936 stories tagged Research · page 12 of 39
Researchers developed Heuresis, a framework to help AI agents explore high-quality, diverse, and novel ideas in machine learning. This could speed up scientific progress by making AI research more efficient and creative.
A small group of Wikipedia editors has shown that targeted edits can influence how AI models discuss animal welfare. Their work highlights the power of curated information in shaping AI behavior.
Researchers developed a method using large language models to automatically generate challenging test problems for AI reasoning. This could lead to better evaluations of AI's ability to apply knowledge to new, harder tasks.
Researchers introduced RIFT-Bench, a dynamic red-teaming benchmark that uses a graph-based methodology to evaluate security vulnerabilities across diverse agentic AI systems. Unlike static tests, it simulates adaptive attacks to find weaknesses before real hackers do, aiming to make autonomous AI safer.
A new study explores whether reinforcement learning (RL) on beneficial behavior can help AI systems generalize alignment beyond their training data, addressing risks like reward hacking and deception in high-stakes settings.
Researchers introduced a new way to align AI models with user preferences without expensive fine-tuning. This method uses simple instructions and a few examples to create natural-language prompts that guide AI behavior.
Researchers developed a neuro-symbolic framework that trains vision-language-action driving models using rule-grounded reasoning traces from classical planners, making autonomous driving decisions more transparent and causally connected to planned motion.
Apple researchers found a single parameter in AI models that can dramatically change how they behave. This discovery could make AI more predictable and easier to control.
A new study found that AI chatbots fail to provide consistent safety for most mental health conditions. Only suicide and self-harm have reliable safeguards in place. Researchers call for better standards and transparency.
Researchers developed SciLens, an AI system that can verify scientific claims by analyzing both text and images. This tool could make it easier for scientists and the public to check the accuracy of research findings.
Scientists developed a new method using Shapley values to measure how adjectives influence AI responses. Their findings reveal how small word choices can significantly impact AI performance and alignment.
Researchers developed PeerCheck, a system to enhance AI-generated academic reviews toward human-level quality. It addresses challenges in the traditional peer review process as academic submissions grow.
Researchers developed PEAR, a dynamic AI debate system that improves reliability by changing roles and reducing biases. This could make AI responses more trustworthy and consistent.
A new research paper introduces a reference architecture for 'agent skills' — reusable, externalized behavioral knowledge that LLM agents can discover, activate, and interpret at runtime. The framework formalizes how skills are bound to context and authority, interpreted by stochastic agents, and recorded as run evidence.
A new study examines how different AI reasoning strategies perform under varying conditions. The findings show that some methods hit performance limits even with more computing power. In plain English, this means AI problem-solving isn't as flexible as we thought.
A new paper finds that offline recommendations to mix AI models from different "families" for diversity may not hold up in real-time interactive settings—the very environments where multi-LLM systems are actually deployed. This discovery could change how multi-AI systems are designed.
Researchers propose a new language to clearly define how AI and humans should work together in software development. This could make teamwork between people and AI more predictable and reliable.
A new paper argues that the success of modern AI vindicates a modest form of associationism: the idea that learning is uniform, gradual, and driven by feedback. This could reshape how we think about both AI and human cognition.
Researchers developed LAGO, a system that helps AI robots plan tasks by understanding human language. This could make robots more useful in everyday settings. The framework predicts intermediate goals from language, reducing errors in planning.
GPT-5 Pro helped immunologist Derya Unutmaz solve a three-year-old puzzle about how T cells behave, specifically regarding an unusual subset of T cells. This breakthrough could accelerate research into cancer and autoimmune diseases.
Researchers created FirstPass, a dataset and AI model trained on real multi-round peer-review dialogues from Nature Communications. This could make AI evaluations of scientific papers more accurate by learning from actual editorial decision-making, not just stylistic mimicry.
A new paper proposes a roadmap for AI that learns through interaction with complex environments, mimicking natural evolution by systematically removing human priors. The authors have released an open-source infrastructure called Darwin Mobile Agent to test this idea using mobile graphical user interfaces as a practical proxy for an open-ended world.
Researchers have developed AlphaMemo, an AI system that helps financial trading agents learn from their past decisions. This could make AI-driven trading more reliable and less prone to repeating errors.
Researchers found that AI models can still make irrational decisions even when trained to align with human values. This 'rational value risk' means models may not always choose the best possible action, even if they understand what's valuable.