New study warns that vision-powered AI agents can hide backdoors. Multimodal large language models (MLLMs) let robots see, reason, and act — but a specific object in view can secretly flip them into an attacker’s plan. The authors introduce BEAT, the first framework to plant such visual backdoors using everyday
AI models for biology with downloadable weights promise faster drug discovery—but they also raise biosecurity concerns. A new study introduces a framework to test whether current “just filter the training data” safety practices actually work. * What’s tested: Models’ understanding of viruses from three angles: reading genetic sequences, predicting
MolChord in a nutshell Designing new medicines often means finding the right small molecule that “fits” a target protein. MolChord is an AI system that acts like a matchmaker, aligning protein shapes with potential drug molecules—and the words scientists use to describe them. * Speaks multiple “languages”: connects 3D structures,
What’s new? AI can call external tools (search, code, etc.) to solve problems, but juggling images, text, and many steps is hard. ToolScope is a new framework that helps multimodal AI plan globally and see locally — like using a telescope plus a toolbox. How it works * Global Navigator: sets
AI
TL;DR We often know what we want, but struggle to tell AI. This “intention gap” wastes time. This paper reframes AI from a passive follower to a Socratic partner that asks the right questions. * What’s new? An agent called Nous learns an inquiry policy that maximizes information gained
AI
Interpreting what large language models “think” is slow and expensive. Each new model often needs custom tools and lots of manual labeling. Atlas-Alignment offers a shortcut. Instead of rebuilding everything, it lines up a new model’s hidden activity with a shared, human-labeled Concept Atlas using only overlapping inputs and
DeepCompress: Smarter reasoning with fewer tokens Large reasoning models often overthink easy questions and underthink hard ones. DeepCompress fixes this with a dual reward strategy that teaches models when to be brief and when to explore. * Adaptive difficulty check: The system classifies each problem as Simple or Hard in real
Can a general-purpose time‑series AI learn brainwaves? This study tests a recent foundation model on electroencephalography (EEG) tasks like motor imagery (for brain–computer interfaces) and sleep stage prediction. * Pretrained on diverse, non‑EEG real‑world time series * Or pretrained entirely on synthetic signals * Then fine‑tuned for EEG
How do we know if an AI is truly reasoning—and where it fails? TempoBench offers a clear answer. TempoBench is a new, formally grounded and verifiable benchmark that lets researchers systematically probe multi-step reasoning, with difficulty you can dial up or down. * Why it matters: Ad-hoc tests capture real
AI
How can AI agents and developers find API and data services they can actually trust—especially when anyone can spin up fake identities to game the rankings? David Shi and Kevin Joo propose TraceRank for the x402 network, where services accept crypto payments for access. Instead of counting raw traffic,
AI
AI agents are getting good at coding and research, but they still struggle with long, domain-specific projects. Training them is hard: full-time human annotation is too costly, while pure trial-and-error rarely finds successful paths. Apollo is a new sampling framework that blends lightweight, asynchronous human guidance with strict action-level filtering.
Empty content