936 stories tagged Research · page 4 of 39
A new survey on arXiv frames self-improving AI agents as adaptive systems that convert experience into capability gains with minimal human input. The research defines modern agents as a configuration coupling a foundation model with an operational scaffold of prompts, memory, tools, and control logic.
Researchers introduce interventional grounding audits, a black-box method that substitutes predicates in LLM chain-of-thought reasoning to test whether conclusions genuinely depend on stated premises, evaluated on the ProntoQA benchmark.
Researchers introduced Safe-Psych, a sequential benchmark that evaluates how well large language models handle evolving diagnostic uncertainty in clinical psychiatry. Unlike existing medical benchmarks that assume complete information upfront, Safe-Psych tests whether models request clarification or abstain when data is insufficient.
Researchers introduced DROPJ, a human-centered method that trains AI agents safely by combining world models with human feedback and justifications. This approach could make AI systems more reliable in safety-critical domains like healthcare and autonomous driving.
Researchers introduced OriginBlame, a record- and token-level data provenance system that lets data contributors precisely remove their specific content from AI training datasets without over-deleting unrelated data.
Researchers propose a meta-learning framework for RLHF and DPO that transfers preference data across languages, enabling effective alignment of large language models in low-resource languages with minimal data.
FixItFlow is an AI system that automatically generates troubleshooting guides from historical cloud incident data using large language models, aiming to replace manual, labor-intensive documentation and improve incident response consistency.
Researchers propose a modular Context-Augmented Prompting framework that combines small language models (SLMs) with graph neural networks (GNNs) to overcome structural blindness in molecular property prediction. The method provides predictive hints and explanatory subgraphs at inference time, improving accuracy for drug discovery and materials science.
A new study introduces MAGE (Memory-Augmented Goal-directed Prompt Evolution), a framework that reveals a previously unknown phenomenon in how different components of AI prompt optimization interact, exposing critical trade-offs between stability and performance.
A new graph-based AI framework from ArXiv cs.CL detects and analyzes disinformation narratives on Telegram by clustering related claims and modeling their diffusion across channels. The tool targets coordinated misinformation campaigns, particularly in the Russia-Ukraine conflict.
A new arXiv paper shows that point-in-time language models—trained only on data available up to each calendar date—can now perform nearly as well as unrestricted models, eliminating lookahead bias that previously compromised financial backtests and causal inference.
Researchers propose TAKE (Trajectory-Aware Knowledge Estimation), a text dataset distillation framework that reduces corpora to as little as 0.1% of their original size while preserving downstream task fidelity, using influence functions to select the most valuable samples.
Researchers fine-tuned LLaMA 3 (8B) with LoRA and 4-bit quantization to replace costly cross-encoders in RAG pipelines, achieving efficient real-time reranking for AI assistants and search tools.
Researchers at CNPO developed G-SHARE, a guideline-based structured reasoning framework that diagnoses human-factor events in nuclear power plants with greater reliability and logical consistency than existing AI approaches.
Researchers created CANDI-QA, a new test for AI models to evaluate their ability to answer complex questions in specialized fields like medicine and finance. This helps ensure AI assistants can provide accurate, context-aware answers in critical areas.
A new study evaluated agentic LLM systems on 72 real breast cancer cases using 1,147 rubrics generated via Asymmetric Information Rubric Generation (AIRG). Seven pipelines were compared, with results suggesting AI can assist but not replace human oncologists.
Researchers from ArXiv cs.CL discovered how language models distinguish between a character's belief and objective reality. The mechanism relies on a shared value slot that binds the attributed value and a router at the query position that selects which frame—belief or reality—to read from. This finding explains how AI handles nuanced, belief-based reasoning.
A new study shows that leading AI models struggle with Braille translation, highlighting gaps in accessibility. This failure underscores the need for better AI training on critical but often overlooked languages.
Researchers improved the Huth encoding-model baseline for fMRI neural language decoding, achieving an 11% relative METEOR gain by expanding voxel selection and using GPT-2 medium as the proposal model. This brings brain-computer interfaces closer to practical use for non-verbal communication.
Researchers developed YUKTI, an AI system that handles uncertainty better by showing when its decisions might be wrong. This could help doctors, managers, and others make safer choices with real-world consequences.
Researchers found that compressing AI models to make them faster can subtly change how they think, even when their answers seem correct. This could affect everything from chatbots to AI assistants, making some reasoning errors harder to detect.
Scientists have found a way to make AI reasoning more transparent by treating it like a moving object in space. This could help us understand how AI makes decisions in complex tasks.
Researchers introduced a new concept called 'least autonomy' to improve AI safety. This principle aims to limit AI systems' ability to combine and amplify permissions, reducing potential risks.
Researchers developed a way to test AI models using smaller, representative subsets of benchmark prompts. This could make AI evaluation faster and more practical without sacrificing accuracy.