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 employing task-specific dynamic compression. The method preserves core gameplay constraints while maximizing information density for design exploration. The framework utilizes GameGoData, comprising 55,060 development trajectories across various game dimensions, and GameGoBench for evaluation. Training a model on this data results in agents that perform comparably to frontier models on 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 execution. The pipeline resolves entities, discovers metadata, enforces read-only SQL, and composes dashboards through various checks. Evaluation showed success in metadata and SQL tasks, and the system successfully executed various dashboard generation tasks. The model's inference accounted for over 97% of the observed runtime in the reported scenarios.
Prefill-decode disaggregation requires managing dynamic imbalances in resource allocation for LLM serving. FluidPD introduces in-place elasticity to maintain Service Level Objectives (SLOs) by adjusting resource distribution without adding hardware. It uses FluidToken to handle transient imbalances by shifting work between prefill and decode workers when slack exists. FluidRole manages sustained imbalances by reassigning running workers between the two roles without restarting the model. This approach improved SLO attainment by up to 94.6% over static methods on production workloads.
Contrastive representations naturally involve a family of geometries that can be specialized based on semantic context, unlike fixed cosine similarity. The paper proposes using contrastive learning, exponential families, and information geometry to link probability distributions over anchors to Bregman geometries. This connection allows for defining "Anchor Divergences" to specify context-specific semantic geometries on learned representations. Modeling the anchor distribution under this framework effectively specifies context-dependent semantic similarity. Experiments demonstrate that anchor divergences provide an efficient method for specifying this context-specific similarity.
RadOnc-Agent is an agentic AI framework that orchestrates workflows across the radiotherapy care pathway using a Large Language Model (LLM) controller. It formalizes radiotherapy into four clinical phases and exposes 26 callable functions via a conversational interface to manage complex tasks. The framework successfully selected the correct function in 98.79% of single-function executions and achieved high completion rates for cross-stage and real-patient workflows. Evaluation demonstrated the technical feasibility of using an LLM to coordinate heterogeneous radiotherapy capabilities. These results establish a foundation for coordinating varied AI capabilities across longitudinal patient workflows.
The study investigates how Vision Transformers ground abstract concepts, such as "angry," when direct referential evidence is scarce. The authors hypothesize a metonymic grounding mechanism where abstract predictions are mediated by concrete anchor concepts like "fire." They apply Transcoders to vision encoders to recover intermediate features associated with concrete semantic labels. Experiments on icon datasets revealed structured metonymic circuits where perceptual primitives precede abstract targets. Causal interventions confirm that these metonymic intermediates are functionally necessary for grounding abstract concepts.
LLM agents lack continuous evolution from experience due to the high cost and parameter restrictions of fine-tuning. External memory helps accumulate experience, but current methods fail to transform it into generalizable knowledge. The SAGA framework proposes a method to achieve experience-grounded knowledge abstraction and utilization in agents. SAGA converts interactions into reusable procedures and principles with explicit conditions, linking them to execution evidence. This creates a feedback loop where accumulated knowledge guides future actions and updates hierarchical memory.
AegisFlow is a multi-agent AI framework designed for autonomous remediation and self-healing in fragile data ecosystems. It uses agents, including a Watchdog and a Repair agent powered by LLMs, to close the loop between detection and resolution. The framework employs Parallel Shadow Patching based on a Monitor, Analyze, Plan, Execute, Knowledge (MAPE-K) loop to verify patches in digital twins. Experimental testing showed significant improvements, reducing Mean Time to Repair by 98.1% and achieving a 92% patch success rate across various failures. This system reallocates data engineering time from firefighting to innovation and is deployment agnostic.
AI research agents currently optimize feedback without adequately governing how that feedback is interpreted and reused. EPOCH introduces an evidence-governed architecture to address this by incorporating explicit task contracts, typed memory, active falsification, and independent replay. This system evaluates candidates against the strength and scope of their supporting claims. EPOCH achieved state-of-the-art performance across several benchmarks, showing significant advances in executable constructions and proof-supported results. The research concludes that evidence governance is crucial for AI agents to produce trustworthy scientific discoveries.
Improved traffic forecasts do not automatically lead to better signal control decisions. A diagnostic study analyzed demand data to evaluate the value of various forecast types, including point forecasts and conformal intervals. Interface audits revealed that only a few controlled intersections allow for multiple effective actions based on the forecasts. The study suggests that forecast value depends on temporal observability, action identifiability, dynamics consistency, and objective alignment. A proposed protocol aims to diagnose where predictive improvements fail to result in operational benefits.
A Bayesian Mirror Architecture (BMA) is a self-referential generative framework involving circular recursion between sensory abstractions, meta-abstractions, and self-latent states. Consciousness is framed as an architectural property resulting from this circular structure rather than an optimization objective. The architecture operates over probability measures using optimal-transport geometry to define stability and coherence in the 2-Wasserstein metric. The framework introduces concepts like self-manifolds and Wasserstein epsilon-necks to interpret choice and learning arising from environmental affordances. Ultimately, stability and agency emerge from variational free-energy minimization within this system.
A machine learning analysis was conducted on a longitudinal cohort to predict Beck Depression Inventory-II (BDI-II) scores at 12 and 24 weeks after mindfulness interventions. The study used demographic data, clinical context, baseline scores, and engagement measures to model depression outcomes. Ridge Regression best predicted 12-week outcomes, while LightGBM performed best for 24-week outcomes. Clinically relevant patterns showed baseline severity is the strongest predictor, short-term results depend on clinical context, and long-term results rely on behavioral adherence and demographics. These findings suggest using interpretable, context-aware modeling for personalized mental-health support.
Fine-tuned geospatial foundation models show reduced accuracy when applied across different continents due to phenological and geographic shifts. Evaluation across 12 countries on three continents revealed a significant drop in overall accuracy from 0.65 in the US to 0.40 in Europe. Performance collapse was linked to misalignment between observation windows and local crop phenology rather than pure spatial transfer. Adjustments to class schemes and observation window compression partially recovered accuracy. Consequently, these models are only reliably transferable where observation windows align with local growing seasons.
KV cache compression is necessary for efficient LLM inference as context windows grow. Current compression methods only consider content relevance for eviction, ignoring sequential positional information. This leads to "sequential forgetting," where verbatim copying is broken when only the head of a sequence is kept. KVFetch introduces a training-free method to temporally prefetch positional successors for score-based compressors. This approach recovers verbatim copying and improves performance on tasks requiring sequential access.
Agentic AI automation requires more than probabilistic reasoning for safe deployment, as epistemic drift can cause behavior to diverge from expert constraints. The paper introduces BRaVeS, a framework called DNRS, to manage this by encoding expert constraints and using depth-aware access mechanisms. It further proposes the Lyapunov-Bounded Consensus Framework (LBCF) to formally bound recovery by mapping risk signals to an abstraction and enforcing shielded state transitions. Simulation results using industrial-control-system data showed that the LBCF achieved finite-step convergence without safety violations. These findings suggest that bounded governance behavior can be enforced in agentic systems through structured safety mechanisms.
The research investigates the resource requirements—state, context, or dynamical criticality—needed for accurate forecasting in sequence models, focusing on long-range temporal dependence. The study establishes approximation bounds for memory decay, showing that forecast error decreases exponentially with $r$, necessitating $r=\Theta(\log^2(1/\tau))$ states for a target error $\tau$. Fractional long memory alters the geometric relationship of forecast error, with finite context yielding an exact $1/L$ leading order term. The work also links predictive geometry to statistical information and demonstrates that vanishing forecasting error pushes recurrences toward criticality. Experimental results confirm these theoretical state and stability predictions across various model types.
SSVEP denoising is challenged by real physiological artifacts which degrade BCI performance. DenoFlow proposes using flow matching, where a field network regresses the velocity along a straight path between contaminated and clean trials. This approach integrates the denoising process forward from the observation, avoiding the need for paired ground truth data. Physiological artifacts from EMG and EOG signals were injected during training to simulate real-world noise. Experiments showed DenoFlow outperformed seven baseline denoising models in both signal fidelity and downstream decoding accuracy.
The framework GenST addresses the challenge of forecasting unobserved node states in spatio-temporal graphs where sensor data is sparse. It utilizes Large Language Models (LLMs) fine-tuned to extract semantic features from node descriptions to compensate for missing spatio-temporal signals. The approach employs a two-stage generative architecture: a Spatio-Temporal VAE followed by a Generative Transformer (GenT). This process reconstructs future states for unobserved nodes by being guided by semantic, geographic, and neighborhood context. Experiments on traffic and non-traffic datasets show GenST significantly outperforms existing methods in zero-shot prediction tasks.
HydroSphere is a data-driven framework for governing and self-healing wastewater infrastructure using machine learning. It integrates a TCN-LSTM model for forecasting, a PPO module for adaptive chemical dosing, and a deep autoencoder for anomaly detection. The system is evaluated using extensive historical water-quality measurements from 1940 to 2023. It incorporates tiered governance and safety bounds to ensure safer and more adaptive management of water systems. This framework supports objectives related to UN Sustainable Development Goals 6 and 13.
Latent multi-agent collaboration can utilize latent communication to avoid high inference costs associated with natural language. Existing methods focus on sender-side state fidelity, which causes overhead and redundancy. KITE is proposed as a training-free framework for task-oriented key-layer KV communication, prioritizing receiver-side task sufficiency. This framework selects task-effective key layers, transmits only the relevant latent working memory, and uses these layers for autoregressive reasoning. Experiments demonstrate that KITE significantly reduces communication volume, speeds up inference, and improves accuracy compared to full-layer communication.
The research studies trust-region optimization for smooth potential-interaction energies defined in the Wasserstein space. The method employs a quadratic model along pushforward curves and a Steihaug-Toint subsolver with an explicit second-variation operator. The analysis proves that under certain conditions, the objective is nonincreasing, and the Wasserstein-gradient norms converge to zero within $O(\varepsilon^{-2})$ outer trials. The results apply to arbitrary initial measures with finite second moments, with complexity independent of particle number for empirical measures. Numerical experiments validate the approach for various energy minimization problems.
A pipeline called CARE (calibrated adaptive rectification and escalation) is introduced to facilitate safe and efficient human-AI collaborative decision-making. CARE combines AI models and human reviewers to ensure decisions are safe and human-aligned by continuously learning from feedback. A novel adaptive calibration module within CARE guarantees risk control at every step for any AI model. Experiments on real-world datasets across driving, language, and robotics show CARE reduces human queries by 25-81% compared to baselines. This method achieves human-aligned decisions while increasing automation.
The research investigates quadratic weak-to-strong generalization in two-layer random feature networks using random matrix theory. The study derives deterministic equivalents for the population errors of optimally trained teacher and student models. For ReLU activations and spherical harmonic targets under a Gaussian universality assumption, the student error exhibits a quadratic improvement over the teacher's error. These results establish the general lower bound for this type of generalization. The analysis also characterizes the regimes for weak-to-strong generalization across various stopping times and targets.
The research focuses on learning transition kernels for time-homogeneous jump-diffusion processes using conditional diffusion models. The goal is to generate new sample paths from training data observed on a high-frequency discrete time grid. Theoretical work establishes non-asymptotic bounds for score estimation error and KL divergence between true and generated paths. Numerical evaluation assesses the method on synthetic data against prior work and then on real-world data for probabilistic forecasting.
Multivariate ordinal data, often with covariates, is common in areas like language model alignment and recommender systems. Standard approaches often treat attributes individually or rely on pairwise comparisons, which can lead to coarsened observations. The authors propose a joint covariate-dependent consecutive ratio Markov random field model to address this issue. This joint modeling yields improved comparisons and allows for the derivation of pairwise or listwise comparison models. A maximum likelihood inference procedure is also developed despite an intractable normalizer.
Gaussian mean-field variational inference in wide Bayesian neural networks suffers from prior dominance due to the KL regularization term. Tempering the likelihood by raising it to the power $1/T$ scales the KL term by $T$, allowing for a balance between likelihood and regularization. The paper investigates the relationship between the tempering parameter $T$ and network width $M$ for single-hidden-layer linear networks. Limiting predictive moments exhibit phase transitions depending on the scaling $T = \tau/M^{c}$. These transitions reveal how the predictive expectation and variance relate to the infinite-width limit, the Neural Network Gaussian Process (NNGP) posterior.
A framework is developed for calibrating and controlling stochastic multiscale bioprocess models within a biological systems-of-systems context. Calibration uses quasi-likelihood estimation and adjoint sensitivity analysis on stochastic differential equation (SDE) models to account for parameter bias. Forward-backward adjoints quantify how calibration uncertainty affects value functions and policy performance. This framework enables policy-directed adaptive experimental design and uncertainty-aware optimization via a Gaussian-averaged objective. An Actor-Simulator algorithm is used to jointly update parameters, select experiments, and optimize control policies, showing improved performance over baselines.
Reflected anchored Langevin dynamics (RALD) is introduced as a reflected diffusion method for converging to non-differentiable targets on constrained domains, addressing limitations of standard first-order Langevin algorithms. The method employs a smooth anchored reference potential and scales the drift and noise covariance of the reflected Langevin dynamics by a state-dependent factor. The RALMC algorithm is derived using Euler-Maruyama discretization with projection. Explicit convergence bounds and iteration complexity in the 2-Wasserstein distance to the target distribution are proven for this algorithm. Numerical experiments validate the theoretical predictions of the method's performance.
Unpaired Canonical Correlation Analysis (UCCA) is a novel method designed to learn linear projections to maximize the correlation of underlying pairings without access to paired data. The method is derived by connecting the Quadratic Assignment Problem (QAP) to Canonical Correlation Analysis. UCCA is presented as the first approach to learn maximally correlated projections in a strictly unpaired setting. Validation on real-world multi-modal datasets shows UCCA significantly outperforms existing unpaired alignment baselines. This research addresses a gap between traditional statistical multiview learning and unpaired data learning.
The paper explores distribution-free guarantees for chain-of-thought (CoT) verifiers, specifically addressing scenarios with small calibration budgets. It introduces the concept of validity by abstention, suggesting that certificates can be valid even if they rarely fire, which allows them to avoid false positives. The research investigates how standard and improved certification methods perform in providing guarantees for CoT verification. Findings show that abstention effectively absorbs failures under benchmark shifts and selection procedures against the verifier. Ultimately, a derived certificate can avoid revealing what matters after deployment while maintaining coverage guarantees.