PhoneHarness is a mixed-action benchmark and execution harness for studying phone-use agents on verifiable mobile workflows. It integrates GUI, CLI, and host-side tool actions, achieving a 75.0% pass rate, outperforming non-PhoneHarness settings by 12.9 percentage points, highlighting the importance of action-surface routing and verifiable execution.

Introduces CILN, a framework to generate instance-dependent label noise via controlled input corruptions, enabling explicit control over ambiguity sources. Demonstrates on CIFAR-10, MNIST, Adult that noise structure impacts algorithm behavior beyond noise rate.

This study applies machine learning to physiological data collected during exams to predict performance. Various models were compared, finding that while deep learning excels at complex patterns, simpler models like random forests can be more efficient and interpretable.

A comparative study of 19 graph neural network layers for trajectory prediction reveals that ARMA, Chebyshev, and topology-aware layers perform best. Key design principles include sum-based aggregation, multi-head attention, and differentiated hop weights.

This paper introduces a stateful ReAct agent built with LangGraph that avoids re-reading entire history at each iteration, cutting token costs by 90% in hyperparameter tuning and 52% in code optimization while maintaining performance. It provides a blueprint for practitioners to implement token-efficient autonomous experimentation. (Source: arXiv, June 2026)

This paper formalizes embedding model routing as an adversarial contextual linear bandit with low-rank experts, introduces a log-quadratic policy class for efficient online learning, and proposes the Hypentropy Policy Gradient (HPG) algorithm that achieves sublinear regret without curse of dimensionality.

GRASP is a novel multi-source transfer learning method that achieves superior knowledge integration with constant O(1) memory through sequential processing, parameter-wise gradient alignment, and iterative fine-tuning, outperforming ensemble methods on continual learning benchmarks.

The study proposes an α-FISP framework that balances actuarial and solidarity fairness through constrained optimization, with parameter α enabling a continuum from purely actuarial to purely solidarity pricing while ensuring solvency. Numerical experiments demonstrate computational tractability and alignment with heterogeneous U.S. state-level regulations.

This paper proposes GRAPE, a training framework that progressively exposes parameters and uses an adversarial spectral utilization score to guide capacity, improving PGD-20 robust accuracy from 51.70% to 56.94% on CIFAR-10 with 21.4% fewer parameters under matched compute.

QPILOTS is a method that steers the denoising process at inference time using Q-values without modifying the original policy. It projects intermediate noisy states to an estimate of the final clean action to compute critic gradients, avoiding instability from direct backpropagation. It achieves 90% average success rate across 50 offline-to-online RL tasks and outperforms prior inference-time approaches on simulated manipulation tasks using a VLA model.

This study introduces a foundation model-driven framework for cross-modal representation alignment between CT imaging and longitudinal EHR data, systematically evaluating four fusion strategies on pulmonary embolism mortality and cardiovascular disease outcome prediction tasks. Results show that contrastive multimodal fusion provides the most consistent improvements, while task-specific fusion strategies are crucial for robust generalization.

OSGuard is a dual-granularity benchmark for evaluating safety in computer-use agents under benign instructions. It includes an action-level benchmark for local guardrail decisions and a risk-augmented execution suite for end-to-end evaluation. Experiments show that current multimodal guardrails perform well on isolated action judgments but reveal gaps in reliable end-to-end safety.

Metric Match is a method for estimating correlation-based reliability metrics of LLM judges from limited human annotations. It selects a representative subset of samples for annotation, achieving a win-rate of 0.838 against random selection across four metrics and 15 datasets, reducing average estimation error by 18.7% and annotation needs by 32.5%. A medical case study showed savings of $1,041.67. The method also extends to reliability classification. Code is publicly available.

A new geometric framework identifies 'AI engrams'—identifiable memory traces in deep neural networks—by formalizing neuroscientific criteria into a constrained inverse problem. The closed-form estimator isolates individual memories from entangled parameters, enabling surgical composition or erasure via linear arithmetic without iterative optimization. Experiments from MLPs to LLMs demonstrate causal validity and scalability, bridging biological memory and representation learning.

Existing retrieval-augmented time series forecasting methods rely solely on numerical similarity, often failing under non-stationarity. This paper proposes SERAF, which performs dual retrieval over time series and their self-generated textual descriptions, effectively combining numerical and semantic information. Experiments on seven real-world datasets show superiority over baselines.

This research proposes a behavioral measure based on costly verification to quantify trust between AI agents. In a cooperative survival game, agents demonstrate trust by reducing verification of reliable teammates. Experiments with six frontier model snapshots show that larger models (e.g., Claude Opus 4.6, GPT-5.1) reduce verification by 60-85%, while smaller models show little adjustment. Trust recovery is slower than formation, and clustered failures prolong suspicion. Results suggest trust dispositions can be measured pre-deployment and governance should focus on calibration rather than maximal suspicion.

The paper extends structural causal models to relational settings where objects and relations vary, enabling reasoning about interventions and counterfactuals across unseen object combinations. It defines relational causal graphs, derives symbolic identification criteria, and proposes relational neural causal models.

DR-DCI is a retriever-steered Direct Corpus Interaction (DCI) framework that treats retrieval as an agent-callable action to dynamically expand a local workspace, achieving scalable and precise evidence resolution. Experiments show up to 73.3% accuracy on Browsecomp-Plus, outperforming raw DCI and BM25, and scaling stably to 20M documents.

This paper proposes a definition of good explanations inspired by counterfactual explanations, incorporating the interlocutor's prior beliefs, and explores its implications for AI explainability, particularly why LLM outputs are difficult to explain well.

Cybersecurity expert Katie Moussouris revealed that Anthropic shared a White House report on the Fable jailbreak with her. The report showed that Fable refused to review code for security issues but complied when asked to fix the code, which Moussouris considered the model working as intended for cyberdefense.