Large language model (LLM) coding agents develop software through experimentation, offering an opportunity to design the serving systems supporting their own operation. Effective serving requires coordinating implementation and operating decisions under session dependencies, latency constraints, and correctness requirements. We introduce ServerBench, a benchmark for agentic design of complete, specialized LLM serving systems. Agents experiment against a common objective, dollars per dollar: latency-qualified token value per dollar of GPU cost. Stateful replay preserves session dependencies, and independent verification checks the submitted system. Across 11 coding-agent models and three deployments, Claude Sonnet 5.5 achieves 2.97 dollars per dollar on Gemma E4B, $4.87\times$ the vLLM reference; 22 of 33 runs yield verified systems, and 13 exceed the strongest baseline. Our analysis identifies three findings: high-value designs coordinate capacity improvements with load selection; agents interpret performance feedback through workload-relevant experiments; and reliable delivery requires verification of the execution paths introduced by optimization. Overall, ServerBench supports systematic study of how agents translate service objectives and feedback into effective, verified serving systems for given task specifications.
We present a toolchain for development of ML models on AWS EC2 F2 cloud FPGA infrastructure from high-level language, implementing a complete Vitis HLS to HDK bridge to AFI workflow that has been absent from the academic literature despite eight years of AWS FPGA availability. We introduce a four-level methodology (FP32 PyTorch, bit-accurate quantized Python, HLS C-simulation, physical FPGA) to aid in localizing errors, including FPGA-specific circuit errors. Together these tools allow coding agents to effectively implement FPGA designs from user directions, testing and debugging as needed. We document the deployment workflow, timing closure strategies, and practical HLS engineering decisions that proved crucial. We validate using TinyBERT fine-tuned for address Named Entity Recognition, implementing the embedding layer and the first of four encoder layers on the FPGA hardware. Our implementation streams weights from host DRAM, yielding poor performance compared to CPU/GPU implementations; we discuss forward paths.
We study which training and calibration data help three open models solve ARC-AGI-2 tasks when the solver runs offline on four L4 GPUs for at most 12 hours. We compare data offline on public tasks and report 189 runs on the 120 hidden semi-private tasks. For the NVARC Qwen3-4B grid model, we built a corpus of 12,000 executable rules in a typed language and 95,993 tasks; a rule is admitted only if its behaviour on a fixed probe set differs from that of every admitted rule, and a task only if its pairs refute a shortlist of nearest competing rules and defeat 12 shortcut solvers. Continued fine-tuning on 1,920 real task records with revised four-GPU scheduling raised the semi-private score from 28.96 (mean of the released model) to 33.06 (mean of five runs, best 34.86); no checkpoint trained with synthetic tasks, even all 95,993 of them, improved on the real-task model. For the Qwen3.8-27B reasoning model, we fine-tuned on 1,036 of its own reasoning traces with correct final answers and compared sets that add traces of other models and synthetic tasks. With an inference procedure built around the model (a thinking cap enforced in the serving engine, a Python program checked on the demonstrations after each answer, a long first attempt per input), the score rose from 26.81 to 44.03, the level the ARC Prize Foundation measured for the base model without a time limit (42.36); most of the gain came from the inference procedure. For the Qwen3.8-Flash-Next mixture-of-experts model (125B parameters, 6B active), GPU memory limited concurrency. Pruning from 512 to 384 experts per layer by router-weighted activation saliency grew the key-value cache from 0.55M to 1.39M tokens; a 296-expert list adapted with ARC routing statistics, keeping experts for program tokens, plus an FP8 indexer cache, reached 1.96M. With a tool that checks programs, the model scored 56.67, our best result.
Efficient benchmarking estimates a system's full benchmark performance from a small number of task executions. Standard approaches rely solely on the binary correctness of executions from a collection of already-evaluated reference systems. An agentic system, however, produces a rich trace for each execution that can reveal substantially more about its capabilities. Leveraging this information requires useful trace representations. We propose representing a trace by extracting a structured summary of it with an LLM and embedding the summary with a standard text encoder. These representations preserve correctness- and task-related signal while minimizing the presence of system identity-, model family- and harness-related signals that dominate raw trace embeddings. Given this representation for each task execution from all reference systems, we construct low-dimensional representations of the systems and predict a new system's score from its position among the reference systems. On SWE-bench Verified and Terminal-Bench 2.0, augmenting an item-response model with this geometry-based prediction reduces mean absolute error from 0.13 to 0.08 (Wilcoxon $p < 10^{-4}$) and from 0.14 to 0.12 (Wilcoxon $p < 10^{-3}$), respectively, when provided a single random executed task. Further, given access to the full set of reference systems, our approach dominates baselines for any given number of executed tasks on both benchmarks.
Generative video models are increasingly used as visual world simulators, yet their primary strength lies in rendering plausible pixels rather than reasoning about how a world should evolve. Coherent world evolution requires more than visual fidelity: when an event occurs, a model must infer its consequences for the entities in the world and then render the resulting state. For example, after a wild beast appears, the model should not only depict the wild beast, but also infer whether a nearby person should flee, take cover, or fight back based on the realized scene. Existing diffusion-based video models lack an explicit mechanism for maintaining world state and reasoning over such event-conditioned consequences, often producing visually plausible but behaviorally static or causally inconsistent outputs. To make this reasoning gap concrete and measurable, we propose a new problem, Out-of-Sight Causal Evolution, which requires a model to render an event, reason about its consequences for the evolving world state, and reveal a causally plausible outcome as the world continues to evolve, even out of sight. As a baseline reference solution, we introduce EvoWorld, a brain-renderer collaboration framework in which a vision-language model serves as the world brain: it observes generated segments, maintains an evolving account of the world, and reasons about plausible state transitions, while a segment-autoregressive video renderer realizes them visually. We further introduce EvoWorld-Data and EvoWorld-Bench for training and evaluating event realization, causal reactions, and revisit consistency. Experiments show that EvoWorld substantially improves event-conditioned causal evolution over recent models while preserving visual quality and consistency, establishing a reference point for reasoning-guided, autonomous world evolution beyond the current field of view.
This paper introduces an innovative virtual reality (VR) exergame system designed to gamify High-Intensity Interval Training (HIIT) workouts through 11 predefined combo moves, including squat-punch combinations, bilateral arm strikes, and diagonal stretch combos. Unlike conventional tracking methods that often rely on wearable motion capture suits or limited headset-based sensors, our system uses a standard smartphone camera placed on a tripod to capture full-body movement. This approach offers a cost-effective and scalable alternative, particularly well-suited to the fast, dynamic motions typical of HIIT routines. The smartphone footage is processed in real time using a pose estimation algorithm, with extracted biomechanical data analyzed by a backend server built on the Model Context Protocol (MCP), Anthropic's open standard for connecting AI models to structured data sources and tools, and a GPT-4o large language model. The system generates personalized assessments of form accuracy, movement efficiency, exercise intensity, and potential injury risks, tailored specifically to high-impact, time-sensitive exercises. A within-subject study involving 20 participants (ages 18-25) compared experiences in VR and Non-VR HIIT sessions. Results indicated significantly higher enjoyment (Mean = 3.5 vs 2.8, p = 0.02), improved immersion/focus (Mean = 3.2 vs 2.8, p = 0.03), and increased likelihood to recommend (Mean = 3.2 vs 2.7, p = 0.04) under the VR condition. Confidence to continue did not differ significantly between conditions (p = 0.15), while perceived intensity remained comparable. The MCP system showed strong agreement with expert evaluation (87.5%, N = 10 sessions, one certified trainer), supporting its reliability as a feedback tool.
Large reasoning models (LRMs) perform well on challenging benchmarks, yet it remains unclear whether they form generalizable abstractions across conceptually related reasoning tasks. The Abstraction and Reasoning Corpus (ARC) provides a controlled testbed for studying abstraction, requiring solvers to infer transformation rules over two-dimensional grids from a small set of examples. While ARC evaluates rule abstraction within individual tasks, different tasks may share higher-level structure despite requiring different transformations$\unicode{x2014}$for example, several tasks may involve reasoning about objects contained within other objects. Benchmarks such as ConceptARC capture this structure by grouping ARC-style tasks into recurring spatial and semantic concepts, but they often contain only a few manually constructed tasks per concept. We introduce ConceptARC at Scale, a generator for creating thousands of ARC-style tasks targeting a specific concept. Using this generator, we construct ConceptARC-100, a dataset of 1,600 new, manually reviewed tasks, with 100 tasks for each of the 16 concepts defined by ConceptARC$\unicode{x2014}$ten times the size of the original benchmark. We evaluate a range of recent LRMs and models trained from scratch on ConceptARC-100 to characterize which concepts they solve reliably and where they struggle. We further analyze the models' internal representations and find clear semantic structure: task representations cluster according to the operations applied to objects, such as recoloring, removal, or translation, even when realized through substantially different rules and scenarios. Together, our results suggest that models form stable abstractions of transformation operations that generalize across diverse tasks.
Dexterous robotic hands are increasingly coupled to large-scale human data and foundation models, yet progress in hardware, representation learning and evaluation is often assessed separately. This survey reviews work from January 2024 to early November 2026, with earlier studies included to establish technical lineages. We organize the literature into five interdependent layers: hardware; data collection and curation; cross-embodiment transfer; learning and foundation models; and evaluation and applications. Human hand-object understanding and tactile sensing provide two cross-cutting perspectives. A system-level architecture, comparative tables and five case analyses connect design choices across these layers. We distinguish within-study experimental comparisons from heterogeneous corpus statistics, vendor specifications and company-reported results. Three conclusions emerge. First, the value of human data depends on kinematic, visual and contact alignment, not recording volume alone. Second, shared action representations and foundation policies do not remove embodiment-specific constraints on force, latency and contact. Third, task completion is insufficient to establish physical correctness or deployment readiness. Evidence remains limited for zero-shot transfer of contact-rich skills between morphologically different hands, long-duration reliability, and cross-sensor tactile generalization. We identify eleven open problems and propose reporting interfaces and controlled comparisons that can make full-stack progress more measurable. The review is a structured narrative synthesis rather than a pooled performance ranking or an exhaustive systematic review.
Responsible AI scholarship has found that ethical principles are frequently not reflected in AI design and development. Conceptualised as a `gap' between principle and practice, this is framed as a problem of implementation, awareness or compliance, and by implication, a deficit of systems or individual behaviour. This paper sets out an alternative explanation. Drawing on fourteen semi-structured interviews with actors designing and governing AI-enabled anti-trafficking systems across law enforcement, commercial, research, and survivor-support contexts, it argues that such apparent gaps emerge from the varied and complex institutional and professional conditions through which actors come to recognise some issues, but not others, as ethical problems requiring action. We show that epistemic choices are intertwined with ethical judgements, and trace four interconnected moments---perception, judgement, situated action and socio-technical materialisation---through which concerns become intelligible, are weighed and acted upon, and are embedded in technical artefacts and governance arrangements. Actors often hold genuine ethical commitments, but these commitments are shaped by institutional logics and professional moral economies: some concerns are materialised, while others are narrowed, deferred or left unrealised. We show that ethical agency is exercised within and between competing constraints of profession, role, prior experience, and institutional settings. The paper argues that responsible AI governance requires purposeful and well-developed epistemic infrastructures that can cut across these constraints to establish institutional conditions through which the knowledge of affected communities and potential harms becomes receivable and actionable in design.
We introduce LifeVal, a benchmark evaluating AI model performance on personalized agentic consumer tasks. LifeVal covers the most valuable life tasks that a personalized digital superassistant could automate, representing over 1.7 trillion US dollars in annual willingness-to-pay across 148 consumer task categories (e.g., managing finances, moving apartments, searching for a job, finding childcare). To construct our activity distribution, we conducted a new survey of willingness-to-pay for a digital superassistant across a demographically representative sample of over 3,800 consumers, cross-referencing our results against the American Time Use Survey (Bureau of Labor Statistics, 2025) and the Multinational Time Use Study (Fisher et al., 2025). For the most valuable consumer activities, we recruited task creators to create AI evaluation tasks personalized using real ChatGPT conversations and memories, reference files, and personal context. Task creators had to pass our expertise threshold (including personally completing the task multiple times and seeing the results through in their real lives for over a year) and also write a rubric, ensuring that our evaluation measured the best personalized completion for that task. All tasks had to pass at least three rounds of human quality-control before acceptance into our final evaluation sets.
Infrastructure security does not establish whether an LLM agent selected an allowed tool, grounded its answer in tool output, or followed a required workflow. AgentTrustCI is a Java/JUnit 5 framework that expresses these application-layer expectations as policy, grounding, and trajectory checks over recorded traces. A deterministic replay implementation supports repeatable build-time tests; a judge interface provides an extension point for optional semantic evaluation. On 55 synthetic cases spanning seven enterprise domains, the full pipeline correctly classifies all 30 injected faults and 25 passing controls, including ten non-trivial controls, and misses none of 12 designated critical cases. Each single-layer configuration correctly classifies 35 cases because the corpus deliberately isolates ten faults per layer. A grounded answer obtained through a disallowed sequence illustrates the additional coverage of trajectory checks. Repeated measurements quantify validator execution separately from process startup and reporting. These results establish coverage of the authored regression suite, not production accuracy. We describe a production-trace validation protocol, optional LLM-judge integration, and maintenance of expectations as workflows evolve. The contribution is a practical pattern for expressing behavioral contracts within an existing Java release pipeline.
Search-augmented reasoning systems can reach correct answers while making errors in intermediate steps. Reliable auditing must check whether evidence $D$ covers the question $Q$, reasoning $T$ addresses $Q$, and $T$ is grounded in $D$. We introduce GraphRail, which represents requirements, evidence, and reasoning as structured claims and compares them without a generative judge or gold answer. Across three multi-hop QA benchmarks and three reasoning frameworks, GraphRail improves error recall over generative verifiers while remaining stable under stochastic extraction. Its diagnoses can also guide online reasoning, providing a traceable way to detect and correct intermediate errors.
Quantum circuit Born machines (QCBMs) can learn distributions that are hard to sample from classically, but their value depends on generalizing from finite training data. Here, we study how hardware noise affects this generalization. We show that, for any noise process, the generalization gap of a QCBM trained with the maximum mean discrepancy is bounded by the complexity of the distributions the noisy model can reach. Scrambled local Pauli noise contracts this complexity but biases the model towards the uniform distribution, so residual noise acts as a regularizer whose optimal strength can be intermediate. We propose GENIE, a framework that realizes a chosen noise level with capped error mitigation and selects it during training from held-out data within a fixed execution budget, with provably bounded regret. In simulations on synthetic and census data, GENIE stays close to the best of four noise-control strategies, whereas every fixed noise level falls short in many settings.
Defenders and attackers race to find the same software flaws. I use the launch of Project Glasswing, which gave a small group of large US tech firms early access to Claude Mythos Preview, a model for finding software vulnerabilities, to study how access to frontier AI affects defenders, external researchers, and attackers. Using the universe of publicly disclosed vulnerabilities, I find that firms with access more than doubled their disclosures relative to Chinese and Russian tech firms without access, disclosed about 150 percent more high-impact vulnerabilities, and fixed reported flaws as quickly as before. As treated firms found more flaws themselves, external researchers, who sell their findings through bug bounties, reported about half as many vulnerabilities in treated firms' code and more in other firms' code, without leaving the market. There is no evidence that attackers exploited the additional vulnerabilities: the share exploited within a month of disclosure fell by about two thirds, mainly because a smaller share was already exploited when disclosed, while the number exploited did not change. Frontier AI helped defenders find and fix more flaws before attackers exploited them, but its restricted release creates an asymmetry in software security between firms in different countries.
A national digital twin was conceived as an ecosystem of twins connected through mappings to a common foundation, but the roadmap that promises to confirm those mappings are valid never says what valid means. We give the test a precise form, coherence and conservativity of the merged ontology, borrowed from logic-based alignment repair, and apply it to the Information Exchange Standard (IES), first pair by pair and then to whole hubs. Crosswalks from IES into thirteen ontologies across seven sector groups, anchored on the building ontologies BOT, Brick, and SAREF4BLDG, pass through separate stages of matching, human review, certification of each SKOS correspondence against the two class hierarchies, and reasoning with HermiT. Pair by pair, the stages catch different failures: certification alone removed all 526 conservativity violations of the raw crosswalks, review alone left 73, and together they made eleven of twelve overlapping pairs coherent and conservative. Pairwise certification does not compose. Merging the three certified building crosswalks into one hub made five Brick classes unsatisfiable, through a homonym between the IES and Brick notions of a floor that only BOT's disjointness axioms expose; rejecting that one row gave a coherent and conservative hub that entails 290 cross-ontology links nobody wrote down. The core hub is inconsistent before any crosswalk is added, because two published standards, the Open Energy Ontology and BFO 2020, contradict each other; with either one left out it is coherent and conservative. A hub, not a pair, is the unit an ecosystem of twins has to certify.
We present a shared framework that adapts Cosmos world foundation models for real-time interactive world simulation in Physical AI. By streaming visual observations in response to actions and scene controls, the framework enables users and embodied agents to interact with evolving virtual environments. It supports both predicting the visual consequences of actions and generating observations from physical state maintained by an external simulator. We demonstrate this approach across camera navigation and robot manipulation, controllable video transfer, and multi-view autonomous vehicle (AV) simulation. These examples provide a shared foundation for developing interactive virtual environments for Physical AI training and evaluation. Our code and model checkpoints are available at https://github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3-examples. The project website is available at https://research.nvidia.com/labs/cosmos-lab/cosmos-simulation.
Continual agents must handle urgent interaction while retaining alternatives, replaying errors, and committing safe long-term changes. A previous digital-life framework distinguishes online (busy) and offline (idle) channels, but leaves open whether these are separate programs or phases of a common process and how their operations should be ordered. Motivated by recent observations of rapidly alternating perceptual representations in primate cortex, we develop a temporal organization layer with phase-tagged micro-epochs, an observable duty cycle, independent content and resource controls, and three nested clocks separating candidate exploration from durable commitment. We give a projected duty-cycle update, state its deterministic descent guarantee and the additional noise term required for a stochastic guarantee, and specify coupled busy/idle execution contracts. Five synthetic experiments (64 seeds each) examine descent and noise, objective misalignment, gain asymmetry, resource coupling, and dropped replay logs. In the gain/noise model the best tested gain ratio is 1.55; decoupled scheduling yields 26.6% versus 11.1% success on a constructed urgent/high-load subset. A frozen-backbone Qwen3-8B prototype further checks orchestration, but its proxy outcomes are not task-benchmark evidence. The results establish a testable scheduling design and identify its conditions of validity, not machine consciousness or general-purpose agent superiority.
AI agents rely on third-party tools, such as Model Context Protocol (MCP) servers, whose publishers can be authenticated and whose execution can be restricted by runtime security policies. Yet authenticity does not guarantee benign behaviour: a legitimately signed tool may perform its advertised function while concealing malicious code that leaks confidential information. We demonstrate this trust failure with a signed MCP tool that embeds sensitive data in requests to a policy-allowed destination. Meta's MCPGuard-Dynamic, a state-of-the-art runtime defence, blocks a direct exfiltration attempt in the same run, yet fails to prevent this attack in every trial. We introduce germline execution, which limits information disclosure without inspecting or trusting tool implementations. Each tool invocation runs in ephemeral copies forked from a data-free germline process, each receiving only information authorised for its destination. Even if malicious code executes, what it can disclose is bounded by the data available to its copy. Under explicit assumptions on copy confinement and a data-free germline, we prove sink noninterference for every signed release and, for releases deterministic within our replay model, show that a verified witness establishes a flow violation. Across 153 attack runs, no unauthorised disclosure was detected under germline execution, compared with 5.9-74.5% of runs under baseline defences and an ablation. Invocations served by a single copy completed in 10-49 ms, 2-28x faster than restarting the tool for every invocation, the conventional way to clear state. These results show that data-isolated execution protects agent toolchains against malicious code concealed within authenticated software.
Category systems often reach their users as text, and readers apply them with tacit repairs. We ask what such a text determines once the repairs are removed, using diabetic retinopathy grading, where one international scale circulates in several documents. We transcribe four texts (the English screening definitions of 2025, the international scale as tabulated in 2017 and as published in 2003, and an Italian guideline of 2015) into one rule language over 24 finding variables, and make every point a text leaves open a named precisification. Exhaustive enumeration of 1,425,807,360 finding profiles, re-derived by an SMT solver and spot-checked by an OWL reasoner, shows that every text that assigns actions leaves some eyes without a category under every reading; that the 2017 tabulation has a proliferative class that is empty when read literally and a macular gap that the 2003 text lacks, so that no function maps the 2003 grades onto the 2017 grades; and that on image findings alone the English and Italian texts order some pairs of eyes in opposite ways. On 927 annotated photographs from three public datasets, with retinal thickening assumed absent, the English text and the 2017 tabulation disagree on referral for 26.8% of the eyes of one dataset (DDR) on which both decide, and in a pre-registered experiment classifiers trained on each text's labels side with their own text. Thickening cannot be seen on a colour photograph; left free, every disagreement involving England becomes undecidable, so it rests on what the camera cannot show. In a second pre-registered study, language models that answered without deliberating repaired most gap and overlap items even when told to read literally, while Claude Sonnet 5.5 and Claude Opus 5.5 reported the defects far more often; within two models, deliberation reduced silent repair without removing it.
Long-term time-series forecasting supports decisions in energy, transportation, and other applications where reliable predictions are needed beyond the observed data. Although forecasting research has advanced model architectures, fixed activation configurations provide limited guidance on how activation choices should vary across datasets and prediction horizons. This study aims to establish when task-specific activation choices improve forecasting accuracy and how previous experiments can guide the selection of suitable configurations for new tasks. We develop a calibrated selection policy that learns a reliable default from prior experiments, adapts the activation function and its location to each task, and retains the default when evidence for a change is insufficient. In an evaluation supported by 1,225 Informer training runs across 17 dataset variants and 51 forecasting tasks, the policy reduces mean squared error by 7.86% on average relative to the conventional fixed-activation baseline and improves 72.55% of tasks. These findings establish activation choice and placement as practical opportunities for improving existing forecasting models and show how accumulated experimental evidence can guide configuration decisions without repeating an exhaustive search for every task.