6924 papers
This paper introduces Token-Conditioned Reinforcement Learning (ToCoRL), a framework that leverages the intrinsic behavioral plasticity of Large Language Models to internalize and stabilize inference-time adaptations, enabling precise control over behavioral modes like switching from reasoning to direct answering without degrading overall capabilities.
This paper identifies value-scale amplification as the primary cause of instability in non-linear value decomposition for offline multi-agent reinforcement learning and proposes a scale-invariant value normalization technique to stabilize training, ultimately providing a practical recipe to unlock the full potential of offline MARL.
This paper introduces Geometrically Constrained Outlier Synthesis (GCOS), a training-time framework that generates virtual outliers in the feature space by respecting in-distribution manifold structures and using conformal shells to improve out-of-distribution detection robustness and provide formal error guarantees.
This paper introduces a novel Meta-RL framework featuring a hybrid actor-critic architecture with shared state representations and parameter-sharing mechanisms that significantly enhances sample efficiency and fast adaptation in non-stationary environments, as validated by superior performance on a decade-long real-world Building Energy Management Systems dataset.
The paper introduces SYNAPSE, a systematic, training-free framework that analyzes and stress-tests Transformer models by extracting layer representations and applying forward-hook interventions to reveal domain-independent internal organization, functional stability through redundant neuron subsets, and specific vulnerabilities to small manipulations.
This paper introduces IronEngine, a general AI assistant platform featuring a unified orchestration core and a three-phase pipeline that integrates diverse backends, adaptive memory, and extensive tooling to achieve high task completion rates while separating planning quality from execution capability.
This paper introduces GRACE, a novel dynamic scaling framework for Class Incremental Learning that adaptively balances model capacity through a cyclic "Grow, Assess, Compress" strategy to achieve state-of-the-art performance while significantly reducing memory overhead compared to purely expansion-based methods.
This prospective feasibility study demonstrates that a conversational AI system (AMIE) can safely and effectively conduct clinical history-taking and generate diagnostic suggestions in a real-world urgent care setting, achieving high patient satisfaction and diagnostic accuracy comparable to primary care providers while requiring no real-time human intervention.
LycheeCluster is a novel KV cache management method that employs structure-aware chunking and hierarchical indexing to transform cache retrieval into a logarithmic-time process, achieving up to a 3.6x inference speedup with minimal performance degradation for long-context LLMs.
This paper investigates world model-based anomaly detection under gradual observation drift, revealing a universal sharp detection threshold that depends on the interaction between detector sensitivity, noise floor, and environment-specific dynamics, while identifying critical failure modes such as the undetectability of sinusoidal drift and agent collapse prior to detection.
This paper presents an adaptive entropy-driven sensor selection policy within a camera-LiDAR particle filter that dynamically switches between modalities to optimize tracking accuracy and continuity for single-vessel surveillance, validated through real-world maritime deployment.
This paper introduces , a novel predictive uncertainty framework that leverages data-driven priors derived from cross-modal similarities and modality-specific corruptions to significantly enhance both the performance and reliability of multimodal in-hospital risk prediction using MIMIC-IV and MIMIC-CXR datasets.
This paper proposes a Conditional Information Bottleneck framework that reframes efficient Chain-of-Thought reasoning as a lossy compression problem, introducing a semantic prior and a reinforcement learning objective to prune redundant tokens while preserving essential logic and improving accuracy.
The paper introduces MUSA-PINN, a multi-scale weak-form Physics-Informed Neural Network that reformulates PDE constraints as integral conservation laws over hierarchical control volumes to overcome convergence pathologies and significantly improve accuracy and mass conservation in fluid flow simulations within complex geometries like Triply Periodic Minimal Surfaces.
This paper proposes a model-based reinforcement learning framework that integrates Lagrangian neural networks into the Dyna architecture to enforce physical laws and improve prediction accuracy, demonstrating that state-estimation-based optimization converges faster than stochastic gradient-based methods during training.
The paper proposes STRIDE, a hybrid dynamics learning framework that combines a Lagrangian Neural Network for energy-consistent rigid-body mechanics with Conditional Flow Matching for stochastic residual interaction forces, achieving significant improvements in long-horizon prediction and contact force accuracy for robotic systems in unstructured environments.
This paper proposes X-AVDT, a robust deepfake detector that leverages internal audio-visual cross-attention cues accessed via DDIM inversion to achieve superior generalization across diverse and evolving synthesis paradigms, supported by the introduction of the new MMDF dataset.
The paper introduces NN-OpInf, a structure-preserving, composable neural network framework for non-intrusive reduced-order modeling that outperforms traditional polynomial methods in accuracy and stability for systems with non-polynomial nonlinearities, albeit at the cost of higher computational training requirements.
This paper introduces PolarBear, a Bayesian racing framework that identifies Pareto-optimal anytime algorithms by adaptively sampling temporal Plackett-Luce rankings to eliminate dominated candidates without requiring objective bounds, normalization, or known optima.
This paper introduces "decalibration," an efficient method that generates credal sets as probability intervals for complex foundation models without requiring computationally expensive retraining, thereby enabling robust uncertainty representation in safety-critical applications.