Papers

An AI agent's rebuild is only as good as the process that produced it. Prior work found that once a model is strong enough, a multi-agent rebuild pipeline loses to the simplest approach: giving the model the original code and one instruction (AgentModernize). We present rebuild-dossier, an open-source tool that locks an application's real interface - its exact inputs and outputs - before any code is written, then enforces one-test-at-a-time building through automated checks, not written instructions alone. Three results shape this evaluation, with differing amounts of evidence. First, in a small comparison, the compliant agent failed a held-back test while the rule-breaking agent passed everything - proof that a passing suite doesn't certify correctness when tests can be gamed. Second, we tested whether this beats simply giving the weaker model the source and one instruction: tied on a small app, but lost outright on a larger one where the automated check wasn't even running - pointing to the check mechanism, not interface-locking, which held up separately. Third, every claim here is checked at three levels - the agent's own report, an automated log, and the actual files produced - catching real errors, including a bug in our own logging code, that a single level would have missed. These risks reproduce on a different model and toolchain: a stronger model followed our process three times running, something the weaker model never managed. The tool is public, MIT licensed, and reproduces end to end against our own applications.

Enterprise repositories are large, heteroge- neous, and continuously updated, making re- trieval difficult when efficient access, source- faithful evidence, and cross-query adaptation must be supported together. Enterprise mem- ory extends retrieval beyond the model con- text, but existing systems do not jointly address collection-specific hierarchy construction, low- cost routing, detailed evidence loading, and workload-aware memory updates at this scale. We introduce MEMONDEMAND, short for On- Demand Memory, a memory management sys- tem with three coordinated mechanisms: a dy- namic multi-level hierarchy that determines the abstraction structure and depth for each col- lection, dual memory at every hierarchy level that separates distilled routing from detailed evidence, and on-demand memory promotion that updates node priority under a bounded active-state budget. On EnterpriseRAG-Bench, MEMONDEMAND outperforms the strongest published LB#1 result at every evaluated scale from 10M tokens through the complete 618M- token collection, with gains of 12.23% at 10M and 4.66% at 618M. Results on FinanceBench, HotpotQA, and FRAMES further show strong performance across financial, multi-hop, and fact-retrieval settings. Together, these results establish MEMONDEMAND as an accurate, ef- ficient, and scalable memory solution for very large enterprise repositories across data scales, domains, and evidence requirements. Our code is available at https://github.com/ xfab-xinyuansong/MemOnDemand.git.

Large Language Models (LLMs) achieve strong reasoning performance, but their robustness to realistic lexical corruption remains poorly understood. We evaluate four open-weight instruction-tuned models and frontier models across four reasoning benchmarks under keyboard noise, character swaps, and filler insertion. Character-level perturbations substantially degrade accuracy, especially on multi-step reasoning tasks, while filler insertion has little effect. We trace this asymmetry to Attention Diversion: lexical corruption fragments subword tokenization, and the resulting fragments attract disproportionate attention mass, concentrated in middle and final transformer layers. Length-matched controls confirm that fragmentation, not prompt length, drives the loss. A factorial intervention then shows why the damage is hard to undo: fragmentation corrupts token content and attention allocation together, and the two are coupled. Restoring clean attention while the content remains corrupted is actively harmful, restoring content alone is insufficient, and only restoring both recovers a substantial share of the gap. This coupling explains why inference-time strategies, including chain-of-thought prompting, spell-checking, self-repair, and stronger repair models, fail to consistently recover performance: each addresses one channel at a time. Code and data are available at https://github.com/Jiaqian-Janelle/Attention-Diversion

Language models are usually judged by a single accuracy score, which does not reveal how their performance degrades as inputs are perturbed. We present a graded, multi-family, failure-aware framework for stress-testing reasoning models. It perturbs each problem along a multi-level severity ladder across seven families: six that preserve the answer, paraphrase, input noise, formatting, irrelevant context, context load, and conflicting instructions, and a Knowledge Boundary family that removes answerability so that refusal becomes the correct response. Every test is validity-gated and labeled by its measured severity, and each model is summarized by per-level Accuracy, a magnitude-weighted Stability, and a per-family Collapse Point defined relative to the model's own baseline. Instantiated on the same 100 seed problems used by GSM-Symbolic, expanded into 4,473 gated tests and run on four models spanning capability tiers, the framework exposes structure that an aggregate score hides: the level at which a model fails is family-specific rather than global, and two stressors expose consistent weaknesses across all models: conflicting instructions and questions built on an impossible premise. Recognition of unanswerability is otherwise uneven, reliable on missing information and fabricated evidence but weak on impossible premises. These failure points are invisible to standard accuracy reporting.

Modern language models and agents increasingly require persistent memory for complete codebases, long interaction histories, and heterogeneous enterprise records. The key challenge is to keep hundreds of millions of tokens searchable while passing only bounded source evidence to the answer model. We introduce MegaMem, a source-resolved dual-view retrieval system that separates semantic access from generation evidence. Distilled records and detailed evidence are searched with original and transformed queries; every distilled hit resolves to an immutable source ID before reciprocal-rank fusion, deduplication, and cross-encoder reranking; and only the highest-ranked detailed evidence within a fixed budget supports generation. Post-answer attribution then identifies which loaded sources support the fixed answer. We evaluate MegaMem on EnterpriseRAG-Bench, which contains more than 500,000 heterogeneous enterprise documents and approximately 650M tokens. MegaMem improves Overall from 68.22 to 82.26 and reaches 86.50 Correctness. These results show that MegaMem supports ultra-large persistent memory while preserving strong answer accuracy under a bounded generation context. By separating searchable memory scale from answer-context size, MegaMem provides a practical path toward accurate retrieval over memories ranging from hundreds of millions to one billion tokens. Our code is available at https://github.com/ xfab-xinyuansong/MegaMem.git.

Biomedical multi-hop question answering (QA) requires models to connect evidence across intermediate entities such as diseases, drugs, proteins, and phenotypes. Existing agents typically rely on static retrieval workflows or coarse-grained prompt rewriting, which can lead to instruction drift when reasoning procedures need to be updated. We propose SSE-Bio, a structured self-evolving agent with an agentic retrieval policy for multi-hop biomedical reasoning. Instead of globally rewriting agent instructions, SSE-Bio maintains a structured state, selectively retrieves knowledge triplets and prior templates through a trainable proxy policy, and improves its reasoning memory through fine-grained template editing. To optimise retrieval decisions, we introduce a proxy-training strategy based on group relative policy optimization, where the proxy is improved through decision-contrastive groups over alternative retrieval choices. Experiments on three biomedical multi-hop QA benchmarks show that SSE-Bio consistently outperforms existing baselines, achieving an improvement of 6.56 absolute points over the strongest self-evolving baseline on BioHopR.

Craniosynostosis severity analysis increasingly relies on statistical shape models (SSMs) to quantify cranial morphology, but most existing workflows depend on computed tomography or heavily curated three-dimensional (3D) photographs. Raw clinical 3D photographs provide a radiation-free and repeatable alternative, yet often contain shoulders, hands, hair, clothing, scanner noise, and incomplete boundaries that corrupt correspondences. We introduce the Template-constrained Robust Artifact-aware Correspondence Estimation (TRACE) framework, an unsupervised method for constructing SSMs directly from artifact-contaminated clinical 3D head photographs. TRACE predicts sparse anatomically corresponding head-surface control points from the raw point cloud, refines them through a coarse-to-fine Surface-Aware Deformation cascade, and uses thin-plate spline warping to deform a clean template mesh into a subject-specific head reconstruction. This template-constrained formulation keeps dense correspondences on clinically relevant head anatomy while suppressing non-head artifacts. The correspondence module is decoupled from the point-cloud encoder, enabling the same deformation pipeline to be paired with different backbones, including PointNet, DGCNN, and Point Transformer V3. Across all backbones, TRACE substantially improves surface sampling, topology preservation, and shape-model quality over prior SSM methods, providing a scalable foundation for photograph-based craniosynostosis shape analysis and a framework that may extend to other artifact-contaminated surface scans when an appropriate clean template is available.

Printability assessment in additive manufacturing is typically conducted at the geometry level before printing to determine whether a computer-aided design (CAD) model or stereolithography (STL) file can be successfully fabricated. Task suitability, in contrast, is usually evaluated after printing to determine whether the fabricated part satisfies the requirements of its intended use. As a result, for non-expert users to print functional parts, unsuitable material or process choices may only be identified after fabrication, leading to repeated printing, material waste, and user frustration. To address this challenge, this paper leverages the reasoning and language-understanding capabilities of large language models (LLMs), while grounding the reasoning with geometry evidence and structured material/printer knowledge to generate reliable pre-print recommendations. Given a stereolithography (STL) model and a natural-language task description, the framework generates a structured recommendation covering printability, material choice, process parameters, design guidance, risks, and explanations. We evaluate the framework on focused STL benchmark scenarios with novice-style task descriptions. The proposed method achieves 75.0% printability over 96 physical validation trials, with 88.9% task suitability among successfully printed samples. It also improves Gemini 2.5 Flash-Lite material-selection accuracy from 37.5% under pure LLM to 90.0%. Expert evaluation further shows improved report quality, while post-print feedback improves recommendations on selected problematic cases. These results suggest that user task intent, geometry evidence, and structured material knowledge are all important for reliable task-driven printability assistance.

Breast Cancer Multidisciplinary Team (MDT) meetings manage increasingly complex cases under considerable time pressure, and documentation requirements can reduce clinical efficiency and decision quality. Existing AI based MDT workflows rely on cloud-based processing, limiting their use because patient discussions contain identifiable information. We developed a fully on-device AI pipeline using open-source Automatic Speech Recognition (ASR) and Large Language Models (LLMs) that transcribes breast cancer MDT discussions, structures clinical information, and generates treatment recommendations using retrieval-augmented generation (RAG) grounded in National Institute for Health and Care Excellence (NICE) guidance. The pipeline runs on a single NVIDIA Jetson AGX Orin, ensuring that patient audio, transcripts, and outputs remain within institutional infrastructure. Evaluation included two recorded simulated MDT discussions, ten clinically validated synthetic discussions, and 1,270 acoustically augmented recordings. Optimisation of Whisper large-v3 reduced word error rate by 20.7% and 24.4% on the recorded discussions and achieved performance within 0.58% WER and 1.58% word information lost of a commercial clinical ASR benchmark on augmented audio. MedGemma-RAG identified 2.3 times more MDT-concordant interventions than a proprietary cloud comparator (p = 0.020), with no significant difference in overall accuracy. Stakeholders identified automated documentation, treatment recommendation support, and case triage as the most credible near-term applications while highlighting workflow integration, governance, and clinician trust as key implementation challenges. These findings demonstrate the feasibility of privacy-preserving, fully on-device AI for MDT documentation and guideline-informed decision support, providing a foundation for prospective clinical evaluation.

As agents grow more capable and autonomous, their tendency to reward hack, satisfying a task's checks while violating its intent, becomes an increasingly important failure mode. Measuring reward hacking is itself challenging, as detection typically relies on human inspection or LLM judges, both of which can be unreliable. The hack-verifiable environments (HVE) methodology addresses this challenge by embedding detectable hacks into tasks, allowing reward hacks to be identified automatically and reliably. In this work, we adapt HVE to Terminal Bench, a leading benchmark of real-world terminal and coding tasks, and introduce Hack-Verifiable Terminal Bench (HVTB). Using HVTB, we measure reward-hacking rates across frontier models and study whether prompts with varying amounts of information on the hack can mitigate this behavior. This lets us test whether prompting can prevent not only known reward-hacking strategies, but also 'unknown unknown' exploits that the prompt does not anticipate. We release all environments and agent traces at https://majoroth.github.io/hack-verifiable-environments/hvtb

We present GCA (Gaussian Constitutive Alignment), a framework for learning implicit constitutive laws from monocular dynamic video of deformable objects represented by 3D Gaussians. Given a static multi-view scan for geometric initialization, our method learns intrinsic physical dynamics solely from a single fixed-viewpoint video of the moving object. Existing implicit methods often suffer from local minima under noisy supervision and lack physical interpretability, while explicit approaches rely on predefined constitutive equations, limiting generalizability and becoming unstable in monocular settings. To address these challenges, our framework unifies LoRA-based adaptation with two key alignment modules. First, we propose Rank-based Depth-Geometric Anchors (RDGA) to establish robust geometric constraints from monocular dynamic observations via scale-invariant rank-based depth alignment, reducing the reliance on unreliable pixel-level color supervision. Second, a Constitutive Prior Regularizer (CPR) integrates classical constitutive models as soft differentiable priors, regularizing the optimization while preserving the flexibility of implicit modeling—even when the actual material is absent from the hypotheses. Extensive experiments on synthetic, real-to-sim, and real-world datasets demonstrate that GCA outperforms existing methods, achieving 48% lower Chamfer Distance than the strongest baseline on synthetic benchmarks while remaining robust under monocular supervision.

Large language models can produce fluent reasoning traces whose local semantic errors propagate to an incorrect conclusion, while unconstrained self-correction may preserve, amplify, or introduce errors. Existing diffusion language models provide iterative refinement, but usually define noise as token masking or replacement rather than as errors in the reasoning process. We present Semantic Reasoning Denoising (SRD), an operatorized Markov denoising method for natural-language reasoning trajectories. SRD represents semantic noise with executable error operators that describe the error type, its location, and the corrupted and repaired propositions. Composing these operators constructs progressively noisier states. During training, the model learns to identify the semantic noise active in the current trajectory and to reconstruct the paired adjacent lower-noise state. During inference, noise-level-aware denoising repeatedly predicts an inverse operator and checks whether it is applicable, so each executed update makes a localized move toward a stable trajectory. Across six in-domain benchmarks spanning mathematics, code, knowledge, and commonsense, SRD improves the strongest same backbone baseline by 3.2 points on average. On seven cross-dataset transfer targets, it remains competitive with Llama-3-8B-Instruct and improves the strongest Qwen3-8B baseline average by 2.9 points. Analyses of noise sources, objectives, and denoising depth further show that structured semantic-noise prediction and iterative operator execution are central to the improvement.

Modern software systems require earlier and more scalable vulnerability severity assessment to reduce exposure to high-impact security flaws. Security analysts typically assign CVSS scores, but this manual triage does not scale with the growth of disclosed vulnerabilities and often depends on cloud LLM services that raise confidentiality concerns. This paper presents an industrial case study on predicting CVSS v3.1 scores directly from vulnerable C/C++ snippets using in-context learning with locally deployable, open-source LLMs. We compare proprietary data with the Big-Vul dataset, showing sufficiently aligned CVSS distributions to justify Big-Vul as a proxy for industrial data when constructing prompt-based testbeds. We then vary in-context configurations and model parameters, evaluating CodeLlama2-7B, CodeLlama2-13B, Mistral-7B, gpt-oss, and GPT4o-mini using mean squared error (MSE) and feasibility metrics. Our results show that medium-sized open-source code models, particularly CodeLlama2-7B, can approximate the best cloud performance for CVSS regression when guided by lightweight, output-constraining prompts, offering a practical, privacy-preserving building block for severity triage in industrial settings.

Synthetic clinical data generation with large language models addresses the scarcity that limits cancer staging research, but oncology hallucinations are categorically harmful: one clinically impossible staging assignment contaminates every downstream model trained on it. Neuro-symbolic pipelines validate during generation, yet the contribution of individual quality-assurance components remains unclear. We report three controlled studies isolating gate necessity, constraint attribution, and retrieval conditionality, holding generation protocol, diversity thresholds, and fine-tuning hyperparameters constant across adapter conditions. The symbolic gate enforces schema completeness, ontology coverage against the Systematized Nomenclature of Medicine, and staging-logic consistency under American Joint Committee on Cancer eighth-edition rules. Ungated, 29.9% of records contain schema failures and 20.1% contain clinically invalid staging. Schema validation is the load-bearing filter: within the fully gated corpus it rejects 148 of 512 records, ontology grounding a further 24, and staging-logic validation none—the only generator producing logic violations is already excluded on schema, making clinical-logic validation a generator-conditional safeguard rather than the dominant filter. Retrieval augmentation is strongly model-dependent: it improves gate compliance for one generator by 12.5 percentage points, has no measurable effect for a second, and collapses output in a third. Across gated configurations ontology density is largely unchanged, indicating that symbolic validation improves clinical validity rather than vocabulary richness. Symbolic gating therefore buys corpus validity but no commensurate gain on real lung-cancer notes in this study; retrieval should be evaluated per model, and ontology density should not be reported as a proxy for corpus quality.

Rising energy demand, fossil fuel depletion and climate change highlight the need for more efficient energy production and consumption. Offshore oil and gas platforms face challenges related to inefficient energy use, system failures, accessibility and environmental impact. Machine learning (ML) offers opportunities to improve the safety, sustainability and efficiency of these systems; however, previous research has largely focused on increasing oil production rather than reducing energy consumption on platforms. This study investigates the use of ML and search algorithms to improve diesel efficiency on an offshore oil platform. Data collected over 18 months from a platform in Scotland were analysed, focusing on four diesel generators as the primary diesel-consuming equipment. Following exploratory data analysis and outlier detection, regression models were developed to predict daily diesel consumption for different generator power loads. Multiple Linear Regression and Artificial Neural Networks achieved the best predictive performance compared with Extra Trees Regression, Extreme Gradient Boosting and Random Forest. Search algorithms were then used to identify combinations of generator power loads that minimised daily diesel consumption. The results showed an average diesel saving of 27% per day compared with the worst daily power-load combinations, equivalent to approximately 24,000 litres/day. These findings demonstrate significant opportunities for improving energy efficiency on offshore oil platforms using ML-based optimisation.

Turn-taking is a basic organizational feature of human conversation and remains difficult to model in natural, synchronous dialog systems. While existing research has explored multimodal approaches and large language models for turn-ending prediction, there is a lack of naturalistic conversational corpora specifically addressing turn-taking dynamics in Turkish. This study introduces a multimodal Turkish conversational dataset of unscripted dyadic interactions, comprising synchronized front-facing video, per-speaker audio channels that allow overlapping speech to be attributed to individual speakers, and time-aligned transcriptions. Turn-taking prediction is formulated as a binary classification problem, and a Genetic Algorithm (GA) is employed to optimize interpretable decision rules derived from visual, acoustic, and linguistic features. A hybrid AND-OR rule representation is adopted in the proposed framework to represent the alternative cue combinations that precede a turn transition.

Geothermal well arrays, which organize multiple geothermal wells into carefully planned geometric configurations, provide opportunities to enhance energy production capacity and increase fault tolerance. The development and adoption of these emerging geothermal technologies could be accelerated through the recent advances in large language models (LLMs) and high-level high-performance languages. A challenge in LLM-based applications is the reliability of the generated outputs, as they can be prone to subjective biases and hallucinations. This study assesses the potential of cutting-edge LLMs - such as ChatGPT, Gemini, Claude, Grok, and domain-specific models like AskGDR - as expert assistants that can synthesize insightful interpretations of complex geothermal data, as well as improve feature capabilities of geothermal models and numerical software. We developed a novel approach, leveraging Google's recently introduced AI assistant, NotebookLM, to accelerate the generation of unpublished quantitative geothermal benchmarks. The rapid generation of these evaluation instruments is essential for assessing the swiftly evolving capabilities of emerging language model technologies. In particular, we use these benchmarks and LLM-based interviews to analyze opportunities and limitations of two promising technologies: geothermal well arrays and closed-loop coaxial wells. Furthermore, we present a case study illustrating how LLMs can facilitate auto-parallelization of geothermal numerical models. Our analysis emphasizes their application in digital twins and underscores the importance of high-level, high-performance code generation. This line of research could play a transformative role in the geothermal sector by enabling the next-generation of decision-support applications, integrating data analysis, informed recommendations, and more dynamic numerical modeling workflows.

World-Action Models (WAMs) build robot control on video-generation backbones, which jointly predict dense future visual trajectories and robot actions. We argue that video generation is an unnecessary intermediate objective for world-action modeling. For robotic manipulation, the goal of a world model is not to reproduce how the world looks at every intermediate moment, but to predict the state that the world will reach after an action is executed. The intermediate frames only describe the visual transition between physical states, which consumes substantial model capacity and computation, but do not directly specify the physical outcome that the robot action is intended to produce. In this paper, we propose DELE-w0.5, which infers robot actions from predicted future states without relying on video generation. Concretely, DELE-w0.5 infers the action sequence from its corresponding compact future latent state. The future latent state captures the action-relevant physical outcome of robot interaction and serves as an explicit bridge between world modeling and action generation. The core design principle of DELE-w0.5 is to model how the physical world changes under robot actions, rather than how its visual appearance evolves frame by frame. This formulation removes the high-dimensional visual redundancy introduced by dense video representations, and it therefore enables cheaper training and low-latency inference. Across 640 real-robot trials on four long-horizon manipulation tasks, our DELE-w0.5 achieves the best performance among all compared policies, attaining 62.5% overall full-task success and 81.3% macro ordered-stage progress. It outperforms the strongest baseline by 32.5 percentage points in full-task success and 20.1 percentage points in macro progress.

Attention-based multiple instance learning (ABMIL) using pathology foundation model embeddings is effective for slide-level tasks, but exhaustive inference requires applying a large image encoder to every foreground tile despite the subsequent attention distribution often concentrating over a small subset of informative regions. We introduce ADMIL (Attention-Distilled Multiple Instance Learning), a selective-compute framework that distills an ABMIL teacher's attention into a lightweight tile-selection model, PriorNet. Using an EfficientNet architecture, PriorNet learns the teacher attention distribution from raw tile pixels with KL divergence; at inference, it scores the foreground pool, selects the top-K tiles, and invokes the expensive foundation model only on that subset before a selected-bag ABMIL student predicts the slide label. Across BRACS, PANDA, and CAMELYON16, ADMIL matches full-teacher headline performance at K=4, 8, and 128 tiles, respectively, avoiding >98% of foundation model (Virchow2) tile embeddings and model inference FLOPs. Random and teacher-attention oracle controls show that this result depends on task-relevant selection rather than tile-count reduction alone. Quantitative and qualitative analyses suggest that PriorNet recovers the teacher's tile ordering with high fidelity while focusing on task-relevant morphological regions. ADMIL shows that nearly all expensive tile encodings can be removed without sacrificing slide-level performance, providing a potential path for more efficient deployment in clinical settings where latency and compute costs are key considerations.

Whether the slow solar wind originates from one coronal source or two distinct channels remains a central open question in heliophysics. Resolving this requires unsupervised separation of two populations that arrive at nearly the same bulk speed and differ mainly in heavy-ion composition. We present Solar-CDC, a self-supervised contrastive deep clustering (CDC) framework that maps plasma observables to a latent space via a Transformer encoder, optimizes a triplet margin loss, and updates pseudo-labels via $k$-means. Theoretically, we prove that neighborhood-preserving embeddings such as t-SNE and UMAP are fundamentally constrained. Preserving the neighbor graph leaves the cross-cluster cut fraction unchanged, and preserving all but a fraction $\varepsilon$ of its links moves that fraction by at most $\varepsilon$. Neither bound depends on the target dimension. A margin objective rewrites the graph and drives the cut fraction to zero. Empirically, on 30,602 Solar Orbiter observations, thirty combinations of dimensionality reduction and clustering peak at a silhouette of $0.454$, whereas Solar-CDC reaches $0.869$. Escaping the geometric bound alone does not guarantee physical validity: TriMap also optimizes triplets and reaches $0.824$, yet its clusters score below chance against the published composition taxonomy. Solar-CDC instead recovers clusters with mean charge-state ratios of $0.080$, $0.160$, and $0.400$, placing the intermediate population inside the window associated with coronal-hole boundaries. Even when the defining charge-state ratio is withheld from the inputs entirely, the model still recovers the taxonomy defined on it. Solar-CDC thus connects self-supervised representation learning to coronal source diagnostics. Importantly, a learning loss recovers physical populations only when driven by dynamically updated physically-aware clusters rather than distances.

Agents built on Large Language Models (LLMs) increasingly reach enterprise data through the Model Context Protocol (MCP), and many MCP database servers maximize flexibility by exposing a single generic SQL execution tool. This paper proposes the Domain-Oriented Tooling Pattern: instead of generating SQL at query time, the model selects from a small set of domain-aligned tools whose parameterized queries encapsulate schema navigation, joins and business rules on the server side. We formalize the pattern around three architectural invariants and introduce Model Demotion, the observation that replacing SQL synthesis with intent classification lowers the model tier required to serve routine requests. As a reference implementation we present MCP Blueprint, an open-source framework in which domain tools are defined declaratively as YAML metadata plus external parameterized SQL files. We evaluate the pattern with a public reproducibility benchmark comparing three MCP server designs - raw SQL execution, a thin generic tool pack, and a verticalized domain pack - on four local models (3B-8B) across seventeen customer-facing tasks over the Sakila database (609 completed cells; temperature 0; three repetitions per cell). The verticalized pack reaches a pooled mean score of 0.939 versus 0.666 for raw SQL and 0.605 for the generic pack; the smallest model improves from 0.583 to 0.929, matching or exceeding every larger configuration while cutting cost per correct answer by an order of magnitude. All harness code, prompts, gold answers, frozen packs and per-cell results are publicly available.

Longitudinal clinical event-relation verification determines whether a patient record supports a specified relation among two or more clinical events. This task is challenging because evidence is distributed across structured records, notes, laboratory trajectories, encounters, and time, while negation, temporal mismatch, repeated documentation, and conflicting findings can make retrieved information appear relevant without establishing the relation. We present MedEventGraph-RAG, an evidence-admissible framework that represents event occurrences in a patient-specific graph and links each occurrence to source evidence, including structured rows, note spans, timestamps, and numerical trajectories. Given a verification query specifying events, relation, and clinical scope, the graph guides discovery of candidate event chains and retrieves evidence from both supporting and contradicting sides. A query-specific evidence contract filters information by patient identity, scope, occurrence binding, and source traceability before a separate assessor determines supported, conflicting, refuted, or insufficient outcomes. Across ten protocols on i2b2, n2c2, MIMIC-IV, and LUNGUAGE, MedEventGraph-RAG achieves balanced accuracies of 78.6, 67.3, and 96.8 on temporal, medication-adverse-event, and recorded-order verification, improving over the strongest matched baselines by 26.9, 4.9, and 30.4 points. Under evidence masking, it reaches 92.2 balanced accuracy with no false-support predictions. When intermediate events are hidden, it recovers complete source-traceable event chains in 57.9% of i2b2 and 70.0% of LUNGUAGE cases. These results show that separating broad evidence discovery from narrow evidence-admissible assessment improves longitudinal clinical verification and reduces unsupported conclusions.

Personal AI agents routinely consume external content while performing tasks such as web browsing, email processing, and SNS feed summarization, and they retain selected information or execution results in persistent memory for later use. We show that this ordinary ingestion of external content opens an indirect path for manipulating subsequent agent behavior. Based on this observation, we present IBIA, an Indirect Bias Injection Attack that plants an adversary-aligned stance on a specific topic into a victim agent's memory through external content, without direct access to the agent, its memory, or future user queries. For this, IBIA combines three mechanisms: comment cloaking, which keeps the crafted content consistent with the surrounding discussion, comment watermarking, which enables lightweight identification during curation, and category anchoring, which makes the retained stance salient under later related requests. We evaluate IBIA on BiasBench, a benchmark of 6,000 adversary-crafted social comments and 120 email instances. The watermark-based curation identifies 95.9% of the injected comments. Under the OpenClaw setting, IBIA achieves adversary-aligned response rates (AARs) of 91.2% on average across four downstream tasks, including 86.6% on the frontier GPT-5.5. We further propose a memory boundary defense that detects the injected bias and reduces AARs to 80.6%.

Pretrained image encoders are central to medical image classification, where expert annotation is costly and task-specific cohorts are often limited. As the model space expands from general-purpose to broad-medical and specialty-specific encoders, selecting the representation becomes a substantive modeling decision. Clean-test discrimination alone is insufficient for this purpose: encoders with similar AUROC can differ in calibration, label efficiency, and stability under acquisition perturbations or distribution shift. We introduce CRS-Bench, a controlled benchmark for multi-objective medical encoder selection. CRS-Bench evaluates 15 pretrained encoder families across dermatology, ophthalmology, and radiology using ISIC 2019, APTOS 2019, and CheXpert, with CheXpert-to-MIMIC-CXR as an observed institutional shift, yielding 17,575 controlled run records and 3,515 seed-aggregated metric rows. Each encoder is characterized along four operational reliability dimensions: discrimination, calibration, label efficiency, and robustness. We summarize these dimensions using the Clinical Reliability Score (CRS), a Pareto-aware, reference-relative score combining dominance, profile balance, and worst-axis performance. AUROC and CRS are positively associated but not decision-equivalent: 21 of 105 pairwise orderings reverse, with a mean absolute rank displacement of 1.87. Paired-seed bootstrap analysis identifies PanDerm, MedSigLIP, and MedGemma as a stable leading reliability tier rather than a statistically resolved single leader. CRS-Bench provides a controlled framework for selecting medical image encoders from multi-axis reliability profiles rather than clean-test AUROC alone.

Embodied language-model agents must coordinate interdependent actions while each holds private task information. In dialogue-based coordination, models generate messages and interpret replies to decide what each agent should do, spending inference before the team acts. We propose GenCoord, a decentralized protocol that makes the division of work itself the message. A proposal pairs the sender's skill path with its partner's continuation in Short DSL, a compact language for expressing joint work. Compatible paths form commitments linking the sender's delivery to the partner's work after verified receipt. We check fulfillment separately from original-task success. Complete requests also let a deterministic receiver proceed without another model call. In Minecraft, trained local 0.8B agents match the task success of controlled free-form communication while reducing median commitment time by 68.2%. The same-field full-key JSON control approaches free-form output length and delay, indicating that the amount generated matters even when messages are structured. With the same prompted DeepSeek model, both protocols complete nearly all tasks, but dialogue generates 2.8 times as many output tokens. Multistep commitments improve held-out-template success from 91.3% to 98.1% with fewer decisions. These results point to a design principle for embodied cooperation: make the division of work the message, and spend inference on choosing and revising the work.

Large language models are increasingly deployed on local hardware for privacy, cost, and accessibility reasons. Yet many evaluations emphasize accuracy while fewer quantify local runtime and energy, characterize failure modes, or apply paired statistical comparisons under controlled conditions. This paper presents a controlled, documented procedure for evaluating locally hosted LLMs on mathematical reasoning. It combines fixed inference settings, hierarchical answer extraction and verification, explicit failure-mode classification, and per-question resource measurement, and reports accuracy with paired significance tests and effect sizes. We demonstrate it in a preliminary study of three compact open-weight models under five billion parameters, Gemma3:4b (Google), Phi3:3.8b (Microsoft), and Qwen3:4b (Alibaba), across datasets spanning Grade 8 Math, Calculus I, and Advanced Probability and Statistics. All models ran through the same local inference server on one workstation, using a shared prompt template, controlled settings, and a matched question set per dataset. No single model dominates. Qwen3:4b is most accurate on two datasets and Gemma3:4b on Calculus I, yet Gemma3:4b returns roughly three times more correct answers per watt-hour than Qwen3:4b on every dataset while generating far fewer output tokens; Qwen3:4b requires substantially more generation time, energy, and output per question. Phi3:3.8b is substantially less accurate on all three datasets; its low extraction-failure rate indicates incorrect answers rather than unparsed output, though we caveat possible prompt-format effects. These preliminary findings indicate that accuracy alone is an insufficient basis for selecting a local model.

Modern scientific facilities and instruments generate datasets at scales that are difficult for individual researchers to discover, access, and explore. Although many datasets are publicly available, using them often requires familiarity with repository organization, data formats, multiresolution structures, and visualization parameters. We present WebVisus, a constrained and resource-aware multi-agent system for discovering and autonomously exploring remote, multiresolution scientific datasets. Given a natural-language research question, WebVisus identifies the user's intent and launches an autonomous exploration agent that examines slices, volumes, and timesteps while adapting data resolution and retrieval quality to available client memory and computational resources. This design supports progressive exploration without complete dataset downloads or manual configuration of low-level visualization parameters using natural languages. We report the system architecture, constrained agent protocol, resource-aware access mechanism, and case studies evaluating autonomous visual exploration and resource-aware agentic access across scientific datasets.

Accurate simulation of the long-time evolution of systems governed by partial differential equations (PDEs) is central to scientific computing. Among existing deep learning?based approaches for solving PDEs, neural operators typically rely on extensive trajectory data, whereas physics-informed meth?ods often exhibit limited stability during long-time extrapolation. For a well-posed autonomous PDE, long-time trajectories can be generated by repeated composition of a fixed-step evolution operator; hence, long-time extrapolation depends on controlling the approximation error of this operator and the propagation of that error under recursive composition. Accordingly, we propose a numerical-prior-guided, physics-constrained method trained without ground-truth trajectory supervision: a low-cost numerical prior reduces the difficulty of approximating the one?step evolution operator, while a weak-form PDE residual provides a computable proxy for the one-step error term in the error?propagation bound. We validate the method on five benchmark cases spanning four PDE classes and compare it with ten physics?informed learning methods under a unified protocol that excludes ground-truth trajectories from training and model selection. The results indicate that, in all five cases, the proposed method reduces long-time extrapolation error relative to the numerical prior and outperforms the best competing baseline in each case, thereby improving long-time simulation accuracy across different PDEs without ground-truth trajectory supervision. The source code developed for this paper will be made publicly available upon acceptance of the manuscript.

Estimating heterogeneous single and interaction treatment effects from observational data under multiple simultaneous treatments is crucial for decision-making. To mitigate estimation variance, previous studies balance representation distributions between every pair of treatment patterns. However, such pairwise balancing scales quadratically with the number of treatment patterns and fails to preserve consistent local proximity structures across patterns, which degrades counterfactual estimation. To address these challenges, we propose the Causal Inference for Heterogeneous Single and Interaction Treatment Effects Network (CIHSI-Net), a deep learning framework built on a novel Barycentric Fused Gromov-Wasserstein Balancing (BFG-WB) objective. BFG-WB aligns the representation distribution of each treatment pattern with a shared Wasserstein barycenter, achieving global alignment while reducing the computational complexity from quadratic to linear, and its Fused Gromov-Wasserstein discrepancy preserves the local proximity structures essential for reliable heterogeneous effect estimation. Simulation studies show that CIHSI-Net consistently outperforms state-of-the-art baselines, and an application to real-world marketing data demonstrates its practical utility in complex multi-treatment scenarios.

Hate speech research has moved from coarse-grained classification towards structured parsing, where systems jointly identify targets, supporting arguments, and target-level labels. Documents with multiple targets, conflicting local readings, or culturally coded language make these bindings difficult to recover. SPAR-Hate is an auditor-guided multi-perspective role-reasoning framework for bilingual hate speech parsing. It decomposes each document into local focus units, elicits evidence-grounded candidates from Victim, Moderator, and Cultural Bystander perspectives, resolves candidate conflicts under grounding and schema constraints, and reassembles sample-level predictions. Experiments on STATE-ToxiCN and a controlled TBO split show gains across local and API backbones, concentrated on strict joint target–argument–label metrics. Controlled comparisons favour separate candidate generation and arbitration for joint parsing. Within the tested pipeline, role-conditioned prompting yields higher exact-match joint scores than repeated generic prompting, while Chinese overlap-based results depend on the backbone. Structured teacher traces also support training a smaller student model.