All papers

This paper introduces a task and benchmark for tracking candidate sets of referent objects across ambiguous interactive dialogue turns.

This paper presents a training-free contrastive decoding framework using attention-guided perturbations to reduce hallucinations in video language models.

This paper analyzes three distinct failure modes where reinforcement learning agents misgeneralize goals despite competent behavior.

This paper proposes a multi-agent system designed to reduce clinical medication errors through dynamic collaborative coordination.

This paper pairs language models with video diffusion to maintain persistent abstract states across dynamic video world interactions.

This paper trains language models using reinforcement learning to generate explicit perception-to-action rules for theory-of-mind reasoning tasks.

This paper introduces a benchmark evaluating whether coding agents can reimplement and train biological AI tools from published scientific papers.

This paper presents a compact spatio-temporal forecasting model designed to generalize across shifting spatial domains and varying observation availability.

This paper proposes a self-play reinforcement learning framework that automatically generates tasks and intermediate verifiable rewards for long-horizon agent training.

The paper presents a continuous trajectory inference method that regularizes population dynamics using neural splines and Wasserstein smoothness.

The paper proposes a sample-based neural framework to learn fixed-size embedding representations for higher-order objects like distributions and graphs.

The paper describes a vision-language model using task-specific low-rank adapters for multimodal medical image parsing.

The paper investigates whether machine translation systems preserve or neutralize media bias across languages.

The paper proposes an energy-based learning method that optimizes acceptance rates for finite-step discrete graph sampling.

The paper proposes a tree search method that optimizes agent skills through contrastive atomic rule editing and execution memory.

The paper proposes a federated learning framework with risk-aware aggregation to prevent malicious updates from compromising global model safety.

The paper evaluates harnesses for legislative document conversion, showing structured multi-agent pipelines narrow performance gaps between models.

The paper provides a theoretical analysis of stability and convergence for semi-gradient temporal-difference learning using mean-field neural networks.

The paper identifies an activation steering vector that controls whether language models rely on internal memory or retrieved external context.

The paper investigates the loss surface of deep networks using a specific polynomial activation function without standard normalization layers.

The paper compares three textual interface designs for optimizing prompts in multimodal audio-visual question answering systems.

The paper proposes a synthetic data pipeline that extracts atomic facts from source texts to guide factual pretraining for language models.

The paper describes an automated system using language models to verify corporate loan applications against internal bank lending policies.

The paper introduces a benchmark to evaluate conversational floor control, turn completion, and interruptions in multi-party spoken dialogue systems.

The paper uses in-context neurofeedback sessions to evaluate whether language models can intentionally control their internal activations.

The paper presents a benchmark evaluating whether conversational agents agree on verifiable plans that satisfy both parties' private conditions.

The paper introduces a reasoning-guided translation framework to scale multilingual safety and reasoning benchmarks while preserving original task meanings.

The paper proposes a retrieval framework that selects suitable anomaly detection models for time-series data using similarity search over existing datasets.

The paper investigates an asynchronous training recipe combining the Muon optimizer, token-averaged loss, and penalties for long-horizon agent tasks.

The paper introduces a distillation method that trains agent policies on explicit runtime control semantics recovered from replayed execution traces.

The paper studies Byzantine robustness in fair federated learning and proposes an update screening method using client trajectory consistency.

The paper introduces a benchmark and evaluation framework assessing context-sensitive code-switching by language models in creative writing tasks.

The paper presents an intervention method using Shapley values to assign credit to individual actions within multi-action agentic retrieval sets.

The paper proposes a framework where an agent policy navigates a global search tree by alternating between reflection and execution roles.

The paper provides finite-sample convergence guarantees for average-reward option Q-learning algorithms under inter- and intra-option settings.

The paper presents a method to consolidate context gains into parametric model weights using prospective queries and low-rank feed-forward updates.

The paper analyzes when first-moment momentum accelerates reinforcement learning with verifiable rewards under exact and sampled policy gradients.

The paper introduces an adaptive perturbation method for sharpness-aware minimization that modulates perturbation radii using local loss curvature.

The paper evaluates failure modes and confidence calibration across frontier and lightweight language models during entity extraction from cyber threat intelligence reports.

The paper analyzes the performance of retrieval-augmented generation architectures on open-domain question answering tasks involving culturally grounded South Asian narratives.

The paper presents a recursive self-improvement framework using scientific agent swarms that mutate hypotheses and recombine code based on experimental feedback.

Proposes a severity-weighted character error rate that charges text transcription errors based on their practical impact on reader understanding.

Implements adaptive action grids in value-based reinforcement learning that selectively refine promising bins while pruning suboptimal actions.

Introduces a counterfactual risk management method to calibrate adaptive decision-making against shifting opponent behaviors within fixed deviation budgets.

Uses explicit three-dimensional vascular guidance in an iterative diffusion pipeline to improve anatomical accuracy in computed tomography super-resolution.

Presents a graph-based framework with decision certificates to determine when multimodal agents should acquire visual evidence under cost constraints.

Introduces a decomposed rubric synthesis method to create training data for open multimodal models that generate visual code artifacts.

Develops and tests a lightweight convolutional neural network to classify potato leaf diseases on field-collected farm images.

Introduces a benchmark based on real online discussions to evaluate both cognitive and behavioral persuasion skills in language models.

Shows that periodically clearing agent search history and restarting from the top candidate prevents automated research agents from plateauing.

Introduces a self-supervised consistency loss across varied tool interfaces to make language model agents robust against notation and naming changes.

Replaces discrete attention lookup in transformers with integration over a continuous memory curve to improve validation likelihood.

Analyzes the price of anarchy and economic efficiency of first-price auctions where AI agents bid to serve user prompts.

Introduces multi-step action forecasting targets alongside actor rollouts to train language agents more effectively from intermediate trajectory steps.

Proposes a hierarchical residual policy learning framework to fine-tune mobile robot manipulation policies with decoupled base and arm adjustments.

Generatively models humanoid robot control policy weights using a latent diffusion framework combined with a linear autoencoder and Diffusion Transformer.

The paper introduces an architecture that separates safe and unsafe visual knowledge during pretraining so unsafe branches can be disabled.

The paper proposes a physics-informed forward-inverse machine learning framework for sparse recipe formulation in industrial dyeing.

The paper develops a graph attention network to transfer prototype facade articulation patterns onto building massing targets.

The paper develops an early-warning pipeline using public balance-sheet filings to predict defined-benefit pension sponsor distress.