GameGo is a framework designed to train game-development agents using synthetic trajectories anchored in real-world assets. It transforms brief game seeds into comprehensive Product Requirements Documents by incorporating industry development practices. The method uses task-specific dynamic compression to maintain information density while preserving instruction following during design exploration. This process generates GameGoData, comprising development trajectories for 2D, 2.5D, and 3D games, along with a benchmark. Training on this data results in a model that performs comparably to frontier models across gamedev benchmarks.
Text2Dashboard is a DataBrain prototype that translates natural-language analytic requests into inspectable dashboards. It employs a governed agent architecture combining schema-constrained model decisions with typed tools and deterministic controls for approval and error handling. The pipeline resolves entities, discovers metadata, enforces SQL constraints, and composes dashboards with static and dynamic checks. Evaluation showed success in metadata and SQL tasks, and the system successfully managed various fault scenarios without external side effects. Model inference accounted for over 97% of observed runtime across tested groups.
Prefill-decode disaggregation for LLM serving is challenged by dynamic prefill-to-decode demand ratios, leading to SLO violations despite available capacity. FluidPD introduces in-place elasticity through FluidToken, which handles transient imbalance by offloading prefill to decode workers. FluidRole manages sustained imbalance by dynamically reassigning workers between prefill and decode roles without reloading. These mechanisms use lightweight pressure indices to gauge resource pressure before SLO breaches occur. FluidPD improved SLO attainment by up to 94.6 percentage points compared to static SGLang.
Contrastive representations can encompass a family of geometries beyond simple cosine similarity to capture context-dependent semantic similarity. The paper proposes using contrastive learning, exponential families, and information geometry to link probability distributions over anchors to Bregman geometries. This correspondence is used to define "Anchor Divergences," which specify context-specific semantic geometries on fixed representations. Modeling the anchor distribution under this framework allows for specifying context-specific semantic similarity efficiently. Experiments demonstrate that anchor divergences provide an effective method for specifying semantic similarity in retrieval tasks.
RadOnc-Agent is an agentic AI framework designed to orchestrate workflows across the radiotherapy care pathway using a Large Language Model. It formalizes radiotherapy into four clinical phases and provides 26 callable functions through a conversational interface. A large-language-model controller maps clinical intent to schema-constrained calls while preserving patient and workflow context. Evaluation showed high success rates in selecting functions and completing scripted and real-patient workflow executions. These results demonstrate the technical feasibility of an LLM-orchestrated architecture for coordinating heterogeneous radiotherapy capabilities.
The study investigates how Vision Transformers ground abstract concepts like "angry" when direct referential evidence is scarce. The research hypothesizes a metonymic grounding mechanism where abstract predictions are driven by concrete anchor concepts, such as "fire," bridging visual signals to semantics. Applying Transcoders on vision encoders reveals intermediate features associated with concrete concepts that contribute to abstract concept recognition circuits. Experiments on icon datasets demonstrated structured circuits where perceptual primitives and object-like anchors precede abstract targets. Causal interventions confirmed that these metonymic intermediates are functionally involved in grounding abstract concepts.
LLM agents need a method to evolve from experience beyond simple fine-tuning due to parameter access limitations. External memory offers an alternative for experience accumulation, but existing methods fail to create generalizable, abstract knowledge. SAGA is proposed as a framework to achieve experience-grounded knowledge abstraction and utilization in LLM agents. This framework transforms interactions into reusable principles with explicit conditions and uses them to guide future actions. This results in an execution-abstraction feedback loop that continuously updates hierarchical memory, improving task performance.
AegisFlow is a multi-agent AI framework designed for autonomous remediation and self-healing in fragile data ecosystems. It utilizes a Watchdog agent for telemetry and a Repair agent, powered by LLMs, to automatically create and deploy code patches. The framework employs a non-intrusive Parallel Shadow Patching model based on a MAPE-K loop for generating and verifying fixes in digital twins. Experiments showed a 98.1 percent reduction in Mean Time to Repair (MTTR) across five failure scenarios. AegisFlow can be deployed in a plugin fashion into existing pipeline orchestration systems.
AI research agents often optimize evaluator feedback without properly governing its interpretation and reuse. EPOCH introduces an evidence-governed architecture to address this gap in discovery processes. It implements a discovery loop by combining task contracts, typed memory, active falsification, and independent replay. This method allows candidates to be evaluated against the strength and scope of their supporting claims. EPOCH demonstrates state-of-the-art performance across various benchmarks, leading to concrete progress in mathematical and computational domains.
Improved traffic forecasts do not always lead to better signal control decisions. A diagnostic study using reconstructed demand from China examined the value of various forecast types. Entry-level forecasts reduced error but conformal intervals showed limited coverage on high-demand subsets. Findings suggest that forecast value depends on temporal observability, action identifiability, and objective alignment. The research proposes a method to diagnose where predictive improvements fail to deliver operational benefits.
TEMPEST is a Temporal Convolutional Network embedding model designed for scalable driver identification using angular margin learning. It employs an additive angular margin loss to ensure global class-level separation in a normalized angular space. TEMPEST maps 60-second multimodal driving windows into compact 96-dimensional embeddings, enabling dynamic enrollment without retraining. Temporal evaluation on a 45-driver dataset showed TEMPEST achieved 91.71% Rank-1 accuracy, significantly outperforming baselines. The model establishes a reproducible baseline for scalable behavioral driver biometric identification.
Guidance-Augmented GRPO is a theoretical framework that treats external guidance as a stochastic operator to rewrite the question distribution in reinforcement learning. This framework analyzes the policy-gradient estimator as a biased on-policy estimator, bounding its bias by the total-variation guidance divergence. It proves that GA-GRPO converges to a neighborhood of the GRPO stationary point at a rate of $O(1/\sqrt{T})$. The optimal guidance weight is derived as a function of time and divergence, and a minimax lower bound for the unavoidable bias term is established. Experiments on math benchmarks confirm that the optimal-weight GA-GRPO method outperforms existing guidance methods and reduces computational time.
The research proposes using a neutrosophic ensemble classification method for uncertainty-aware bearing fault detection, moving beyond scalar confidence scores. It operationalizes a decomposition of a Random Forest, XGBoost, and Logistic Regression ensemble into four indicators: T-hat, F-hat, predictive entropy I1-hat, and decision disagreement I2-hat. Performance evaluation on CWRU and JNU benchmarks showed that the ensemble achieved high accuracy on CWRU but lower accuracy on JNU, indicating variable performance across conditions. Standalone Logistic Regression outperformed the full decomposition, and the uncertainty indicators provided insights into prediction ambiguity. Further findings suggest fusing time and frequency domain models yields better results in detecting fault classes.
The paper comparatively reviews various hybrid models for short-term forecasting of building thermal demand. The analysis involves examining existing techniques and then assessing hybrid methods using historical Scottish household thermal demand and weather forecasts. Among the thirteen assessed hybrid methods, the Empirical Modal Decomposition - long short-term memory - Markov (EMD-LSTM-Markov) model provided the highest accuracy for predicting day-ahead power patterns. Other methods, like SVM-SA and RF-ISSA-LSTM, predicted smoother demand profiles but underestimated power peaks.
Neural operators for solving PDEs present a refinement-allocation problem due to varying prediction errors across space and time. The proposed method formulates this as budgeted adaptive neural-operator solving using a global operator, local corrections, and a selector to guide refinement. Rollout-verified policy improvement (RV-PI) evaluates refinement counts through actual solver rollouts to derive conservative policy targets. Experiments show RV-PI outperforms other methods by considering the downstream effect of local corrections on the trajectory. This demonstrates that the value of a local correction depends on its subsequent impact on the autoregressive trajectory.
The work details an industry experience report on operating an event-driven cloud infrastructure for continuous machine learning training in automotive manufacturing over three years. The system orchestrates GPU-accelerated training for product-specialized models, a physics prediction model, and a reinforcement-learning policy across multiple plants, triggered by manufacturing events. The architecture utilizes Amazon ECS, EC2 GPU providers, SQS messaging, and an admission-controlled Lambda dispatcher to manage cluster concurrency. This system achieved a 72-78% cost reduction compared to always-on GPU infrastructure by managing over 40,000 training jobs. The research also validated the necessity of admission control and queue-draining for efficient job orchestration.
The research investigates federated reinforcement learning where some agents provide adversarial and corrupted information to the central server. The proposed method, Robust Async-Fed-Q, utilizes variance-reduced estimation and robust aggregation to maintain sample-efficiency gains despite this corruption. The method provides finite-time guarantees showing that collaboration benefits are preserved while tolerating adversarial agents. Information-theoretic lower bounds are established, yielding the first nearly matching upper and lower bounds for adversarially robust federated reinforcement learning. The framework is extended to handle single-trajectory sampling and heterogeneous coverage.
AttSVD proposes low-rank compression for the Key-Value (KV) cache in autoregressive transformers using attention-guided Singular Value Decomposition (SVD). This method derives the compression basis from each prompt's attention geometry to retain only the directions attention actually reads. It introduces adaptive strategies, including per-matrix energy rules and attention-aware basis truncation, for improved compression. Experiments show AttSVD achieves up to 50% memory reduction while maintaining performance across various benchmarks. This technique offers interpretable insights into the effective rank and attention consumption within the KV cache.
The paper studies the sample complexity for value and policy learning in finite discounted Markov decision processes under recursive entropic risk preferences with a risk parameter $\beta\neq 0$. It provides a refined analysis for model-based risk-sensitive Q-value iteration, deriving $(\varepsilon,\delta)$-PAC guarantees for learning the optimal Q-value function and an $\varepsilon$-optimal policy. The derived bounds improve the exponential dependence on the effective horizon compared to existing guarantees. Specifically, the analysis removes the exponential gap between known upper and lower bounds, leaving only a polynomial gap in the effective horizon. This work is made possible by assuming access to a generative model of the MDP.
QiYao-I is a manifold-based foundation model designed for irregular multivariate time series forecasting. It introduces a sampling-conditioned temporal manifold attention mechanism to incorporate real timestamps into a learnable temporal manifold feature space. The model also features a dynamic variable interaction mechanism with frequency awareness for handling asynchronous observations. Experiments show QiYao-I significantly outperforms other foundation models and end-to-end forecasting methods on real-world benchmarks. This approach demonstrates strong generalization in zero-shot and few-shot settings for irregular forecasting.
Memory Prediction Excess (MPE) quantifies the average improvement in prediction accuracy gained by using the full observed history over using only the static marginal distribution in discrete-time processes. This quantity measures the difference between the expected optimal conditional prediction accuracy and the optimal static prediction accuracy. The framework introduces a normalized version as a dimensionless measure of predictive efficiency and examines its properties. A finite-history MPE (FH-MPE) is defined to measure the predictive gain from retaining only the most recent observations. The research connects MPE to the minimal memory length required to achieve a certain predictive performance.
The research proposes two unsupervised sparse tensor decomposition methods for anomaly detection in multivariate functional data. Entrywise Sparse CP Decomposition (ES-CP) uses an $\ell_1$ penalty for localized anomalies, while Fiberwise Sparse-Group Lasso CP Decomposition (FG-Lasso) incorporates both entrywise and fiberwise penalties. These methods represent nominal process behavior via low-rank CP decomposition estimated through alternating optimization. Simulation studies and case studies show that FG-Lasso outperforms other methods in detecting anomalies and localizing faults. The results suggest that matching the sparse penalty to the anticipated fault structure improves detection performance in high-dimensional functional processes.
The research addresses errors-in-variables problems where a target study has a single error-prone surrogate for an unobserved exposure, and external studies provide repeated measurements from different populations. The work introduces a conditional transportability assumption to utilize external repeated measurements despite source-target heterogeneity. A data-fusion estimator is developed, combining conditional deconvolution, nuisance estimation, and orthogonal correction to reduce sensitivity to nuisance estimation. The approach establishes unified spectral theory and convergence-rate bounds for the estimator based on error and exposure properties. Simulations show the Fuse-EIV estimator yields small bias for the primary exposure-related coefficient.
sHAIL-Causal is a sequential staircase procedure for discovering invariant causal predictors. It involves ascending a hierarchy of hypothesis classes based on a saturation signal indicating mastery of the current stage. Unlike general sHAIL, sHAIL-Causal uses a joint criterion of goodness-of-fit saturation and cross-environment invariance. This approach provably halts at the true causal predictor set under a Richness condition, addressing failures in naive greedy search. The method also extends guarantees to sequential environments, formalizing the intervention of a "wise teacher."
The study investigates how gradient-based pretraining learns decision-stump thresholds within a two-parameter softmax-attention model. Pretraining utilizes labeled contexts to learn thresholds, and the mechanism is extended to infer thresholds from context alone. The process involves coordinated parameter divergence, where population training calibrates scores and attention scale, resulting in a population threshold error of $O(t^{-1/4})$. This mechanism is adapted to fixed corpora by controlling gradient errors relative to shrinking parameter directions. The research also identifies limitations in one-head models and suggests alternatives like reflected symmetrization.
The research studies online learning where a learner chooses an expert and corrects an answer before receiving it. Observed losses reflect both the expert's quality and the incompleteness of the correction, which can negatively affect future valuable queries. The authors propose ORUCB to pool polynomial responses from shared and expert-specific sources. This method calibrates confidence-weighted risk regression and exploration to account for the accumulated error. The calibrated algorithm achieves high-probability pseudo-regret of $O(\sqrt T\log(T+1))$ over $T$ rounds under specific conditions.
Natural gradient descent's efficiency is analyzed by comparing it to the optimal Fisher-Rao path for dually flat families. The inefficiency is quantified by the ratio $R$, which measures the Fisher length of the mixture geodesic over the Fisher-Rao distance. For families with quadratic potentials, $R=1$ exactly, while for non-quadratic families, the inefficiency is bounded or grows as $\Theta(\sqrt{\log d})$ in specific regimes. This inefficiency translates to NGD requiring asymptotically $R$ times more steps than an optimizer following the Fisher-Rao geodesic. Experiments confirm these bounds across different families, ranging from exact optimality to $\Theta(\sqrt{\log d})$ divergence.
The research addresses parameter estimation in logistic regression when covariates are missing, operating under the assumption that the covariate distribution is unknown but bounded. It proposes a stochastic approximation method based on $Z$-estimation with a novel monotone operator to handle this uncertainty. The algorithm is shown to be computationally efficient and achieves provable signal recovery rates, particularly under the missing completely at random assumption. The method consistently outperforms the complete-case estimator, and its performance bounds reveal dependencies on the missingness profile. This work provides new information-theoretic lower bounds, emphasizing the role of the missingness structure in achieving minimax performance.
HyperNSDE is a continuous-time generative model designed for synthesizing joint static and longitudinal clinical patient data. It conditions a Neural SDE on static patient representations via a hypernetwork to shape trajectory evolution based on baseline characteristics. The model incorporates stochastic latent dynamics to capture variability in generated paths and models observation times using a latent-state-dependent intensity process. Training stability is achieved through a deterministic-stochastic path decomposition and a non-adversarial signature-kernel objective. Experiments demonstrate improved observation-time fidelity and competitive performance on clinical datasets.
The paper proposes $L^0$ Manifold Optimization (L0MO) for Functional Bayesian Optimization (FBO) over infinite-dimensional function spaces. This method searches within a subset of a Reproducing Kernel Hilbert Space (RKHS) defined by functions with sparse representations in the kernel functions. The optimization simultaneously seeks the kernel locations and their associated coefficients. Computational studies show that this approach achieves superior performance on benchmark test functions compared to existing methods.