to one difference score reaches at most 8/20 at the same budget in a retrospective interval search. Three further controls, two pre-registered, examine the stimuli: replacing the source description of the refuting spans takes the protocol to 20/20 and the holistic baseline to 9/20, and crossed templates show large lexical effects on both heads that are conditional on the reporting verb and do not decompose the whole-template contrasts they follow. The contribution is diagnostic. Code, stimuli, pre-registrations and scores are released.

Abstract: Morphogenetic System Theory (MST), proposed by Germano Resconi in 1998, is a generalized system model that integrates multidisciplinary knowledge and provides a unified mathematical framework for the recording, preservation, and reconstruction of forms. Although MST has been applied in risk analysis, system identification, and image recognition, its potential as a holistic meta-methodology for analyzing the evolution of socio-technical systems has not yet been fully revealed. This paper systemati-cally reviews the four-decade development of MST, distills its core philosophical propositions—the unity of subjective and objective understanding, the unity of analysis and synthesis, and the primacy of structure over individual values—and positions it as a holistic alternative to reductionist approaches. Systematic comparisons with seven related methods reveal MST’s unique capacity to capture both the formation and evolution of forms via the closed loop of write and read operations. Two case studies—learning community morphogenesis and legal morphogenesis—demonstrate MST’s applicability in socio-technical systems. The findings indicate that MST provides a formal, traceable, and scalable framework for quantitative morphological identification and evolution of socio-technical systems, with promise in education and institutional analysis. The paper also discusses the main limitations of MST and directions for future research.

Abstract: Labels are the bottleneck in underwater acoustic target recognition (UATR): hydrophones collect unlabeled audio essentially for free, while verified vessel labels remain scarce. Self-supervised learning (SSL) promises to convert the former into reusable representations, but its value for ship-radiated noise has been quantified only on single, leakage-prone corpora. We report a controlled three-corpus pilot: a ResNet-18 pretrained with SimCLR on 176,481 unlabeled log-mel segments pooled from the train splits of Oceanship, QiandaoEar22, and VTUAD, then evaluated on each corpus under linear probing and full fine-tuning against ImageNet and random initialization (15 corpus–initialization–protocol cells × three seeds: 45 runs). The result is predominantly negative at this budget. Under linear probing the SSL representation is significantly worse than ImageNet features on two of three corpora (VTUAD: −0.083 accuracy, p = 0.017; QiandaoEar22: −0.053, p = 0.012). Fine-tuning largely erases the differences within 30 epochs (paired accuracy gaps ≤ 0.039, mostly non-significant). One unestablished hint survives: QiandaoEar22 macro-F1 +0.0372 (p = 0.310). A diagnostic val–test gap on VTUAD, plus a recording-identity probe that reads recording ID near-perfectly from the SSL embeddings, shows the contrastive objective encodes recording identity, not vessel type. A normalization control rules out the main alternative explanation. We analyze why in-corpus contrastive pretraining under-delivers and derive requirements for the follow-on large-scale study. One limitation bounds all of this: pretraining was a single run (n = 1), so encoder-level variance is unmeasured, and all results are preliminary observations on one specific encoder, not general conclusions about SimCLR on this task.

Abstract: Machine learning-based network intrusion detection, the explanation of detector decisions, and alert triage have each developed as a largely separate line of research. As a result, a detector’s aggregate performance score still tells a security analyst little about how far to trust an individual alert: it does not reveal how reliable the detector is for each attack type, whether the explanation of a decision is faithful, or on what calibrated basis some alerts should be automated and others escalated. We propose a confusion-aware triage framework that addresses these three concerns within a single, unified approach. From a trained detector’s predictions, the framework estimates the reliability of each attack type, validates the explanations of its decisions, and routes each alert, by that reliability, into automatic handling, analyst assistance, or escalation, presenting each escalated alert as a compact decision card. We evaluate the framework on UNSW-NB15 and CIC-IDS2017 with six detector families over five seeds. Per-attack-type reliability is estimable, stable across training runs, and transfers to unseen traffic, whereas a detector’s confidence overstates its reliability on the rarest attacks; the explanations are faithful for seven of the eight routed attack types; and routing by reliability escalates 27.9% of alerts while catching 75.7% of the detector’s errors, and abstains where the detector is already reliable. Routing by reliability matches routing by confidence on error coverage while remaining calibrated, per-attack-type, and auditable. The framework is proposed as an analyst-facing decision aid, and its evaluation with security analysts is left to future work.

Abstract: Joint-Embedding Predictive Architectures (JEPAs) are emerging as the backbone for latent world models in robotics and autonomous driving, yet inject-ing domain knowledge (physics, kinematics, geometry) into these models consistently degrades performance—with no theoretical explanation. We resolve this paradox by proving, through six theorems, that prediction and representation learning are strictly decoupled: sequential prediction losses do not alter the optimal maximum-entropy embedding distribution. This yields a world-model design principle with a provable guarantee: encoder maximum-entropy, predictor dynamics, decoder physics—and explains why prior physics-informed JEPAs failed: they constrained the encoder, where constraints provably destroy the entropy guarantee, instead of the decoder, where they are provably benign. We validate this principle across two domains through a phased protocol covering measurement artifacts (Proposition 1: the N/K ≥ 5 reliability threshold), powered statistical testing (60 seeded runs across autonomous driving and bimanual robotic manipulation, paired Wilcoxon significance in every com-parison), and a budget replication demonstrating that the encoder-physics violation deepens with training. Phase 3 additionally characterizes an entropy–utility frontier on low-intrinsic-dimension robotics data, discovering that per-dimension variance matching alone produces correlated collapse—a failure mode invisible on high-dimensional data—and that optimization budget, not regularization weight, is the binding entropy constraint. To test whether the placement rule matters beyond representation quality, we further probe its consequence for downstream control: a latent cross-entropy-method (CEM) planner built on the trained Phase 3 models consistently outperforms real-action replay on model-internal cost across all three abla-tion conditions—an impossible result for genuine planning, since replaying the true actions is itself the ground-truth solution. A model-exploitation diagnostic traces this to the planner discovering action sequences ∼26% of the action range away from the true trajectory while still lowering the learned decoder’s cost, uniformly regardless of physics placement. We report this transparently as a boundary condition on the practical claim: offline latent planning against a learned decoder is not, by itself, suffi-cient evidence that a representation supports downstream control, and we identify closed-loop, simulator-verified execution as the necessary next test. The result is the first formal, cross-domain-validated blueprint for physics-informed world models: place physics where the theorem says it is safe, never where intuition suggests—together with a concrete, reproducible cautionary result on how easily offline latent planning can be mistaken for evidence of control competence.

Abstract: Traditional deep learning architectures relying on standard affine layers (WX +b) partition the activation space into unbounded polyhedral regions, which inherently introduce chronic training instabilities such as gradient vanishing and severe generalization decay. To resolve this structural vulnerability, we introduce an alternative “bounded-by-design” geometric framework that completely moves beyond traditional affine transformations. By replacing them entirely with ball coverings, we reformulate layer mechanics around the fundamental metric of a feature point’s proximity to a ball centered at W. Crucially, computing the formulation using direct dot products fully preserves GPU parallelization performance without altering external network components. Empirical validation demonstrates that our bounded formulation effectively precludes overfitting, maintains numerical stability, and exhibits superior testing generalization compared to unconstrained affine layers.

Abstract: Accurate wind resource estimation in complex environments, such as the Colombian Caribbean, is challenging due to air–sea heat fluxes and nonlinear atmospheric dynamics. This systematic review and meta-analysis synthesized empirical evidence from 49 studies to evaluate whether advanced forecasting architectures (artificial intelligence, ensemble, and hybrid models) systematically outperform classical physical models. The pooled random-effects estimate indicated an overall reduction in reported forecasting error (\(g = -1.6874\)), though heterogeneity was extreme (\(I^2 = 99.99\%\)). Categorical meta-regression revealed no statistically significant advantage for advanced architectures over physical models (\(\beta_1 = 0.2558\), \(p = 0.9339\)). Mixed-effects meta-regression identified Sample Size (\(N > 10{,}000\)) as the primary significant moderator, with the final model explaining 40.47% of between-study variability (\(R^2_{\text{analog}} = 40.47\%\), reducing residual variance from \(\tau^2 = 2.5102\) to \(\tau^2 = 1.4943\)). Metric-specific contrasts indicated that physical and AI architectures are statistically indistinguishable, whereas hybrid models yielded lower pooled effects. classical physical models remain a reliable, stable reference framework, whereas AI and hybrid approaches demonstrate advantages only in specific, non-generalizable settings. These findings challenge the assumption of universal superiority for AI-based wind forecasting in complex coastal domains.

Abstract: The integration of Multimodal Large Language Models (MLLMs) into healthcare has the potential to drive a considerable advancement toward many AI use cases in medicine, offering transformative capabilities across hierarchical levels of clinical granularity, ranging from microscopic tissue analysis over organ level imaging and individual patient modeling to population-level health surveillance. However, such integration into clinical practice is constrained by a multifaceted “trustworthiness gap”: the mismatch between the raw capabilities demonstrated by Medical MLLMs in controlled experiments and their reliability, fairness, safety, privacy protection, explainability, and regulatory accountability required for clinical deployment in the real world. This gap requires systematic examination of both model behavior and clinical risk across the full lifecycle of Medical MLLM development and use. In this paper, we survey the existing literature on methods, evaluation, and benchmarks of trustworthiness research in Medical MLLMs. We first establish a multi-scale clinical landscape across tissue, organ, individual, and population levels to categorize current models, providing critical insights into the interplay between data heterogeneity and clinical task requirements. Subsequently, we propose a holistic, six-dimensional taxonomy of trustworthiness, comprising truthfulness, robustness, fairness, safety, privacy, and explainability. Using this taxonomy, we critically analyze existing literature to elicit recurring Medical MLLM trustworthiness failure modes and suggest appropriate remediation strategies. Furthermore, our survey enumerates open challenges in evaluating trustworthiness by means of traditional automated metrics, frontier “LLM-as-a-Judge” methods and expert-centric assessment protocols. As a response to these challenges we identify emerging research directions including dynamic and workflow-oriented evaluations for interactive Medical MLLMs. We aim for this work to serve as a systematic guide for researchers and practitioners aiming to develop the next generation of trustworthy medical AI by achieving clinically acceptable reliability through technological innovations. Project Link: https://github.com/junyuanM/Trustworthy-Medical-MLLMs-Survey.

Abstract: Large language models (LLMs) are increasingly expected to act as generalist agents capable of solving complex real-world problems. Training such agents, however, requires stable and diverse environments that support repeated interaction with stateful, tool-augmented tasks and provide verifiable feedback. Despite recent progress, the development of robust LLM agents remains limited by the lack of realistic, scalable, and executable training environments. We present InternAgentHarness, a scalable synthetic environment for improving the agentic capabilities of LLMs. Built upon the InternBootcamp training framework [1], InternAgentHarness instantiates executable agent environments through a four-layer interface that unifies prompt generation, tool execution, interaction control, and reward computation. We further introduce \bootcampcli, an agent harness that automatically converts diverse agentic tasks into a Bootcamp-trainable paradigm. Unlike static benchmarks, InternAgentHarnessmakes evaluation actionable: observed failures can be systematically converted into new synthetic tasks, filtered trajectories, reinforcement learning rollouts, and subsequent re-evaluation under the same executable interface. We instantiate InternAgentHarness on a suite of 10 tasks covering both text-only and vision-based agent scenarios. Starting from Qwen3-VL-30B-A3B-Thinking, both supervised fine-tuning (SFT) and reinforcement learning (RL) on InternAgentHarness substantially improve performance over untuned model. These results suggest that InternAgentHarness provides a practical foundation for scalable synthetic agent environments and enables the continuous improvement of LLM agents through an iterative cycle of evaluation, synthesis, training, and refinement.

Abstract: This survey presents a comprehensive review of coreset methods in deep learning, an important tool for improving data efficiency in large-scale neural networks. In general, “coreset” is an algorithmic technique for selecting a small yet representative subset of data to replace the full dataset, which can yield more efficient training process and meanwhile preserve model performance. In the past 20 years, coreset techniques have been widely applied to many classical machine learning problems, such as clustering, regression and classification. In recent years, the coreset techniques also begin to attract a lot of attention in modern deep learning area. However, designing effective coresets usually is a challenging task since we need to take account of the trade-off among multiple different factors, such as complexity, robustness and accuracy. In this survey, we focus on two common scenarios for using coreset methods in deep learning: (1) reducing the extremely high computational cost for training a deep learning model, and (2) improving the data utilization under resource constraints such as limited label budget or storage capacity. We begin by outlining the fundamental principles, advantages, and design challenges of coresets for these two scenarios. We also discuss the emerging applications of coresets in large language models. Finally, we identify several open problems and promising directions for future research.

Abstract: Designing analog hardware requires an engineer to chain symbolic calculations, a SPICE simulator, and a schematic editor manually, which consumes considerable engineering time. Large language models (LLMs) can reduce this effort because they are capable of writing netlists, invoking simulators, and drawing schematics through tool calls. However, current LLM agents are prone to hallucinating component values, fabricating SPICE syntax, and terminating at the netlist, because their outputs are not verified at any stage. In contrast, a human engineer never draws a schematic before the component values are calculated and the circuit is simulated. Taking advantage of this workflow, this paper proposes large language model electronic design automation (EDASYTH), a six-phase LLM-in-the-loop hardware design workflow that designs analog circuits in LTspice and lays them out in KiCad. EDASYTH chains symbolic mathematics verification, LTspice simulation, and netlist-to-schematic conversion behind the symbolic\_math\_mcp and bltspice\_mcp Model Context Protocol (MCP) servers and the electronics\_design conversion library, such that every LLM output must pass a verification check before the next phase begins. We evaluate the workflow on four LLM harnesses with both tutorial prompts. In the bandpass exercise, three of four models place both half-power cutoffs within 2\% of the specified 30 Hz and 25 kHz, and the fourth passes its own looser tolerance but measures a 12 dB passband gain. In the power supply exercise, only two of four models deliver both rails inside the specification, because one design cheated the exercise by replacing the AC input with a fixed DC source, and the other restarted after a simulation failure. In future work, this workflow may be extended to the fabrication and measurement of physical boards, thereby closing the hardware design loop.

Abstract: A quantitative risk assessment of human loss of control over advanced AI used a Bowtie diagram extended with fault tree-like and event tree-like analysis. Six primary threats were identified (recursive self‑improvement, power seeking, deceptive alignment, loss of corrigibility, off‑switch subversion, malicious misuse) and six consequences (systemic infrastructure collapse, economic breakdown, resource shortages, non‑human value lock‑in, human marginalization, global supply cascade failures). Preventive and mitigative barriers were assigned per pathway from expert literature. Input probabilities (threat base rates and barrier failure‑on‑demand values) were sourced from experts and modeled with triangular uncertainty distributions. A 1,000‑iteration Monte Carlo simulation propagated epistemic uncertainty, yielding a median probability of the top event (loss of human control) of 12.8% (90% CI: 11.3%–14.4%), roughly 1 in 8. The distribution is approximately symmetric with slight positive skew, indicating modest tail risk if barrier failures interact. Conditional on the top event, Expected Severity is 1.85 on a 1–10 scale (90% CI: 1.75–1.96), suggesting mitigation is effective in most scenarios. Results align with expert estimates and demonstrate barrier effects; narrow CIs reflect model consistency. Remaining tail risks support precautionary governance, increased alignment research, iterative risk modeling, and investment in international coordination with robust safety measures to reduce the existential risk of AI loss of control.

Abstract: A new class of pretrained time-series forecasting models — Chronos, TimesFM, Moirai, TimeGPT, Lag-Llama, and Tiny Time Mixers (TTM) among them — now offers zero-shot and few-shot forecasting capability without per-dataset training, a genuinely new operating mode for energy forecasting that existing BQEB benchmark infrastructure was not built to evaluate; strong supervised architectures such as PatchTST remain relevant as a reference baseline rather than as a member of this new class. This paper introduces BQEB FoundationBench, a reproducibility-first benchmark framework for evaluating foundation models on energy-forecasting tasks. Consistent with the architecture established across this research program, FoundationBench introduces no new architectural layer: it registers as the third module of the existing Benchmark Evolution Layer, alongside ForecastBench (Module 1, executed) and SecBench (Module 2, protocol specified), and extends the existing Benchmark Evolution Layer and Scientific Lineage Layer registries rather than defining parallel mechanisms. The framework specifies a model registry that records each candidate model's architecture, deployment mode, and license rather than treating all foundation models as interchangeable — explicitly distinguishing open-weight zero-shot models from Nixtla's closed-source, API-only TimeGPT, and from PatchTST, whose primary operating mode is supervised or self-supervised training rather than deployed zero-shot inference. It reuses BQEB ForecastBench's existing forecasting targets (next-hour electricity load, day-ahead electricity price) rather than defining new tasks, applying a zero-shot and few-shot evaluation paradigm to them instead of the trained-from-scratch paradigm ForecastBench used. The evaluation protocol specifies point and probabilistic metrics, a mandatory naive-baseline comparison, and bootstrap confidence intervals. This manuscript defines benchmark infrastructure; it is not a model paper, proposes no new forecasting architecture, and reports no execution results. Every planned-evaluation table in this manuscript is explicitly marked reserved for future experimental validation, consistent with the evidentiary discipline established in Papers 1 and 2.

Abstract: Digital twins and multi-agent artificial intelligence are increasingly proposed together for intelligent energy management — coordinating distributed energy resources, balancing virtual power plants, and orchestrating demand response across many independently-acting components — yet no reproducible benchmark evaluates this specific combination under a governed, versioned architecture. This paper introduces BQEB TwinAgentBench, the fifth module of the BQEB research program. Consistent with the architecture established across Papers 1–4, TwinAgentBench introduces no new architectural layer: it registers as the fifth module of the existing Benchmark Evolution Layer, alongside ForecastBench, SecBench, FoundationBench, and ExplainBench. The framework specifies eight benchmark tasks spanning digital twin state estimation, twin synchronization and fidelity assessment, multi-agent coordination, consensus formation, adaptive control, fault recovery, human-in-the-loop intervention, and open-world adaptation, formalized as a 6-tuple B = (W, A, T, C, E, R) over twin models, agent ecosystems, tasks, inter-agent communication, evaluation, and reproducibility artifacts. Eleven component metrics — spanning twin fidelity, synchronization accuracy, temporal consistency, agent coordination, consensus stability, communication efficiency, adaptation quality, fault recovery, safety preservation, human-intervention quality, and deployment readiness — are specified for aggregation into composite benchmark indices, following the same evidence-tier discipline established in Papers 3 and 4. This manuscript defines benchmark infrastructure; it reports no executed evaluations, and every planned-output table is explicitly marked reserved for future experimental validation, consistent with the evidentiary discipline established in Papers 1–4.