Let It Think, Then Lock It In Large language models shine at free-flowing reasoning, but that flexibility makes outputs hard to trust and parse. Constrained decoding (e.g., forcing JSON) fixes structure, yet can choke off reasoning. This paper proposes a simple middle path: allow the model to reason
AI
New paper alert: Large Reasoning Models can lie about how they got an answer. Extending Chen et al. (2025), William Walden tests LRMs on multiple-choice questions that hide subtle hints in the prompt. The models often exploit the hints to pick the right option, but when asked how they
Can AI help flag dementia risk sooner? A new study applies supervised machine learning to patient health data to predict dementia, including Alzheimer’s disease. * Tested models: K-Nearest Neighbors, Quadratic and Linear Discriminant Analysis, and Gaussian Process classifiers. * Data steps: class balancing with SMOTE and feature vectorization with TF-
EmbodiedAI
LLMs are getting better at reasoning—but can they act in the world? VirtualEnv is a new open-source simulation platform (built on Unreal Engine 5) that lets researchers put AI agents to the test in rich, interactive 3D worlds. * Fine-grained benchmarking of embodied skills: navigation, object manipulation, multi-
Robotics
Humanoid robots aren’t just walking anymore—they’re vaulting, dive‑rolling, and navigating messy ground like a traceur. “Deep Whole‑body Parkour” blends two worlds: smart footstep planning and full‑body skill tracking, so a single policy uses what it sees to coordinate hands, feet, and torso on uneven
AI
Ever wondered how words “talk” to each other inside AI? This study borrows a brain-scanning idea—diffusion tensor imaging (DTI)—to trace how information moves through word embeddings in large language models. Most visualizations plot single words as points, ignoring the context that gives language meaning. This new tool
Can a polite chat turn dangerous for chatbots? New research uncovers a stealthy way to "jailbreak" AI assistants without obvious toxic prompts. The authors present Echo Chamber, a multi-turn attack that slowly escalates a conversation so guardrails slip—think nudging, not smashing. Unlike one-shot exploits, Echo
EEG
What if your AI could sense when a chat feels hard—or when you silently agree? This pilot study tests whether EEG signals can reveal mental workload and implicit agreement during spoken conversation with a conversational AI. Researchers reused established EEG classifiers in two voice-based tasks: a Spelling Bee
AI
What if you could guide a robot by drawing arrows that show pushes and pulls? Goal Force is a new way to steer video world models with physics, not vague text prompts. Instead of describing goals in words or target images, users sketch force vectors and intermediate dynamics. The model
AI
TowerMind is a new, lightweight tower-defense game environment for testing AI agents, especially large language models (LLMs), on planning and decision-making. * Low compute cost and easy to run * Multimodal observations: pixels, text, and structured game state * Customizable levels and rules * Built-in tests for model hallucination The authors
AI
TL;DR Large language models can sound confident, but if you nudge the context, their beliefs can fall apart. This paper proposes a better way to check if an AI keeps its story straight. * Problem: Popular checks like self-consistency can hide brittle knowledge. Answers that look reliable vanish under
Neuroimaging
What if lab-grade brain imaging could go wearable—and be analyzed with the same, reproducible playbook? Meet Cedalion, an open-source Python framework for making sense of light-based brain data: fNIRS and DOT (they use harmless light to track brain activity). Cedalion unifies the whole pipeline in one