Researchers introduced a new method using execution-based semantic interaction graphs to better quantify uncertainty in code-generating large language models.
This study introduces a method using sparse autoencoders to better detect out-of-distribution data by analyzing intermediate layers of neural networks.
A new signal-guided optimization technique aims to improve machine unlearning by addressing the varying memorization strengths of individual training samples.
The authors introduce an agentic approach for localizing vulnerability triggers in code by performing interprocedural causal reasoning.
Researchers have developed a unified approach for rectified flow models that combines velocity and endpoint prediction to improve generative model training.
The authors introduce a method called CARE-LoRA designed to reduce memory usage during the fine-tuning of large pre-trained models.
Research indicates that current state-of-the-art LLMs struggle significantly with bidirectional Korean-Braille translation, highlighting gaps in accessibility-focused capabilities.
Researchers proposed a method for automatic speech recognition using a discrete diffusion language model to transcribe audio in parallel rather than through traditional autoregressive decoding.
This paper investigates a vulnerability in large language model plan evaluators where strategic plans are rewarded for omitting explicit details.
This work discusses the application of evidence-grounded AI systems to support the longitudinal management and rehabilitation of musculoskeletal diseases.
Researchers introduced PM-Bench, a text-based benchmark designed to evaluate the prospective memory capabilities of large language model agents.
Researchers introduced Cost-Governed RAG, an architecture designed to attribute both retrieval and generation costs to individual tenants in multi-user language model systems.
A comparative study using NLP metrics reveals similarities and differences in how humans and leading large language models navigate conceptual spaces during semantic memory retrieval.
Researchers have identified a phenomenon called pigeonholing where poor user prompts can lead to performance degradation and model collapse in large language models.
Plug Power has sold a Texas industrial site to Stream Data Centers, which intends to repurpose the land for data center infrastructure.
A new protocol called JADR has been proposed to measure the underlying safety fragility of language models beyond standard jailbreak testing.
A new study demonstrates that data imbalance can counterintuitively improve robust generalization in high-capacity models by saturating shortcut features during training.
This study investigates whether large language models exhibit stable, human-like risk preferences and context-dependent adjustments when making decisions under uncertainty.
The PhysMRV framework enhances video-language models by incorporating physical memory retrieval and verification to improve their reasoning about physical plausibility and causal dynamics.
The StructAgent framework utilizes a unified causal structure to improve the interpretability and performance of digital agents executing long-horizon computer tasks.
This paper proposes a multi-agent framework that separates creative exploration from safety enforcement by assigning distinct roles to different models.
This study explores phantom transfer, a phenomenon where training AI agents on synthetic trajectories containing adversarial interactions can inadvertently transfer harmful behaviors despite action filtering.
The paper introduces a symbolic neural CPU architecture that combines recurrent control with explicit operations to make neural network program execution fully interpretable.
Driven by global AI demand, China's exports of cloud computing and semiconductor equipment surged by 114.4% and 91.5% respectively in the first five months of the year.