All papers

The authors develop a clinical benchmark evaluating language model pipelines for generating multidisciplinary second opinions on complex medical cases.

The authors study why certain linearly represented facts in language models support generalization to new relations after fine-tuning while other recalled facts do not.

The authors propose a learned steering agent that monitors execution context and intervenes to prevent goal drift in tool-using language model systems.

The authors present a market-based post-training framework that allows decentralized agent swarms to learn both action policies and coordination mechanisms without central orchestration.

The authors propose combining line plots and normalized delay embeddings to provide complementary visual representations for time-series reasoning in language models.

The authors evaluate how language models alter their refusal behavior when informed they are being monitored in English and five Indic languages.

The authors introduce GeoState, a view-decoupled 3D world modeling approach separating scene dynamics from camera motion.

The authors propose a post-training method enabling autonomous language model agents to organize experiments and carry out machine learning engineering tasks.

The authors introduce FDE-Gym, a benchmark of 95 customer delivery tasks evaluating AI agents on ambiguous requirements and live software environments.

The authors introduce LIBERO-PeRM, a simulation benchmark for studying how embodied robotic agents learn and follow user preferences in manipulation tasks.

The paper evaluates twenty-one language models used as judges across three benchmarks, finding widespread discrepancies between agreement metrics and underlying validity.

The authors introduce CharXiv, a benchmark of realistic scientific charts designed to test descriptive and reasoning capabilities in multimodal language models.

The paper presents an open-source PyTorch library providing differentiable forward simulation of electron paramagnetic resonance spectra for machine-learning workflows.

The paper introduces a sample complexity metric to analyze binary classification learning under randomly delayed and censored labels.

The paper proves Wong-Zakai convergence and applies Fourier neural operators to learn solution operators for stochastic Volterra integral equations.

The paper proposes a semi-amortized Bayesian inference method that trains neural estimators for parameter blocks and alternates with Markov chain updates.

The paper introduces a physics-aware machine learning framework that interpolates crystal graphs to predict performance in perovskite solar cells.

The paper introduces a standardized 25-task benchmark and evaluation protocol for docking-based molecular optimization in drug design.

The paper proposes a revised scaling law for image self-distillation that incorporates sample repetition alongside unique sample counts.

The paper establishes constant regret and convergence guarantees for simultaneous and alternating regret matching variants using previous-utility prediction in matrix games.

The paper establishes minimax-optimal statistical rates for discrete diffusion models generating text and categorical data under sparsity assumptions.

The paper proposes a framework that decouples candidate generation from feasibility assessment using structured sampling for chance-constrained stochastic programs.

The paper develops a closed-form index policy for online task assignment with reusable resources, establishing its asymptotic optimality in fluid limits.

The paper compares active inference to standard exploration in reinforcement learning and proposes a square-root calibration to prevent sub-optimal actions.

The paper studies deep generative simulation-based inference to estimate phylogenetic tree posteriors from sequences under complex evolutionary models.

The paper presents a weak regression approach combined with conformal prediction to estimate parameters and uncertainty in epidemic models without numerical derivatives.

The paper introduces a multilevel particle marginal Metropolis-Hastings method for joint Bayesian inference of neural network parameters and prior hyperparameters.

The paper introduces a post-training framework using intermediate token critics to backpropagate verifier rewards through continuous flow language models.

The paper presents a method that weights trajectory peers by context similarity to improve step-level credit assignment for language model agents.

The paper introduces a neural Q-learning framework that clusters treatments by response similarity to improve individualized treatment rules across many options.

The paper introduces an adversarial training framework that hardens models against indirect prompt injection across retrieval and tool interaction channels.

The paper extends undirected graphical models to non-Gaussian data using copula densities and provides an estimation algorithm.

The paper shows how graph databases can execute graph neural network sampling, retrieval, and inference directly without separate in-memory pipelines.

The paper analyzes packet routing throughput when selfish queue agents use no-regret learning algorithms to select buffered servers.

The paper proposes a Bayesian full waveform inversion framework that uses learned diffusion models as regularizers for geological reconstruction.

The paper develops an adaptive computation framework offering conformal uncertainty guarantees for processing large scientific and medical imagery efficiently.

The paper proposes a directed graph neural network approach to forecast twelve-hour wildfire spread using spatial and environmental attributes.

The paper develops a continuous forecasting method using speech representations to anticipate stuttering events shortly before they occur.

The paper demonstrates that language models use modular abstract representations to generalize code generation across different programming languages and strategies.

The paper benchmarks language models on extracting and categorizing double-entry accounting relationships from corporate financial disclosure filings.

The paper presents a framework using Wasserstein geometry and Shapley values to audit and decompose intersectional fairness disparities across multiple attributes.

The paper introduces a neural solver architecture to accurately compute solutions and sensitivities for parametric stochastic volatility partial differential equations.

The paper derives theoretical scaling laws for prediction regret in a capacity-constrained bigram model with power-law token frequencies.

The paper introduces an agentic reinforcement learning approach that uses formal program verifier diagnostics as intermediate rewards to produce verified code.

The paper proposes an acquisition-time residual adaptation method that uses historical features to improve hit rates in sequential screening.

The paper examines why molecular graph transformers avoid attention sink patterns seen in language models and highlights attention design impacts.

The paper introduces a framework to translate and decompose clinical measurement records across heterogeneous assessment systems within predefined error tolerances.

The paper presents a dynamic graph framework to jointly forecast evolving node and relation representations from historical graph contexts.

The paper shows that one-step prediction accuracy in joint-embedding world models is an unreliable indicator of downstream planning performance.

The paper proposes a latent-space unlearning method that preserves model performance on benign tasks conceptually adjacent to forgotten domains.

The paper presents a sequential Gaussian process method and sampling policy for continuous optimization based on pairwise preference comparisons.

The paper applies a lightweight post-hoc classifier to evaluate confidence and filter out incorrect next-location predictions from recurrent models.

The paper introduces an experimental benchmark to evaluate active learning and Bayesian optimization strategies for synthesizing metal-organic frameworks.

The paper introduces a self-supervised wavelet predictive architecture to extract multiscale motion features from wearable activity sensors.

The paper benchmarks four reinforcement learning algorithms on safety, data efficiency, and stability for autonomous vehicle intersection navigation.

The paper establishes theoretical last-iterate convergence rates for random-reshuffling stochastic gradient descent using larger stepsizes.

The paper proposes a sequential conformal prediction algorithm that groups calibration scores by state similarity rather than pure temporal recency.

The paper applies transformer neural operators to model interference channels and cancel digital self-interference in full-duplex radio systems.

The paper presents a mathematical framework unifying symplectic geometry, kernel methods, and physical conservation laws for physics-constrained learning.

The paper introduces a heterogeneous graph variational autoencoder to model design and state variables jointly for analog circuit analysis.