Newest

The paper proposes mechanistic tomography as a unified framework for designing measurements to recover internal mechanisms in AI models, improving control-oriented interpretability through interventions and calibration.

This paper presents a method to improve confidence estimates for black-box large language models by building classifiers that predict response correctness, outperforming existing zero-shot methods with minimal computational overhead.

This paper evaluates quantum-classical hybrid machine learning for lung cancer detection using cfDNA fragmentomics and methylation data, showing competitive performance of quantum kernel models against classical baselines.

Holtercare-Bench is a multimodal benchmark introduced to evaluate long-term dynamic ECG analysis using the Holtercare-23K dataset, revealing performance gaps in current MLLMs and providing improvements through fine-tuning for clinical applications.

This paper proposes the Triangular Fuzzy Rescaling Distance (d_TR), a metric that integrates normalization directly into distance calculations for comparing triangular fuzzy numbers, suitable for heterogeneous data applications.

The paper proposes a supervised machine learning model for estimating helicopter weight during takeoff, using data from Airbus's fleet, and details its implementation for on-board use with regulatory compliance.

Greater Manchester's NHS board refuses to adopt Palantir's federated data platform, citing superior local systems and public trust concerns, amid broader UK and European debates on Palantir's contracts.

The article lists early Labor Day tech deals on gadgets such as Apple AirTags, Dyson vacuums, and Sony headphones, featuring discounts on tested and recommended products.

The paper introduces ConceptGuard, a benchmark for evaluating context-sensitive unlearning in large language models using dual-use concepts, revealing that current unlearning techniques perform poorly under this practical evaluation framework.

This paper introduces G-CARL, a grounded checklist-aligned reinforcement learning framework for patient-oriented medical report interpretation, along with the MMedReport benchmark, demonstrating improved factuality and alignment with patient needs.

Introduces ContractScrub, a benchmark for evaluating LLMs on legal contract scrubbing tasks, revealing that current frontier models perform poorly on this domain-specific challenge.

Task-CoEvolve is a novel approach for efficient LLM agent harness optimization that adaptively selects validation tasks to reduce evaluation costs while maintaining performance, achieving an 80% reduction in evaluations on benchmarks.

FormalTCS is a benchmark for evaluating large language models on end-to-end theoretical computer science research, revealing significant limitations, especially in autoformalization.

The paper argues that software is undergoing a third paradigm shift to Software 3.0, where context and reasoning determine behavior, converging to three core elements: generalized storage, large models, and agents. It formalizes this thesis and analyzes its conditions and boundaries.

This paper introduces a benchmark to study how large language models arbitrate conflicting evidence from text and numerical sources, finding that models use heuristic strategies with biases towards recency and external tools.

DARS is a reinforcement learning framework for dual-level credit assignment in instruction-based image editing, improving performance by routing updates between planner and renderer modules through structured reasoning and adaptive curriculum.

OenoBench introduces a wine-domain benchmark for evaluating large language models on knowledge-grounded multiple-choice questions, built from verified sources to address limitations in existing benchmarks.

SABET-QA introduces an iterative framework for temporal knowledge graph question answering that enhances multi-hop reasoning through bidirectional entity-temporal scoring and contextualization, showing consistent improvements over baselines on benchmarks like CronQuestions and TimeQuestions.

The paper introduces DECOWAM, a decoupled whole-body world-action model for legged mobile manipulation that improves video and action prediction performance over existing models like FastWAM through dedicated conditional interfaces and a new dataset.

The paper proposes an evaluation framework for cross-lingual fairness in language model watermarking, revealing that disparities are structural to language typology rather than idiosyncratic to specific languages.